Generating a differentially private dataset using a generative model

Abstract

Disclosed are techniques for generating a differentially private dataset using a generative model. In particular, disclosed are apparatuses, systems, and techniques for training a generative model based at least in part on a private dataset. In at least one embodiment, the generative model is trained based at least in part on a differentially private Sinkhorn algorithm, e.g., using backpropagation with gradient descent to determine a gradient of a set of parameters of the generative model and modifying the set of parameters based at least in part on the gradient.

Description

[0001] Claiming priority

[0002] This application claims the benefit of U.S. Provisional Application No. 63 / 086,780, filed October 2, 2020, entitled “Learning Data Distributions from Private Data with Differential Privacy Using Generative Models,” the entire contents of which are incorporated herein by reference. Background Technology

[0003] Training neural networks and similar models for specific applications can be limited by constraints on training data. For example, availability and access to training data may be restricted in various privacy-sensitive domains. Furthermore, training data can be distributed among multiple entities with different privacy concerns. Generative models trained on privacy-sensitive data can provide indirect access to that data; however, existing techniques are time-consuming and costly to produce, and may inadvertently expose privacy data or otherwise provide suboptimal results. Attached Figure Description

[0004] Figure 1 An example of a differential privacy generative model according to at least one embodiment is shown;

[0005] Figure 2 An example is shown of training a model using training data generated by a differential privacy generative model, according to at least one embodiment;

[0006] Figure 3 An example of a forward pass for generating a differential privacy generative model, according to at least one embodiment, is shown;

[0007] Figure 4 An example of backpropagation for generating a differential privacy generative model is shown according to at least one embodiment;

[0008] Figure 5 An example method for training a model using training data generated by a differential privacy generative model, according to at least one embodiment, is shown.

[0009] Figure 6 An example method for generating a differential privacy generative model according to at least one embodiment is shown;

[0010] Figure 7 An example method for backpropagation for generating a differential privacy generative model, according to at least one embodiment, is shown;

[0011] Figure 8A The inference and / or training logic according to at least one embodiment is illustrated;

[0012] Figure 8B The inference and / or training logic according to at least one embodiment is illustrated;

[0013] Figure 9 The training and deployment of a neural network according to at least one embodiment are illustrated;

[0014] Figure 10 An example data center system according to at least one embodiment is shown;

[0015] Figure 11A An example of an autonomous vehicle according to at least one embodiment is shown;

[0016] Figure 11B The illustration shows an embodiment according to at least one of the embodiments. Figure 11A Examples of camera positions and field of view for autonomous vehicles;

[0017] Figure 11C This is an illustration based on at least one embodiment. Figure 11A A block diagram of an example system architecture for an autonomous vehicle;

[0018] Figure 11D The illustration, according to at least one embodiment, is for one or more cloud-based servers and Figure 11A A diagram of a system for communication between autonomous vehicles;

[0019] Figure 12 This is a block diagram illustrating a computer system according to at least one embodiment;

[0020] Figure 13 This is a block diagram illustrating a computer system according to at least one embodiment;

[0021] Figure 14 A computer system according to at least one embodiment is shown;

[0022] Figure 15 A computer system according to at least one embodiment is shown;

[0023] Figure 16A A computer system according to at least one embodiment is shown;

[0024] Figure 16B A computer system according to at least one embodiment is shown;

[0025] Figure 16C A computer system according to at least one embodiment is shown;

[0026] Figure 16D A computer system according to at least one embodiment is shown;

[0027] Figure 16E and Figure 16F A shared programming model according to at least one embodiment is shown;

[0028] Figure 17 An exemplary integrated circuit and a related graphics processor according to at least one embodiment are shown;

[0029] Figure 18A and Figure 18B An exemplary integrated circuit and an associated graphics processor according to at least one embodiment are shown;

[0030] Figure 19A and Figure 19B Additional exemplary graphics processor logic according to at least one embodiment is shown;

[0031] Figure 20 A computer system according to at least one embodiment is shown;

[0032] Figure 21A A parallel processor according to at least one embodiment is shown;

[0033] Figure 21B A partitioning unit according to at least one embodiment is shown;

[0034] Figure 21C A processing cluster according to at least one embodiment is shown;

[0035] Figure 21D A graphics multiprocessor according to at least one embodiment is shown;

[0036] Figure 22 A multi-graphics processing unit (GPU) system according to at least one embodiment is illustrated;

[0037] Figure 23 A graphics processor according to at least one embodiment is shown;

[0038] Figure 24 It is a block diagram illustrating a processor microarchitecture for a processor according to at least one embodiment;

[0039] Figure 25 A deep learning application processor according to at least one embodiment is shown;

[0040] Figure 26 A block diagram of an example neuromorphic processor is shown according to at least one embodiment;

[0041] Figure 27At least a portion of a graphics processor according to one or more embodiments is shown;

[0042] Figure 28 At least a portion of a graphics processor according to one or more embodiments is shown;

[0043] Figure 29 At least a portion of a graphics processor according to one or more embodiments is shown;

[0044] Figure 30 It is a block diagram of a graphics processing engine of a graphics processor according to at least one embodiment;

[0045] Figure 31 It is a block diagram of at least a portion of a graphics processor core according to at least one embodiment;

[0046] Figure 32A and Figure 32B The diagram illustrates thread execution logic according to at least one embodiment, which includes an array of processing elements of a graphics processor core;

[0047] Figure 33 A parallel processing unit (“PPU”) according to at least one embodiment is shown.

[0048] Figure 34 A general-purpose processing cluster (“GPC”) according to at least one embodiment is illustrated.

[0049] Figure 35 A memory partitioning unit of a parallel processing unit (“PPU”) according to at least one embodiment is shown;

[0050] Figure 36 A streaming multiprocessor according to at least one embodiment is shown;

[0051] Figure 37 This is an example data flow diagram of an advanced computing pipeline according to at least one embodiment;

[0052] Figure 38 This is a system diagram of an example system for training, adapting, instantiating, and deploying machine learning models in an advanced computing pipeline, according to at least one embodiment;

[0053] Figure 39 Example illustrations of an advanced computing pipeline for processing imaging data according to at least one embodiment;

[0054] Figure 40A Includes example data flow diagrams of virtual instruments supporting ultrasound equipment according to at least one embodiment;

[0055] Figure 40BIncludes example data flow diagrams of virtual instruments supporting CT scanners according to at least one embodiment;

[0056] Figure 41A A data flow diagram illustrating the process for training a machine learning model according to at least one embodiment is shown; and

[0057] Figure 41B This is an example illustration of a client-server architecture that utilizes a pre-trained annotation model to enhance an annotation tool, according to at least one embodiment. Detailed Implementation

[0058] Embodiments of this disclosure relate to differentially private generative models using optimal transmission. Systems and methods are disclosed that generate generative models trained under privacy constraints on privacy data, providing indirect access to that data. In various embodiments, the resulting generative model (e.g., a differentially private generative model) produces training data similar to the privacy data, which can be used to train neural networks or similar models for downstream applications without exposing the privacy data. Furthermore, the systems and methods described in this disclosure leverage a computationally efficient approximation of the optimal transmission distance between the data generated by the generative model and the privacy data to improve the performance of the generative model.

[0059] Compared to conventional systems such as those described above, the systems and methods described in this disclosure utilize optimal transfer-based generative approaches to learn data distributions from privacy-preserving data while maintaining differential privacy. In one example, the optimal transfer-based generative approach includes a differential privacy (DP) Sinkhorn algorithm. In such an example, the generative model is trained, at least by minimizing the Sinkhorn divergence and backpropagation gradients to modify the parameters of the generative model. In one embodiment, the gradients of the parameters represent a set of changes to the parameters of the generative model. For example, this set of changes includes adding, subtracting, or otherwise modifying the values ​​of the generative model parameters.

[0060] In various embodiments, during the training phase, the generative model generates a first data batch, at least in part, based on latent codes or other random or pseudo-random inputs. Additionally, during the training phase, a second data batch may also be generated from samples of privacy data. In such embodiments, a cost matrix is ​​generated, at least in part, based on the cost between element pairs from the first data batch (e.g., data generated by the generative model) and the second data batch (e.g., data sampled from privacy data). In one embodiment, a Sinkhorn loss is then computed to train the generative model using backpropagation and gradient descent. For example, the Sinkhorn loss is backpropagated to the first data batch using gradient computation methods. In various embodiments, to enhance privacy, pruning or otherwise adding noise to the gradients during backpropagation is used. Furthermore, during the training phase, in various embodiments, the gradients are further backpropagated to the generative model and used to update the parameters of the generative model. The training phase can be repeated based on various constraints. In one example, the training phase is repeated until the generative model converges. In one embodiment, convergence is achieved when the generated data matches or more closely matches the privacy dataset. In other examples, the number of times the training phase is repeated is determined based on various factors, such as privacy constraints.

[0061] Figure 1 An example of an environment 100 according to at least one embodiment is shown, in which a set of data curators 110A, 110B, and 110C generate a set of differential privacy generative models 104A, 104B, and 104C. In various embodiments, this set of differential privacy generative models 104A, 104B, and 104C is used by a data consumer 116 to generate training data to train an untrained neural network 108 or other models. As described in more detail below, in one embodiment, the set of differential privacy generative models 104A, 104B, and 104C is trained using corresponding privacy datasets 102A, 102B, and 102C. For example, differential privacy generative model 104A is trained using a privacy dataset 102A maintained by data curator 110A. In various embodiments, privacy datasets 102A, 102B, and 102C are maintained by privacy barriers 112A, 112B, or 112C.

[0062] In various embodiments, a group of data managers 110A, 110B, and 110C includes different entities that collect and / or maintain data, including privacy data maintained in privacy datasets 102A, 102B, or 102C. In one example, the group of data managers 110A, 110B, and 110C includes a fashion studio that produces clothing, and privacy datasets 102A, 102B, and 102C include images of fashion items produced by the fashion studio. In another example, the group of data managers 110A, 110B, and 110C includes an autonomous vehicle producer, and privacy datasets 102A, 102B, and 102C include images captured by the autonomous vehicle, route data, performance data, object detection data, or other data for which the group of data managers 110A, 110B, and 110C considers privacy.

[0063] In various embodiments, privacy datasets 102A, 102B, and 102C include a variety of different data, such as images, data objects, databases, source code, medical information, personal information, payment information, metadata, user data, personally identifiable information, user activity, network traffic, social media data, or any other information that entities (e.g., data managers 110A, 110B, and 110C) attempt to control or restrict access to. The set of privacy barriers 112A, 112B, and 112C may include logical and physical barriers to prevent the privacy datasets 102A, 102B, and 102C from being exposed. For example, the set of privacy barriers 112A, 112B, and 112C may include private networks, firewalls, physical walls, encryption, or other security measures to prevent and / or restrict access to the privacy datasets 102A, 102B, and 102C.

[0064] Furthermore, in various embodiments, the set of privacy barriers 112A, 112B, and 112C also includes privacy constraints described in more detail below. For example, a privacy constraint refers to considerations and / or protections for protecting user privacy and / or data privacy. In various embodiments, privacy constraints are used in conjunction with hyperparameters of the differential privacy Sinkhorn algorithm, as discussed in detail below. Additionally, in various embodiments, hyperparameters include pruning constants, noise amplitude, batch size, and number of iterations. In such embodiments, combinations of these values ​​(e.g., hyperparameters) are used to satisfy or break privacy constraints. Using pruning, noise, and other hyperparameters to satisfy privacy constraints includes privacy mechanisms. In some embodiments, other privacy mechanisms are used to satisfy privacy constraints, which may include different and / or additional hyperparameters. In one example, a first privacy mechanism limits the number of times privacy datasets 102A, 102B, and 102C can be accessed.

[0065] In various embodiments, the set of differential privacy-generating models 104A, 104B, and 104C includes neural networks or other models trained using methods and systems described in more detail below. In one embodiment, the set of differential privacy-generating models 104A, 104B, and 104C generates training data during inference operations, which includes features of privacy datasets 102A, 102B, and 102C, which enable the training of neural networks (e.g., untrained neural network 108). Figure 1 As shown, the set of differential privacy generative models 104A, 104B, and 104C is provided to data consumer 116. However, in other embodiments, the set of differential privacy generative models 104A, 104B, and 104C is maintained by corresponding data managers 110A, 110B, and 110C, and training data is provided to data consumer 116.

[0066] In various embodiments, data consumer 116 includes any entity that uses one or more differential privacy generative models 104A, 104B, and 104C to train an untrained neural network 108, such as Figure 1 As shown. In one example, data consumer 116 includes entities that include one or more data managers 110A, 110B, and 110C. In yet another example, data consumer 116 is a third party requesting access to one or more privacy datasets 102A, 102B, and / or 102C. In various embodiments, untrained neural network 108 includes any neural network, machine learning algorithm, or other model. For example, untrained neural network 108 includes a model that, when trained, performs data-centric applications such as recommendation systems, personalized medicine, facial recognition, speech synthesis, object detection, autonomous driving, and other applications. In various embodiments, the set of differential privacy-generative models 104A, 104B, and 104C, trained using privacy constraints on privacy data (e.g., privacy datasets 102A, 102B, or 102C), provides data consumer 116 with indirect access to the privacy data.

[0067] In various embodiments, differential privacy provides a definition of privacy that quantifies the amount of information disclosed by users (e.g., data consumer 116 and / or data managers 110A, 110B, and 110C) involved in any data dissemination (e.g., training a generative model or using training data). Furthermore, although three generative models are shown (e.g., differentially private generative models 104A, 104B, and 104C), a single generative model can be iteratively trained by one or more of data managers 110A, 110B, or 110C and provided to data consumer 116.

[0068] Figure 2An example of an environment 200 according to at least one embodiment is shown, in which a data manager 210 generates a differential privacy generative model 204. As described below, in one embodiment, the generative model 204 includes a parametric function having multiple parameters that are tuned to train the generative model 204 to produce a training dataset 206 similar to the privacy dataset 202. In various embodiments, the privacy dataset 202 includes any data whose access is restricted by the data manager 210, for example, by using a privacy barrier 212. The data manager 210 can be any entity that collects, generates, and / or stores the privacy dataset 202. In one example, the data manager 210 includes a computing resource service provider that stores the privacy dataset 202 on behalf of clients and / or third parties. Furthermore, although not explicitly stated... Figure 2 As shown, data manager 210 may include computing resources (e.g., storage devices, servers, processors, memory) for performing the various functions and operations described in the various embodiments. As mentioned above, privacy barrier 212 may include any logical or physical restrictions on access to privacy dataset 202.

[0069] In various embodiments, Represents the sample space (e.g., a representation of the possible outcomes of the generative model 204 included in the training dataset 206). Indicates to All possible measurements, and This represents the potential space. In such an embodiment, the data manager 210 uses independently sampled, finite-sized observation sets. 204 generative models were trained (of which (Representing generative models 204) This makes it have a noise source Its induced distribution Similar to the observed privacy dataset 202 (denoted by ν). Furthermore, in various embodiments, by... The generative model 204 represents a parameterized model. The trainable parameter function.

[0070] Combined with the following text Figure 3 and Figure 4As described, in one embodiment, training the generative model 204 includes employing an optimal transfer generative learning method using backpropagation, which uses gradient descent to minimize the variation in distance between elements of the privacy dataset 202 (e.g., images, data objects, customer profiles, etc. stored in the privacy dataset 202) and the data generated by the generative model 204. In one example, the distance between real data (e.g., privacy data included in the privacy dataset 202) and generated data (e.g., training dataset 206 or other data generated by the generative model 204) is defined by the Wasserstein distance. In one embodiment, the distance between two distributions (e.g., the privacy dataset 202 and the data generated by the generative model 204) is measured as the expectation of a pointwise cost function between sample pairs distributed according to the optimal transfer plan.

[0071] In various embodiments, the cost function is defined as and Furthermore, the entropy-regularized Wasserstein distance (ERWD) is defined as:

[0072]

[0073] in .

[0074] Furthermore, in various embodiments, such as combining Figure 3 and Figure 4 In the described embodiments, the Sinkhorn divergence is used to offset the entropy bias introduced by ERWD. In one embodiment, the Sinkhorn divergence between two measurements μ and ν is defined as:

[0075]

[0076] Once the Sinkhorn divergence is calculated, in one embodiment, gradient computation is performed by obtaining the Jacobian vector product (e.g., dot product) between the cost matrix (e.g., Jacobian) and the transport weights (e.g., vectors). In one example, X has M data points (e.g., generated data 308), Y has N data points (e.g., real data 312 sampled from a privacy dataset), and the cost matrix 310 is an M×N matrix composed of the element-wise costs of X and Y. Furthermore, during gradient descent, in various embodiments, to preserve differential privacy, the parametric gradients are clipped with a constant and Gaussian noise is added. In various embodiments, differential privacy measures the degree to which the output of the generative model 204 can deviate between adjacent input datasets—each dataset differing by one entry. In one example, differential privacy is defined as a stochastic mechanism. It has a domain and Range, satisfying ( α , ϵ Differential privacy, if for any adjacent d , d' ∈ D It believes

[0077]

[0078] in yes α The divergence of order, for any condition satisfying... M Also satisfies .

[0079] In various embodiments, the generative model 204 is trained using various privacy constraints (e.g., in addition to or as alternatives to pruning and noise). In one example, the number of iterations of the training method (e.g., differential privacy Sinkhorn iterations) is finite. In various embodiments, once the generative model 204 is trained, it is used to generate a training dataset 206 and / or provided to a data consumer 216 so that the data consumer 216 can generate the training dataset 206. Furthermore, in at least one embodiment, the training dataset 206 is used to train an untrained neural network 208 to generate a trained neural network 214.

[0080] refer to Figure 3 , Figure 3 These are examples of forward iterations of a training method using the differential privacy Sinkhorn algorithm to train parameters of a generative model according to some embodiments of this disclosure. It should be understood that such and other arrangements described herein are presented only as examples. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, functional groups, etc.) may be used in addition to or instead of those shown, and some elements may be omitted entirely. Furthermore, many of the elements described herein are functional entities that can be implemented as discrete or distributed components or in combination with other components, and implemented in any suitable combination and location. The various functions described herein as being performed by entities can be performed by hardware, firmware, and / or software. For example, various functions can be performed by a processor executing instructions stored in memory.

[0081] Now for reference Figure 3-7Each box of methods 300, 400, 500, 600, and 700 described herein includes a computational process that can be executed using any combination of hardware, firmware, and / or software. For example, various functions can be executed by a processor that executes instructions stored in memory. The method can also be embodied as computer-executable instructions stored on a computer storage medium. These methods can be provided by standalone applications, services, or managed services (standalone or in combination with another managed service) or plug-ins to another product, to name a few. Furthermore, by way of example, regarding... Figure 2 Data manager 210 describes method 300. However, these methods may be performed additionally or alternatively by any system or any combination of systems, including but not limited to those described herein. Furthermore, the operations of these methods may be performed in a different order, in parallel or serially, or omitted entirely.

[0082] Figure 3 This is a flowchart illustrating a method 300 for performing a forward pass of a differential privacy Sinkhorn algorithm according to some embodiments of the present disclosure. In various embodiments, sensitive data, including data from a privacy dataset 302, is sampled and used as training data. In such embodiments, this training data (e.g., sensitive data sampled from privacy dataset 302) is combined with non-sensitive data generated by a generative model 304 in a cost matrix 310. The backward pass of the differential privacy Sinkhorn algorithm (described below in conjunction with...) Figure 4 In a more detailed description, a privacy barrier is added by cropping the gradient of the generated data and adding noise.

[0083] In various embodiments, the Sinkhorn loss between each batch of real data 312 (e.g., data obtained from privacy dataset 302) and the generated data 308 (e.g., data generated by generative model 304 based at least in part on latent value 306) is used as an objective (e.g., the direction in which the parameters of generative model 304 are moved). In one embodiment, the Sinkhorn loss computed on batches N of generated data 308 and real data 312 M is defined as:

[0084]

[0085] Where X is the uniformly sampled batch of the generated data 308, and Y is the uniformly sampled batch of the real data 312. For the two samples... and , It is the cost matrix between A and B. It is an approximate optimal transmission plan that minimizes the ERWD calculated on A and B.

[0086] In various embodiments, Figure 3 and Figure 4 The training methods described herein are performed by biasing the optimal transport position to couple data of the same class together under class conditions. For example, if privacy dataset 302 includes images of objects with different classes (e.g., cats and dogs), methods 300 and 400 can be performed using class conditioning. In embodiments that include classes, generative model 304 generates generated data 308 within a given class by including class information in latent values ​​306 at least. In other embodiments that include classes, generative model 304 generates generated data 308 within a given class by using at least one layer of class-conditioned batching. In various embodiments, class-conditioned data includes images with class labels. In one example, the dataset consists of images and class labels for the images. In other examples, the generator (e.g., a model trained using a differential privacy Sinkhorn algorithm) produces categorical data and / or continuous non-image data. Furthermore, in various embodiments that include classes, the Sinkhorn loss 316 is modified at least by increasing the cost between data objects of different classes, thereby shifting the weights of the transport plan to class-matched data objects. Furthermore, in various embodiments, a scaling constant is used. This is used to determine the importance of category similarity relative to data objects (e.g., real data 312 and generated data 308). In one example, when used to determine the Sinkhorn loss 316, the category conditional value is given as follows:

[0087]

[0088] In various embodiments, uniformly sampled class labels are used for generated data 308, while true class labels are used for true data 312. In the equation above, let... and Represents the category labels for x and y, and and This indicates concatenation. Furthermore, in the equation above, onehot() provides a one-to-k encoding of the categorical variables used with generative model 304. In various embodiments, Used to calculate Sinkhorn loss 316 (represented as) And the optimal transmission distance is the cost of transmitting the generated data 308 and the joint distribution of category labels to the distribution of real data 312 and real category labels.

[0089] Methods 300 and 400 describe examples of the differential privacy Sinkhorn algorithm, where L is the number of classes, X is the sample space, M is the size of the privacy dataset 302, and backpropagation is the inverse pattern automatic differentiation function of the Jacobian vector product computed with 'out', 'in', and 'gradient weights' as inputs. In one embodiment, the differential privacy Sinkhorn algorithm is defined by the following pseudocode:

[0090] Input: Privacy dataset 302 Sampling rate q, noise scale σ, clipping factor Δ, generator Learning rate α Entropy regularization ϵ , and the total number of steps T.

[0091] Output:

[0092] arrive

[0093] sample ,

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101] Finish

[0102] In one embodiment, The algorithm is defined by the following pseudocode:

[0103] enter: Sampling rate

[0104] Output: ,

[0105]

[0106]

[0107] In one embodiment, Sinkhorn Algorithm Defined by the following pseudocode:

[0108] enter:

[0109] Output:

[0110] Perform when convergence is not achieved

[0111]

[0112]

[0113] exist End of time

[0114] In one embodiment, method 300 includes obtaining a latent value at block 306. In various embodiments, as described above, the latent value is defined as Z. For example, this latent value may be provided as input to a generative model 304 used to generate the generated data 308. In embodiments that include category conditions, the value defined above... It is also provided as input to generative model 304 to generate generated data 308 with category conditions. In various embodiments, generative model 304 is a parameterized function that generates data simulating a privacy dataset. Furthermore, in embodiments utilizing the above algorithm, the output of generative model 304 is defined as generated data 308 of X, where... .

[0115] In various embodiments, the generated data 308 is combined with the real data 312 to generate the cost matrix 310. As described in the algorithm above, in one embodiment, the real data 312 is obtained from samples Y of the privacy dataset 302. In one embodiment, samples Y of the privacy dataset 302 are Poisson samples of the privacy dataset 302. In other embodiments, the real data 312 is obtained using any random, pseudo-random, or other sampling of the privacy dataset 302. Furthermore, the real data 312 may include a subset of the privacy dataset 302.

[0116] In one embodiment, for two samples (e.g., generated data 308 and real data 312)... and , It is the cost matrix between A and B, and It is an approximately optimal transmission plan that minimizes the ERWD calculated on A and B. In such an embodiment, Is using The discretized version of the following dual formula is used for iterative dual potential calculation:

[0117]

[0118] Where ⊗ denotes the product measure and ⊕ denotes the outer sum, the output is a matrix of element-wise sums of each vector. In one example, the optimal transport plan... With dual forces Relevant. Using this equation, in one embodiment, optimal f and g Used to obtain through primal-dual relations Furthermore, in such an embodiment, f and g The first-order optimality condition is provided by the following theorem:

[0119]

[0120]

[0121] In various embodiments, in order to optimize f and g The Sinkhorn algorithm described above was applied. For example... Figure 3 As shown, in various embodiments, a single iteration of the Sinkhorn algorithm is Sinkhorn iteration 314. For example, Sinkhorn iteration 314 includes determining an optimal transport plan for the generated data 308 and the real data 312, at least in part based on the cost matrix 310.

[0122] In various embodiments, the Sinkhorn loss 316 is defined by the following equation:

[0123]

[0124] Referring to the Sinkhorn algorithm described above, in one embodiment, the Sinkhorn loss 316 is defined as... In such an embodiment, The term is omitted because it depends only on the privacy data and there are no parameters in the generative model that would affect the privacy data. Furthermore, in one embodiment, the gradient of the parameters... This represents a set of changes to the parameters of the generative model 304. In various embodiments, this set of changes (e.g., gradients) is modified by a value α to control parameters that might be prevented from converging due to overcorrection (e.g., preventing the generated data from matching privacy data more closely). In various embodiments, the gradient (e.g., ...) is adjusted according to privacy constraints. and / or This allows for noise reduction, clipping, or other modifications. For example, noise can be added to the gradient. This is to prevent the exposure of private datasets.

[0125] In various embodiments, other variations of the loss function are used in conjunction with methods 300 and 400. In one example, when the Sinkhorn loss is computed using a sample-based estimate, a semi-debiased Sinkhorn loss function provides control over the trade-off between bias and variance (e.g., at least by computed an estimate of the “precise” Sinkhorn loss using a finite number of samples, where the estimate has some bias and variance that affects model training). In such embodiments, this indirectly affects the samples generated by the generator. In one embodiment, a semi-debiased Sinkhorn loss is computed on batches of real and generated data (e.g., images) to train the generative model 304. As discussed in more detail below, in one embodiment, privacy protection is enforced during backpropagation by cropping and adding noise to the gradients of the generated data. In various embodiments, for a mixture of scores p∈[0,1] and a natural number n, Given generated samples and m real samples The semi-debiased Sinkhorn loss is defined by the following equation:

[0126]

[0127] in Let X represent the continuous samples from a to b-1. Referring to the semi-biased Sinkhorn loss equation, in one embodiment, when p=1, the semi-biased Sinkhorn loss equation is equal to the fully biased Sinkhorn loss (e.g., (with separate batches of generated data X and X'), while when p=0, the semi-biased Sinkhorn loss equation reverts to the fully biased Sinkhorn loss (e.g., (There is no second separate batch of generated data X'). As shown in the algorithm below, a semi-debiased Sinkhorn loss is computed, while additional gradients are modified by adding noise and pruning to achieve the chosen differential privacy mechanism:

[0128] Input: Privacy dataset Sampling rate q, noise scale σ, clipping factor Δ, generator Learning rate α, entropy regularization ϵ, de-resampled score

[0129] Debiased resampling score

[0130] Eliminate DC bias fractional sampling

[0131] Remove the DC bias fraction and re-sampling

[0132] Unable to load all results

[0133] Retry

[0134] Retrying...

[0135] Retrying...

[0136] Score p, total number of steps T.

[0137] Output:

[0138]

[0139] conduct

[0140] sample ,

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149] Finish

[0150] Figure 4 An example of the reverse pass of the Sinkhorn algorithm according to an embodiment is shown. (Refer to...) Figure 4In one embodiment, backpropagation generates gradient descent of dataset 408 to adjust the parameters of generative model 404. In various embodiments, generative model 404 is trained at least in part based on Differentially Private Stochastic Gradient Descent descent. In one embodiment, generative model 404 includes any generative model as described above, such as generative model 304. In various embodiments, the gradient of generated data 408 is computed at least in part based on Sinkhorn loss 416. In one example, the gradient of generated data 408 is defined as... ,in Define the Sinkhorn loss as 416.

[0151] In various embodiments, the gradient of generated data 408 defines modifications to the parameters of generative model 404 such that generative model 404 generates data that matches the privacy data. Furthermore, in one embodiment, the gradient is computed by taking the Jacobian vector product between cost matrix 410 and a set of transport weights generated by Sinkhorn iteration 414. In the example, cost matrix 410 and Sinkhorn iteration 414 include cost matrix 310 and Sinkhorn iteration 314, as described above. In one example, the gradient defines modifications to the pixels of the image included in generated data 408 to match the image included in the privacy dataset.

[0152] In various embodiments, pruning and / or adding noise 412 is added to the gradient to protect the privacy of the dataset. Pruning and / or adding noise 412 can create a privacy barrier to prevent the exposure of private data. In one embodiment, instead of adding pruning and / or adding noise 412 to the gradient of the parameter θ (e.g., the parameters of the generative model 404), pruning and / or adding noise 412 is added to a gradient defined as... The gradient of the generated data 408 (e.g., the gradient of the Sinkhorn loss 416) is then backpropagated to the parameters θ. In various embodiments, this preserves privacy because the inputs to the generative model 404 (e.g., latent codes) are randomly sampled and independent of the data, leaving a unique connection to the privacy dataset at the output. In one embodiment, in each Sinkhorn iteration 414, gradient descent is performed by backpropagating the gradients. To update the parameters. Furthermore, in such an embodiment, clipping and / or adding noise 412 is applied to each item, making... Therefore, the sensitivity is By adding Gaussian noise with a scale of 2Cσ, this mechanism satisfies differential privacy. .

[0153] For given θ and loss function Parameterized parameter functions (For example, generative model 404), in one embodiment, gradients are used. Update parameter θ. In various embodiments, when it is determined that clipping and / or adding noise 412 is under gradient perturbation, the gradient is first adjusted by a constant Δ. Cut to norm, and then from Sample noise and add it.

[0154] Figure 5 This is a flowchart illustrating a method 500 for training a model, at least in part, based on data obtained from a generative model trained using a privacy dataset, according to some embodiments of the present disclosure. In block B502, method 500 includes providing an untrained generative model to a data manager. In various embodiments, the data manager maintains a privacy dataset protected by privacy barriers as described above. In other embodiments, the generative model is trained partially or fully. In one example, the generative model is partially trained by another data manager before being provided to the data manager. In one embodiment, providing the generative model to the data manager includes providing executable code to the data manager that takes privacy data from a privacy dataset maintained by the data manager as input.

[0155] In box B504, method 500 includes training a generative model at least partially based on privacy data and one or more privacy constraints. In one embodiment, the generative model is trained using a differential privacy Sinkhorn algorithm. For example, the generative model is a parameterized function, and the generative model is trained by at least determining an optimal transport schedule applied to parameters with backpropagation gradient descent. In various embodiments, the privacy constraint indicates the number of training iterations to be performed. In one example, the privacy constraint indicates that ten training iterations may be performed if convergence is not reached. In yet another example, the privacy constraint indicates the amount of pruning and / or noise to be added during training.

[0156] Once the model is trained, method 500 at box B506 includes obtaining training data from the trained generative model. In one example, once the model is trained, a data manager can distribute the training data generated by the trained generative model, which simulates a privacy dataset maintained by the data manager. In other embodiments, the data manager can provide access to the trained generative model. According to one embodiment, method 500 at box B508 includes training the model at least in part based on the training data generated by the generative model.

[0157] Figure 6This is a flowchart illustrating a method 600 for training a generative model trained on a privacy dataset using a forward pass of a differential privacy Sinkhorn algorithm, according to some embodiments of the present disclosure. In block B602, method 600 includes providing a latent code as input to the generative model. In various embodiments, the latent code comprises random or pseudo-random values ​​provided as input to the generative model, which uses this input to generate a batch of data. For example, the latent code is extracted from a Gaussian distribution. In block B604, method 600 includes obtaining a batch of data from the generative model based at least partially on the latent code. In one example, the generative model generates a set of images (e.g., facial images, clothing items, etc.) based at least partially on the latent code. In one embodiment, at block B606, method 600 includes obtaining samples of privacy data. In one example, random or pseudo-random samples of privacy data are obtained from a privacy dataset maintained by a data administrator. In one embodiment, the sampling includes Poisson samples.

[0158] In block B608, method 600 includes determining a cost matrix based at least in part on samples of the data batch and privacy data. In various embodiments, the cost matrix indicates the element-wise cost between pairs of elements in the generated data batch and the privacy data. In block B610, method 600 includes determining an optimal transport plan based at least in part on the cost matrix. In various embodiments, the optimal transport plan is determined during Sinkhorn iterations. Furthermore, in one embodiment, optimal transport is obtained by minimizing the distance (e.g., the distance between samples of the data batch (e.g., data generated by a generative model) and the privacy data (e.g., the Wasserstein distance). In one example, Sinkhorn iterations compute the optimal transport plan such that the product of the optimal transport plan and the cost matrix is ​​minimized. In one embodiment, the optimal transport plan describes the weights of the cost matrix. Figure 6 As shown, method 600 combines Figure 7 The method described continues at box B702 of method 700.

[0159] Figure 7 This is a flowchart illustrating a method 700 for training a generative model trained on a privacy dataset based on backpropagation of a differential privacy Sinkhorn algorithm, according to some embodiments of the present disclosure. In block B702, method 700 includes determining the gradient of the Sinkhorn loss. In various embodiments, the Sinkhorn loss is a measure of how well the generated data matches the privacy data and / or real data. In one example, the Sinkhorn loss is a weighted average of the cost matrix multiplied by the optimal transport schedule. In one embodiment, the Sinkhorn loss is determined at least in part based on the equation defined above.

[0160] In one embodiment, at box B705, method 700 includes pruning and / or adding noise to the gradient of the Sinkhorn loss. In various embodiments, data privacy is maintained by modifying the gradient of the Sinkhorn loss. Furthermore, in one embodiment, the amount of pruning and / or noise added is determined at least in part based on privacy constraints. In one example, where privacy of a privacy-sensitive dataset maintained by a data administrator is important, the amount of pruning and / or noise added to the gradient can be increased to protect data privacy. At box B706, method 700 includes backpropagating the gradient of the Sinkhorn loss to a batch of data. In various embodiments, the gradient of the batch of data is determined by multiplying at least the gradient of the cost matrix by an optimal transport plan.

[0161] In various embodiments, at block B708, method 700 includes updating the parameters of the generative model at least in part based on backpropagation of the Sinkhorn loss to a batch of data. For example, the result of backpropagation using gradient descent (after pruning and adding noise) is the gradient of the parameters of the generative model, which indicates the modification of the parameter values ​​so that the data produced by the generative model more closely matches the real data. In various embodiments, adjusting the parameters includes adding or subtracting a factor of the gradient from the parameters. For example, α A value of 1 might prevent the parameters from converging due to taking excessively large step sizes. In various embodiments, the gradient of the parameters is also optimized for machine learning using one or more functions. In such an embodiment, updating the parameters is defined as: , where Optimize() defines the gradient optimization function. Furthermore, in one embodiment, the selection... α The value is used to match the magnitude of the optimization function.

[0162] In box B710, method 700 includes determining whether convergence has been achieved. In various embodiments, convergence is achieved if the data generated by the generative model matches real data (e.g., privacy data). In one example, the gradient values ​​of the parameters are zero or within a zero threshold. In another example, convergence is determined at least in part based on the cost matrix, the Sinkhorn loss, and / or the gradients of these values. Figure 7 As shown, if convergence is achieved, the training of the generative model is complete, and the generative model is provided in box B712. In one example, the generative model is provided to the data consumer as described above.

[0163] If convergence is not achieved (e.g., training of the generative model continues), then in box B714, method 700 determines whether the privacy constraint has been reached. In various embodiments, the training algorithm (e.g., in conjunction with the above) Figure 3 and Figure 4The Sinkhorn algorithm described has an iteration limit. For example, the number of iterations (e.g., the number of times privacy data is sampled) is limited to ten. In various other embodiments, other privacy constraints are considered in addition to or as an alternative to the number of iterations, such as the degree to which the generated data matches the real data, the amount of pruning and / or noise added, the size of the privacy dataset, the size of the sampled privacy dataset, or other considerations. Return to Figure 7 If the privacy constraint is met, method 700 returns to box B712 and the training of the generative model is complete. However, if the privacy constraint is not met, method 700 returns to the previous step. Figure 6 Box B602 of the described method 600. In various embodiments, training of the generative model continues with another iteration of the Sinkhorn algorithm described above.

[0164] Reasoning and training logic

[0165] Figure 8A Inference and / or training logic 815 is shown for performing inference and / or training operations associated with one or more embodiments. The following is in conjunction with... Figure 8A and / or Figure 8B Provide details about reasoning and / or training logic 815.

[0166] In at least one embodiment, inference and / or training logic 815 may include, but is not limited to, code and / or data storage 801 for storing forward and / or output weights and / or input / output data, and / or other parameters configuring neurons or layers of a neural network trained for and / or used for inference in one or more embodiments. In at least one embodiment, training logic 815 may include or be coupled to code and / or data storage 801 for storing graph code or other software to control timing and / or sequence, wherein weight and / or other parameter information is loaded to configure logic, including integer and / or floating-point units (collectively, arithmetic logic units (ALUs)). In at least one embodiment, code (such as graph code) loads weight or other parameter information into the processor ALU based on the architecture of the neural network to which the code corresponds. In at least one embodiment, code and / or data storage 801 stores weight parameters and / or input / output data of each layer of a neural network trained or used in one or more embodiments during forward propagation of input / output data and / or weight parameters during training and / or inference using one or more embodiments. In at least one embodiment, any portion of the code and / or data storage 801 may be included within other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory.

[0167] In at least one embodiment, any portion of the code and / or data storage 801 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, the code and / or data storage 801 may be a cache memory, dynamic random-addressable memory (“DRAM”), static random-addressable memory (“SRAM”), non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice of whether the code and / or data storage 801 is internal or external to the processor, for example, or composed of DRAM, SRAM, flash memory, or some other storage type, may depend on the available on-chip or off-chip storage space, the latency requirements of the training and / or inference functions being performed, the batch size of the data used in the inference and / or training of the neural network, or some combination of these factors.

[0168] In at least one embodiment, the inference and / or training logic 815 may include, but is not limited to, code and / or data storage 805 for storing backpropagation and / or output weights and / or input / output data neural networks corresponding to neurons or layers of a neural network trained and / or used for inference in one or more embodiments. In at least one embodiment, during training and / or inference using one or more embodiments, the code and / or data storage 805 stores weight parameters and / or input / output data for each layer of a neural network trained or used in one or more embodiments during backpropagation of input / output data and / or weight parameters. In at least one embodiment, the training logic 815 may include or be coupled to code and / or data storage 805 for storing graph code or other software to control timing and / or sequence, wherein weight and / or other parameter information is loaded to configure logic including integer and / or floating-point units (collectively, arithmetic logic units (ALUs)).

[0169] In at least one embodiment, code (such as graph code) causes the architecture of the neural network corresponding to that code to load weights or other parameter information into the processor ALU. In at least one embodiment, any portion of the code and / or data storage 805 may be included together with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. In at least one embodiment, any portion of the code and / or data storage 805 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, the code and / or data storage 805 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice between the code and / or data storage 805 being internal or external to the processor, for example, whether it consists of DRAM, SRAM, flash memory, or some other type of storage, depends on whether the available storage is on-chip or off-chip, the latency requirements of the training and / or inference functions being performed, the data batch size used in the inference and / or training of the neural network, or some combination of these factors.

[0170] In at least one embodiment, code and / or data storage 801 and code and / or data storage 805 may be separate storage structures. In at least one embodiment, code and / or data storage 801 and code and / or data storage 805 may be the same storage structure. In at least one embodiment, code and / or data storage 801 and code and / or data storage 805 may be partially combined and partially separated. In at least one embodiment, any portion of code and / or data storage 801 and code and / or data storage 805 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory.

[0171] In at least one embodiment, the inference and / or training logic 815 may include, but is not limited to, one or more arithmetic logic units (“ALUs”) 810 (including integer and / or floating-point units) for performing logical and / or mathematical operations at least in part based on or instructed by training and / or inference code (e.g., graph code), the results of which may produce activations (e.g., output values ​​from layers or neurons within a neural network) stored in activation storage 820, which are functions of input / output and / or weight parameter data stored in code and / or data storage 801 and / or code and / or data storage 805. In at least one embodiment, activation is activated in response to execution instructions or other code, linear algebraic and / or matrix-based mathematical generation performed by ALU 810, and the activation is stored in activation storage 820, wherein weight values ​​stored in code and / or data storage 805 and / or code and / or data storage 801 are used as operands with other values, such as bias values, gradient information, momentum values, or other parameters or hyperparameters, and any or all of these can be stored in code and / or data storage 805 or code and / or data storage 801 or other on-chip or off-chip storage.

[0172] In at least one embodiment, one or more processors or other hardware logic devices or circuits include one or more ALUs 810, while in another embodiment, one or more ALUs 810 may be located outside the processor or other hardware logic device or the circuitry that uses them (e.g., a coprocessor). In at least one embodiment, one or more ALUs 810 may be included within an execution unit of a processor, or otherwise included in a group of ALUs accessible by the execution unit of the processor, which may be within the same processor or distributed among different processors of different types (e.g., a central processing unit, a graphics processing unit, a fixed-function unit, etc.). In at least one embodiment, code and / or data storage 801, code and / or data storage 805, and activation storage 820 may share a processor or other hardware logic device or circuitry, while in another embodiment, they may be located in different processors or other hardware logic devices or circuitry, or in some combination of the same and different processors or other hardware logic devices or circuitry. In at least one embodiment, any portion of activation storage 820 may be included together with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. Furthermore, inference and / or training code may be stored together with other code accessible to the processor or other hardware logic or circuitry, and may be retrieved and / or processed using the processor’s fetch, decode, schedule, execute, exit, and / or other logic circuitry.

[0173] In at least one embodiment, the active memory 820 may be a cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other memory. In at least one embodiment, the active memory 820 may be wholly or partially located inside or outside one or more processors or other logic circuits. In at least one embodiment, the choice of whether the active memory 820 is internal to or external to the processor may depend on the availability of on-chip or off-chip storage, the latency requirements for training and / or inference functions, the batch size of data used in inference and / or training the neural network, or some combination of these factors. For example, it may include DRAM, SRAM, flash memory, or other memory types.

[0174] In at least one embodiment, Figure 8A The inference and / or training logic 815 shown can be used in conjunction with an application-specific integrated circuit (“ASIC”), such as a TensorFlow® processing unit from Google, an inference processing unit (IPU) from Graphcore™, or a Nervana® (e.g., “Lake Crest”) processor from Intel Corp. In at least one embodiment, Figure 8A The inference and / or training logic 815 shown can be used in conjunction with central processing unit (“CPU”) hardware, graphics processing unit (“GPU”) hardware, or other hardware such as field programmable gate array (“FPGA”).

[0175] Figure 8B An inference and / or training logic 815 according to at least one embodiment is illustrated. In at least one embodiment, the inference and / or training logic 815 may include, but is not limited to, hardware logic, wherein computational resources are dedicated or otherwise uniquely used in conjunction with weight values ​​or other information corresponding to one or more layers of neurons within a neural network. In at least one embodiment, Figure 8B The inference and / or training logic 815 shown can be used in conjunction with an application-specific integrated circuit (ASIC), such as a TensorFlow® processing unit from Google, an inference processing unit (IPU) from Graphcore™, or a Nervana® (e.g., “Lake Crest”) processor from Intel Corp. In at least one embodiment, Figure 8BThe inference and / or training logic 815 shown can be used in conjunction with central processing unit (CPU) hardware, graphics processing unit (GPU) hardware, or other hardware (e.g., field-programmable gate array (FPGA)). In at least one embodiment, the inference and / or training logic 815 includes, but is not limited to, code and / or data storage 801 and code and / or data storage 805, which can be used to store code (e.g., graph code), weight values, and / or other information, including bias values, gradient information, momentum values, and / or other parameter or hyperparameter information. Figure 8B In at least one embodiment shown, each of code and / or data storage 801 and code and / or data storage 805 is associated with dedicated computing resources (e.g., computing hardware 802 and computing hardware 806), respectively. In at least one embodiment, each of computing hardware 802 and computing hardware 806 includes one or more ALUs that perform mathematical functions (e.g., linear algebraic functions) only on the information stored in code and / or data storage 801 and code and / or data storage 805, respectively, and the results of the function execution are stored in activation memory 820.

[0176] In at least one embodiment, each of the code and / or data storage 801 and 805 and the corresponding computing hardware 802 and 806 corresponds to a different layer of the neural network, such that activation obtained from one “store / computation pair 801 / 802” of the code and / or data storage 801 and computing hardware 802 provides input as input to the next “store / computation pair 805 / 806” of the code and / or data storage 805 and computing hardware 806, in order to reflect the conceptual organization of the neural network. In at least one embodiment, each store / computation pair 801 / 802 and 805 / 806 may correspond to more than one neural network layer. In at least one embodiment, additional store / computation pairs (not shown) may be included in the inference and / or training logic 815 after or in parallel with the store / computation pairs 801 / 802 and 805 / 806.

[0177] Neural network training and deployment

[0178] Figure 9Training and deployment of a deep neural network according to at least one embodiment are illustrated. In at least one embodiment, an untrained neural network 906 is trained using a training dataset 902. In at least one embodiment, the training framework 904 is the PyTorch framework, while in other embodiments, the training framework 904 is TensorFlow, Boost, Caffe, Microsoft Cognitive Toolkit / CNTK, MXNet, Chainer, Keras, Deeplearning4j, or other training frameworks. In at least one embodiment, the training framework 904 trains the untrained neural network 906 and enables it to be trained using the processing resources described herein to generate a trained neural network 908. In at least one embodiment, the weights may be randomly selected or pre-trained using a deep belief network. In at least one embodiment, training may be performed in a supervised, partially supervised, or unsupervised manner.

[0179] In at least one embodiment, supervised learning is used to train an untrained neural network 906, wherein training dataset 902 includes inputs paired with desired outputs for input, or wherein training dataset 902 includes inputs with known outputs and neural network 906 is manually graded output. In at least one embodiment, the untrained neural network 906 is trained in a supervised manner, and inputs from training dataset 902 are processed, and the resulting output is compared with a set of expected or desired outputs. In at least one embodiment, errors are then propagated back through the untrained neural network 906. In at least one embodiment, training framework 904 adjusts the weights controlling the untrained neural network 906. In at least one embodiment, training framework 904 includes tools for monitoring the degree to which the untrained neural network 906 converges to a model (e.g., a trained neural network 908) adapted to generate the correct answer (e.g., result 914) based on input data (e.g., a new dataset 912). In at least one embodiment, training framework 904 repeatedly trains the untrained neural network 906 while adjusting the weights to improve the output of the untrained neural network 906 using a loss function and adjustment algorithms (e.g., stochastic gradient descent). In at least one embodiment, the training framework 904 trains an untrained neural network 906 until the untrained neural network 906 reaches the desired accuracy. In at least one embodiment, the trained neural network 908 can then be deployed to perform any number of machine learning operations.

[0180] In at least one embodiment, unsupervised learning is used to train an untrained neural network 906, wherein the untrained neural network 906 attempts to train itself using unlabeled data. In at least one embodiment, the unsupervised learning training dataset 902 will include input data without any associated output data or "ground truth" data. In at least one embodiment, the untrained neural network 906 can learn groupings within the training dataset 902 and can determine how each input relates to the untrained dataset 902. In at least one embodiment, unsupervised training can be used to generate a self-organizing graph in a trained neural network 908, which is capable of performing operations useful for reducing the dimensionality of the new dataset 912. In at least one embodiment, unsupervised training can also be used to perform anomaly detection, which allows the identification of data points in the new dataset 912 that deviate from the normal patterns of the new dataset 912.

[0181] In at least one embodiment, semi-supervised learning can be used, a technique in which a mixture of labeled and unlabeled data is included in the training dataset 902. In at least one embodiment, the training framework 904 can be used to perform incremental learning, for example, through transfer learning techniques. In at least one embodiment, incremental learning enables the trained neural network 908 to adapt to a new dataset 912 without forgetting the knowledge injected into the trained neural network 908 during initial training.

[0182] Data Center

[0183] Figure 10 An example data center 1000 that can be used in at least one embodiment is shown. In at least one embodiment, the data center 1000 includes a data center infrastructure layer 1010, a framework layer 1020, a software layer 1030, and an application layer 1040.

[0184] In at least one embodiment, such as Figure 10As shown, the data center infrastructure layer 1010 may include a resource coordinator 1012, packet computing resources 1014, and node computing resources (“nodes CR”) 1016(1)-1016(N), where “N” represents a positive integer (which may be an integer “N” different from the integers used in other diagrams). In at least one embodiment, nodes CR 1016(1)-1016(N) may include, but are not limited to, any number of central processing units (“CPUs”) or other processors (including accelerators, field-programmable gate arrays (FPGAs), graphics processors, etc.), memory storage devices 1018(1)-1018(N) (e.g., dynamic read-only memory, solid-state drives, or disk drives), network input / output (“NW I / O”) devices, network switches, virtual machines (“VMs”), power modules, and cooling modules, etc. In at least one embodiment, one or more of nodes CR 1016(1)-1016(N) may be servers having one or more of the aforementioned computing resources.

[0185] In at least one embodiment, the grouped computing resource 1014 may include individual groups (not shown) of node CRs housed within one or more racks, or a plurality of racks (also not shown) housed within data centers in various geographical locations. In at least one embodiment, the individual groups of node CRs within the grouped computing resource 1014 may include computing, networking, memory, or storage resources that can be configured or allocated to support groups of one or more workloads. In at least one embodiment, several node CRs, including CPUs or processors, may be grouped within one or more racks to provide computing resources to support one or more workloads. In at least one embodiment, the one or more racks may also include any number of power modules, cooling modules, and network switches, in any combination.

[0186] In at least one embodiment, resource coordinator 1012 may be configured or otherwise control one or more nodes CR1016(1)-1016(N) and / or grouped computing resources 1014. In at least one embodiment, resource coordinator 1012 may include a Software Design Infrastructure (“SDI”) management entity for data center 1000. In at least one embodiment, resource coordinator 1012 may include hardware, software, or some combination thereof.

[0187] In at least one embodiment, such as Figure 10As shown, framework layer 1020 includes a job scheduler 1022, a configuration manager 1024, a resource manager 1026, and a distributed file system 1028. In at least one embodiment, framework layer 1020 may include a framework of software 1032 supporting software layer 1030 and / or one or more applications 1042 supporting application layer 1040. In at least one embodiment, software 1032 or application 1042 may respectively include web-based service software or applications, such as services or applications provided by Amazon Web Services, Google Cloud, and Microsoft Azure. In at least one embodiment, framework layer 1020 may be, but is not limited to, a free and open-source software web application framework, such as Apache Spark, which can utilize distributed file system 1028 for large-scale data processing (e.g., "big data"). TM (Hereinafter referred to as "Spark"). In at least one embodiment, the job scheduler 1022 may include a Spark driver to facilitate the scheduling of workloads supported by various layers of the data center 1000. In at least one embodiment, the configuration manager 1024 may be able to configure different layers, such as the software layer 1030 and the framework layer 1020, which includes Spark and a distributed file system 1028 for supporting large-scale data processing. In at least one embodiment, the resource manager 1026 is able to manage cluster or group computing resources mapped to or allocated to support the distributed file system 1028 and the job scheduler 1022. In at least one embodiment, the cluster or group computing resources may include group computing resources 1014 on the data center infrastructure layer 1010. In at least one embodiment, the resource manager 1026 may coordinate with the resource coordinator 1012 to manage these mapped or allocated computing resources.

[0188] In at least one embodiment, the software 1032 included in the software layer 1030 may include software used by at least a portion of the nodes CR1016(1)-1016(N), the grouped computing resources 1014, and / or the distributed file system 1028 of the framework layer 1020. In at least one embodiment, one or more types of software may include, but are not limited to, Internet web page search software, email virus scanning software, database software, and streaming video content software.

[0189] In at least one embodiment, one or more applications 1042 included in application layer 1040 may include one or more types of applications used by at least a portion of nodes CR 1016(1)-1016(N), grouped computing resources 1014, and / or the distributed file system 1028 of framework layer 1020. In at least one embodiment, one or more types of applications may include, but are not limited to, any number of genomics applications, cognitive computing, applications, and machine learning applications, including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), or other machine learning applications used in conjunction with one or more embodiments.

[0190] In at least one embodiment, any of the configuration manager 1024, resource manager 1026, and resource coordinator 1012 can perform any number and type of self-modification actions based on any amount and type of data acquired in any technically feasible manner. In at least one embodiment, self-modification actions can mitigate potentially poor configuration decisions by data center operators of data center 1000 and can prevent underutilization and / or poor performance of the data center.

[0191] In at least one embodiment, data center 1000 may include tools, services, software, or other resources to train one or more machine learning models or to use one or more machine learning models to predict or infer information according to one or more embodiments described herein. For example, in at least one embodiment, a machine learning model can be trained by calculating weight parameters based on a neural network architecture using the software and computing resources described above with respect to data center 1000. In at least one embodiment, information can be inferred or predicted using trained machine learning models corresponding to one or more neural networks using the resources described above with respect to data center 1000, by using weight parameters calculated through one or more training techniques described herein.

[0192] In at least one embodiment, the data center may use a CPU, application-specific integrated circuit (ASIC), GPU, FPGA, or other hardware to utilize the aforementioned resources to perform training and / or inference. Furthermore, one or more of the aforementioned software and / or hardware resources may be configured as a service to allow a user to train or perform information inference, such as image recognition, speech recognition, or other artificial intelligence services.

[0193] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8BDetails are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be implemented in the system. Figure 10 Used in this context for inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0194] In various embodiments, inference and / or training logic 815 is used to train the generative model as described above. For example, data center 1000 maintains a privacy dataset in memory and uses tools, services, software, or other resources to train one or more generative models to generate training data to simulate the privacy data as described above. Furthermore, in one embodiment, data center 1000 trains one or more machine learning models at least in part based on training data generated by the generative models.

[0195] Autonomous vehicles

[0196] Figure 11A An example of an autonomous vehicle 1100 according to at least one embodiment is shown. In at least one embodiment, the autonomous vehicle 1100 (which may alternatively be referred to herein as "vehicle 1100") may be, but is not limited to, a passenger vehicle, such as a car, truck, bus, and / or another type of vehicle capable of accommodating one or more passengers. In at least one embodiment, vehicle 1100 may be a semi-tractor-trailer for hauling goods. In at least one embodiment, vehicle 1100 may be an aircraft, robotic vehicle, or other type of vehicle.

[0197] Autonomous vehicles can be described according to the levels of automation defined by the National Highway Traffic Safety Administration (“NHTSA”) and the Society of Automotive Engineers (“SAE”) under the U.S. Department of Transportation, in their “Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles” (e.g., standard number J3016-201806 published June 15, 2018; standard number J3016-201609 published September 30, 2016; and previous and future versions of this standard). In one or more embodiments, vehicle 1100 may be able to function according to one or more of the levels of autonomous driving, from Level 1 to Level 5. For example, in at least one embodiment, vehicle 1100 may be able to perform conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5).

[0198] In at least one embodiment, vehicle 1100 may include, but is not limited to, components such as chassis, body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other vehicle components. In at least one embodiment, vehicle 1100 may include, but is not limited to, propulsion system 1150, such as an internal combustion engine, a hybrid powertrain, an all-electric motor, and / or another type of propulsion system. In at least one embodiment, propulsion system 1150 may be connected to the drivetrain of vehicle 1100, which may include, but is not limited to, a transmission, to enable propulsion of vehicle 1100. In at least one embodiment, propulsion system 1150 may be controlled in response to receiving a signal from throttle / accelerator 1152.

[0199] In at least one embodiment, when the propulsion system 1150 is operating (e.g., when the vehicle 1100 is traveling), the steering system 1154 (which may include, but is not limited to, a steering wheel) is used to steer the vehicle 1100 (e.g., along a desired path or route). In at least one embodiment, the steering system 1154 may receive signals from the steering actuator 1156. In at least one embodiment, the steering wheel may be optional for fully automated (Level 5) functionality. In at least one embodiment, the brake sensor system 1146 may be used to operate the vehicle brakes in response to signals received from the brake actuator 1148 and / or brake sensors.

[0200] In at least one embodiment, controller 1136 may include, but is not limited to, one or more system-on-chips (“SoCs”). Figure 11AA controller 1136 (not shown) and / or a graphics processing unit (“GPU”) provides signals (e.g., signals representing commands) to one or more components and / or systems of vehicle 1100. For example, in at least one embodiment, controller 1136 may send signals to operate vehicle braking via brake actuator 1148, to operate steering system 1154 via one or more steering actuators 1156, and to operate propulsion system 1150 via one or more throttles / accelerators 1152. In at least one embodiment, one or more controllers 1136 may include one or more onboard (e.g., integrated) computing devices that process sensor signals and output operating commands (e.g., signals representing commands) to enable autonomous driving and / or assist a driver in driving vehicle 1100. In at least one embodiment, one or more controllers 1136 may include a first controller for autonomous driving functions, a second controller for functional safety functions, a third controller for artificial intelligence functions (e.g., computer vision), a fourth controller for infotainment functions, a fifth controller for redundancy in emergency situations, and / or other controllers. In at least one embodiment, a single controller may handle two or more of the functions described above, and two or more controllers may handle a single function and / or any combination thereof.

[0201] In at least one embodiment, one or more controllers 1136 provide signals for controlling one or more components and / or systems of vehicle 1100 in response to sensor data received from one or more sensors (e.g., sensor inputs). In at least one embodiment, sensor data can be received from sensors, including but not limited to one or more Global Navigation Satellite System (“GNSS”) sensors 1158 (e.g., one or more Global Positioning System sensors), one or more RADAR sensors 1160, one or more ultrasonic sensors 1162, one or more LIDAR sensors 1164, one or more inertial measurement unit (IMU) sensors 1166 (e.g., one or more accelerometers, one or more gyroscopes, one or more magnetic compasses, one or more magnetometers, etc.), one or more microphones 1196, one or more stereo cameras 1168, one or more wide-angle cameras 1170 (e.g., fisheye cameras), one or more infrared cameras 1172, one or more surround cameras 1174 (e.g., 360-degree cameras), and remote cameras (…). Figure 11A (not shown in the image), medium-range camera ( Figure 11A(Not shown in the image) One or more speed sensors 1144 (e.g., for measuring the speed of vehicle 1100), one or more vibration sensors 1142, one or more steering sensors 1140, one or more brake sensors (e.g., as part of brake sensor system 1146) and / or other sensor types are received.

[0202] In at least one embodiment, one or more controllers 1136 may receive input (e.g., represented by input data) from the instrument panel 1132 of the vehicle 1100 and provide output (e.g., represented by output data, display data, etc.) via a human-machine interface (“HMI”) display 1134, a voice signaler, a speaker, and / or other components of the vehicle 1100. In at least one embodiment, the output may include information such as vehicle speed, velocity, time, map data (e.g., high-definition map). Figure 11A The HMI display 1134 may display information about the presence of one or more objects (e.g., road signs, warning signs, traffic light changes, etc.) and / or information about driving operations that the vehicle has already made, is making, or will make (e.g., changing lanes now, exiting exit 34B within two miles, etc.). For example, in at least one embodiment, the HMI display 1134 may display information about the presence of one or more objects (e.g., road signs, warning signs, traffic light changes, etc.) and / or information about driving operations that the vehicle has made, is making, or will make (e.g., changing lanes now, exiting exit 34B within two miles, etc.).

[0203] In at least one embodiment, vehicle 1100 further includes a network interface 1124 that can communicate over one or more networks using one or more wireless antennas 1126 and / or one or more modems. For example, in at least one embodiment, network interface 1124 may be able to communicate over Long Term Evolution (“LTE”), Wideband Code Division Multiple Access (“WCDMA”), Universal Mobile Telecommunications System (“UMTS”), Global System for Mobile Communications (“GSM”), IMT-CDMA Multicarrier (“CDMA2000”) networks, etc. In at least one embodiment, one or more wireless antennas 1126 may also enable communication between objects in the context (e.g., vehicles, mobile devices) using one or more local area networks (e.g., Bluetooth, Bluetooth Low Energy (LE), Z-Wave, ZigBee, etc.) and / or one or more low-power wide area networks (hereinafter “LPWAN”) (e.g., LoRaWAN, SigFox, etc. protocols).

[0204] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8Aand / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be implemented in the system. Figure 11A The operation is used to infer or predict the operation based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0205] Figure 11B The illustration shows an embodiment according to at least one of the embodiments. Figure 11A Examples of camera positions and fields of view for an autonomous vehicle 1100. In at least one embodiment, the camera and its respective field of view are an example embodiment and are not intended to be limiting. For example, in at least one embodiment, additional and / or alternative cameras may be included and / or the cameras may be located at different positions on the vehicle 1100.

[0206] In at least one embodiment, the camera type used for the camera may include, but is not limited to, a digital camera suitable for use with components and / or systems of vehicle 1100. In at least one embodiment, one or more cameras may operate at Automotive Safety Integrity Level (“ASIL”) B and / or other ASILs. In at least one embodiment, the camera type may have any image capture rate, such as 60 frames per second (fps), 1220 fps, 240 fps, etc. In at least one embodiment, the camera may be able to use a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In at least one embodiment, the color filter array may include a red-to-clear (“RCCC”) color filter array, a red-to-clear-blue (“RCCB”) color filter array, a red-blue-green (“RBGC”) color filter array, a Foveon X3 color filter array, a Bayer sensor (“RGGB”) color filter array, a monochrome sensor color filter array, and / or other types of color filter arrays. In at least one embodiment, a transparent pixel camera, such as a camera with RCCC, RCCB, and / or RBGC color filter arrays, may be used to improve photosensitivity.

[0207] In at least one embodiment, one or more cameras may be used to perform advanced driver assistance system (“ADAS”) functions (e.g., as part of a redundancy or fail-safe design). For example, in at least one embodiment, a multi-function mono camera may be installed to provide functions including lane departure warning, traffic sign assist, and intelligent headlight control. In at least one embodiment, one or more cameras (e.g., all cameras) may simultaneously record and provide image data (e.g., video).

[0208] In at least one embodiment, one or more cameras may be mounted in a mounting assembly, such as a custom-designed (three-dimensional (“3D”-printed) assembly, to cut out stray light and reflections within the vehicle 1100 (e.g., reflections from the dashboard in the windshield mirror), which may interfere with the camera’s image data capture capabilities. Regarding the rearview mirror mounting assembly, in at least one embodiment, the rearview mirror assembly may be 3D-printed custom-made such that the camera mounting plate matches the shape of the rearview mirror. In at least one embodiment, one or more cameras may be integrated into the rearview mirror. In at least one embodiment, for side-view cameras, one or more cameras may also be integrated within four pillars at each corner of the cabin.

[0209] In at least one embodiment, a camera (e.g., a forward-facing camera) having a field of view including a portion of the context in front of the vehicle 1100 can be used for surround view and, with the assistance of one or more controllers 1136 and / or control SoCs, to help identify the forward path and obstacles, thereby providing information crucial for generating an occupancy grid and / or determining a preferred vehicle path. In at least one embodiment, the forward-facing camera can be used to perform many ADAS functions similar to LIDAR, including but not limited to emergency braking, pedestrian detection, and collision avoidance. In at least one embodiment, the forward-facing camera can also be used for ADAS functions and systems, including but not limited to lane departure warning (“LDW”), adaptive cruise control (“ACC”), and / or other functions (e.g., traffic sign recognition).

[0210] In at least one embodiment, various cameras can be used in a forward-facing configuration, including, for example, a monocular camera platform including a CMOS (“complementary metal-oxide-semiconductor” color imager. In at least one embodiment, the wide-angle camera 1170 can be used to sense objects entering from the periphery (e.g., pedestrians, people crossing the street, or bicycles). Although in Figure 11B Only one wide-angle camera 1170 is shown; however, in other embodiments, the vehicle 1100 may have any number (including zero) of wide-angle cameras. In at least one embodiment, any number of remote cameras 1198 (e.g., a pair of remote stereo cameras) can be used for depth-based object detection, especially for objects for which a neural network has not yet been trained. In at least one embodiment, the remote camera 1198 can also be used for object detection and classification, as well as basic object tracking.

[0211] In at least one embodiment, any number of stereo cameras 1168 may also be included in a forward configuration. In at least one embodiment, one or more stereo cameras 1168 may include an integrated control unit comprising a scalable processing unit that may provide programmable logic (“FPGA”) and a multi-core microprocessor with a controller area network (“CAN”) or Ethernet interface integrated on a single chip. In at least one embodiment, such a unit may be used to generate a 3D map of the context of vehicle 1100, including distance estimates for all points in the image. In at least one embodiment, one or more stereo cameras 1168 may include, but are not limited to, a compact stereo vision sensor, which may include, but is not limited to, two camera lenses (one on the left and one on the right) and an image processing chip that can measure the distance from vehicle 1100 to a target object and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. In at least one embodiment, other types of stereo cameras 1168 may also be used in addition to those described herein.

[0212] In at least one embodiment, a camera (e.g., a side-view camera) having a field of view that includes a portion of the context of the side of vehicle 1100 can be used for surround viewing, thereby providing information for creating and updating the occupied grid, and generating a side collision warning. For example, in at least one embodiment, a surround camera 1174 (e.g., as...) Figure 11B The four surround cameras shown can be positioned on vehicle 1100. In at least one embodiment, one or more surround cameras 1174 can include, but are not limited to, any number and combination of wide-angle cameras, one or more fisheye lenses, one or more 360-degree cameras, and / or similar cameras. For example, in at least one embodiment, four fisheye lens cameras can be located at the front, rear, and sides of vehicle 1100. In at least one embodiment, vehicle 1100 can use three surround cameras 1174 (e.g., left, right, and rear) and can utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround-view camera.

[0213] In at least one embodiment, a camera (e.g., a rear-view camera) having a field of view that includes a portion of the context behind the vehicle 1100 can be used for parking assistance, surround view, rear collision warning, and creating and updating occupancy raster. In at least one embodiment, a wide variety of cameras can be used, including but not limited to cameras that are also suitable as one or more forward-facing cameras (e.g., long-range camera 1198 and / or one or more mid-range cameras 1176, one or more stereo cameras 1168, one or more infrared cameras 1172, etc.), as described herein.

[0214] In at least one embodiment, inference and / or training logic 815 is used to enable two or more subsequent layers of one or more neural networks to execute on separate computational resources using the output from previous layers of one or more neural networks. In at least one embodiment, inference and / or training logic 815 provides results from previous layers to two or more subsequent layers by determining whether to provide the same output from previous layers to two or more subsequent layers.

[0215] Figure 11C The illustration shows an embodiment according to at least one of the embodiments. Figure 11A A block diagram of an example system architecture for an autonomous vehicle 1100. In at least one embodiment, Figure 11C Each of one or more components, one or more features, and one or more systems of vehicle 1100 is shown connected via bus 1102. In at least one embodiment, bus 1102 may include, but is not limited to, a CAN data interface (which may alternatively be referred to herein as “CAN bus”). In at least one embodiment, CAN may be a network within vehicle 1100 used to help control various features and functions of vehicle 1100, such as brake actuation, acceleration, braking, steering, windshield wipers, etc. In one embodiment, bus 1102 may be configured to have dozens or even hundreds of nodes, each node having its own unique identifier (e.g., CAN ID). In at least one embodiment, bus 1102 can be read to find steering wheel angle, ground speed, engine rotation speed (“RPM”), button position, and / or other vehicle status indicators. In at least one embodiment, bus 1102 may be an ASIL B compliant CAN bus.

[0216] In at least one embodiment, FlexRay and / or Ethernet protocols may be used in addition to or from CAN. In at least one embodiment, there may be any number of molded buses 1102, which may include, but are not limited to, zero or more CAN buses, zero or more FlexRay buses, zero or more Ethernet buses, and / or zero or more other types of buses using other protocols. In at least one embodiment, two or more buses may be used to perform different functions and / or may be used for redundancy. For example, a first bus may be used for a collision avoidance function, and a second bus may be used for actuation control. In at least one embodiment, each of any number of System-on-Chip (“SoC”) 1104 (e.g., SoC 1104(A) and SoC 1104(B)), each of one or more controllers 1136, and / or each computer within the vehicle may access the same input data (e.g., input from sensors of vehicle 1100) and may be connected to a common bus, such as a CAN bus.

[0217] In at least one embodiment, vehicle 1100 may include one or more controllers 1136, such as those described herein. Figure 11A As described above. In at least one embodiment, controller 1136 can be used for a variety of functions. In at least one embodiment, controller 1136 can be coupled to any of various other components and systems of vehicle 1100 and can be used to control vehicle 1100, artificial intelligence of vehicle 1100, infotainment and / or other functions of vehicle 1100.

[0218] In at least one embodiment, vehicle 1100 may include any number of SoCs 1104. In at least one embodiment, each of the SoCs 1104 may include, but is not limited to, a central processing unit (“one or more CPUs”) 1106, a graphics processing unit (“one or more GPUs”) 1108, one or more processors 1110, one or more caches 1112, one or more accelerators 1114, one or more data storage 1116, and / or other components and features not shown. In at least one embodiment, one or more SoCs 1104 may be used to control vehicle 1100 on various platforms and systems. For example, in at least one embodiment, one or more SoCs 1104 may be combined with a high-definition (“HD”) map 1122 in a system (e.g., the system of vehicle 1100), the HD map 1122 being accessible from one or more servers via a network interface 1124. Figure 11C (Not shown in the image) Get map refresh and / or update.

[0219] In at least one embodiment, one or more CPUs 1106 may include CPU clusters or CPU complexes (which may alternatively be referred to herein as “CCPLEX”). In at least one embodiment, one or more CPUs 1106 may include multiple cores and / or a secondary (“L2”) cache. For example, in at least one embodiment, one or more CPUs 1106 may include eight cores in an intercoupled multiprocessor configuration. In at least one embodiment, one or more CPUs 1106 may include four dual-core clusters, each cluster having a dedicated L2 cache (e.g., 2MB L2 cache). In at least one embodiment, one or more CPUs 1106 (e.g., CCPLEX) may be configured to support simultaneous cluster operation, such that any combination of clusters of one or more CPUs 1106 can be active at any given time.

[0220] In at least one embodiment, one or more CPUs 1106 may implement power management functions, including but not limited to one or more of the following features: automatic clock gating of individual hardware modules to conserve dynamic power when idle; clock gating of each core when the core is not actively executing instructions due to executing Wait for Interrupt (“WFI”) / Event Wait (“WFE”) instructions; independent power supply for each core; independent clock gating for each core cluster when all cores are clock-gated or power-gated; and / or independent power gating for each core cluster when all cores are power-gated. In at least one embodiment, one or more CPUs 1106 may further implement an enhanced algorithm for managing power states, wherein allowed power states and expected wake-up times are specified, and the hardware / microcode determines the optimal power state for cores, clusters, and CCPLEX inputs. In at least one embodiment, the processing core may support a simplified power state input sequence in software, wherein the work is offloaded to the microcode.

[0221] In at least one embodiment, one or more GPUs 1108 may include integrated GPUs (or “iGPUs” herein). In at least one embodiment, one or more GPUs 1108 may be programmable and efficient for parallel workloads. In at least one embodiment, one or more GPUs 1108 may use an enhanced tensor instruction set. In one embodiment, one or more GPUs 1108 may include one or more streaming microprocessors, wherein each streaming microprocessor may include a Level 1 (“L1”) cache (e.g., an L1 cache with at least 96 KB of storage capacity), and two or more streaming microprocessors may share an L2 cache (e.g., an L2 cache with 512 KB of storage capacity). In at least one embodiment, one or more GPUs 1108 may include at least eight streaming microprocessors. In at least one embodiment, one or more GPUs 1108 may use a computation application programming interface (API). In at least one embodiment, one or more GPUs 1108 may use one or more parallel computing platforms and / or programming models (e.g., NVIDIA’s CUDA model).

[0222] In at least one embodiment, one or more GPUs 1108 may be power-optimized for optimal performance in automotive and embedded use cases. For example, in one embodiment, one or more GPUs 1108 may be fabricated on a FinFET (“FinFET”) circuit. In at least one embodiment, each streaming microprocessor may include multiple mixed-precision processing cores divided into multiple blocks. For example, but not limited to, 64 PF32 cores and 32 PF64 cores may be divided into four processing blocks. In at least one embodiment, each processing block may be allocated 16 FP32 cores, 8 FP64 cores, 16 INT32 cores, two mixed-precision NVIDIA Tensor cores for deep learning matrix arithmetic, a level-zero (“L0”) instruction cache, a thread bundle scheduler, a dispatch unit, and / or a 64 KB register file. In at least one embodiment, the streaming microprocessor may include independent parallel integer and floating-point data paths to provide efficient execution of workloads that mix computation and addressing operations. In at least one embodiment, the streaming microprocessor may include independent thread scheduling capabilities to enable finer-grained synchronization and cooperation between parallel threads. In at least one embodiment, the streaming microprocessor may include a combined L1 data cache and shared memory unit to improve performance while simplifying programming.

[0223] In at least one embodiment, one or more GPUs 1108 may include high-bandwidth memory (“HBM”) and / or a 16 GB HBM2 memory subsystem to provide a peak storage bandwidth of approximately 900 GB / s in some examples. In at least one embodiment, in addition to or instead of HBM memory, synchronous graphics random access memory (“SGRAM”), such as graphics double data rate type five synchronous random access memory (“GDDR5”), may be used.

[0224] In at least one embodiment, one or more GPUs 1108 may include unified memory technology. In at least one embodiment, address translation service (“ATS”) support can be used to allow one or more GPUs 1108 to directly access the page tables of one or more CPUs 1106. In at least one embodiment, when a memory management unit (“MMU”) of one or more GPUs 1108 experiences a miss, an address translation request can be sent to one or more CPUs 1106. In response, in at least one embodiment, two CPUs of one or more CPUs 1106 can look up the virtual-physical mapping of the address in their page tables and transfer the translation back to one or more GPUs 1108. In at least one embodiment, unified memory technology can allow a single unified virtual address space to be used for the memory of both one or more CPUs 1106 and one or more GPUs 1108, thereby simplifying the programming of one or more GPUs 1108 and the porting of applications to one or more GPUs 1108.

[0225] In at least one embodiment, one or more GPUs 1108 may include any number of access counters that can track the frequency of memory accesses by one or more GPUs 1108 to other processors. In at least one embodiment, one or more access counters can help ensure that memory pages are moved to the physical memory of the processor that accesses the pages most frequently, thereby improving the efficiency of shared memory ranges between processors.

[0226] In at least one embodiment, one or more SoCs 1104 may include any number of caches 1112, including those described herein. For example, in at least one embodiment, one or more caches 1112 may include a Level 3 (“L3”) cache available for one or more CPUs 1106 and one or more GPUs 1108 (e.g., connected to CPUs 1106 and GPUs 1108). In at least one embodiment, one or more caches 1112 may include a write-back cache that can, for example, track the state of a line using a cache coherence protocol (e.g., MEI, MESI, MSI, etc.). In at least one embodiment, although a smaller cache size may be used, according to an embodiment, the L3 cache may include 4 MB of memory or more.

[0227] In at least one embodiment, one or more SoCs 1104 may include one or more accelerators 1114 (e.g., hardware accelerators, software accelerators, or combinations thereof). In at least one embodiment, one or more SoCs 1104 may include a hardware acceleration cluster, which may include optimized hardware accelerators and / or large on-chip memory. In at least one embodiment, large on-chip memory (e.g., 4MB of SRAM) enables the hardware acceleration cluster to accelerate neural networks and other computations. In at least one embodiment, the hardware acceleration cluster may be used to supplement one or more GPUs 1108 and offload some tasks from one or more GPUs 1108 (e.g., freeing up more cycles from one or more GPUs 1108 to perform other tasks). In at least one embodiment, one or more accelerators 1114 may be used for a target workload (e.g., perceptual, convolutional neural network (“CNN”), recurrent neural network (“RNN”), etc.) that is sufficiently stable to withstand acceleration testing. In at least one embodiment, the CNN may include region-based or region convolutional neural networks (“RCNN”) and fast RCNN (e.g., for object detection) or other types of CNNs.

[0228] In at least one embodiment, one or more accelerators 1114 (e.g., a hardware acceleration cluster) may include one or more deep learning accelerators (“DLAs”). In at least one embodiment, one or more DLAs may include, but are not limited to, one or more Tensor Processing Units (“TPUs”), which may be configured to provide an additional 10 trillion operations per second for deep learning applications and inference. In at least one embodiment, a TPU may be an accelerator configured and optimized for performing image processing functions (e.g., for CNNs, RCNNs, etc.). In at least one embodiment, one or more DLAs may be further optimized for specific sets of neural network types and floating-point operations and inference. In at least one embodiment, one or more DLAs are designed to provide higher performance per millimeter than typical general-purpose GPUs and typically significantly outperform CPUs. In at least one embodiment, one or more TPUs may perform several functions, including single-instance convolution functions supporting, for example, INT8, INT16, and FP16 data types for features and weights, as well as post-processor functions. In at least one embodiment, one or more DLAs can execute neural networks, particularly CNNs, quickly and efficiently on processed or unprocessed data for any of the various functions, including, but not limited to: CNNs for object recognition and detection using data from camera sensors; CNNs for distance estimation using data from camera sensors; CNNs for emergency vehicle detection, recognition, and identification using data from microphones; CNNs for face recognition and vehicle owner recognition using data from camera sensors; and / or CNNs for safety and / or safety-related events.

[0229] In at least one embodiment, the DLA can perform any function of one or more GPUs 1108, and by using inference accelerators, for example, the designer can target one or more DLAs or one or more GPUs 1108 for any function. For example, in at least one embodiment, the designer can concentrate the CNN processing and floating-point operations on one or more DLAs, leaving other functions to one or more GPUs 1108 and / or one or more accelerators 1114.

[0230] In at least one embodiment, one or more accelerators 1114 may include programmable vision accelerators (“PVAs”), which may alternatively be referred to herein as computer vision accelerators. In at least one embodiment, one or more PVAs may be designed and configured to accelerate computer vision algorithms for advanced driver assistance systems (“ADAS”) 1138, autonomous driving, augmented reality (“AR”) applications, and / or virtual reality (“VR”) applications. In at least one embodiment, one or more PVAs may strike a balance between performance and flexibility. For example, in at least one embodiment, each of one or more PVAs may include, for example, but not limited to, any number of reduced instruction set computer (“RISC”) cores, direct memory access (“DMA”), and / or any number of vector processors.

[0231] In at least one embodiment, the RISC core can interact with an image sensor (e.g., the image sensor of any camera described herein), an image signal processor, etc. In at least one embodiment, each RISC core may include any number of memories. In at least one embodiment, the RISC core may use any of a variety of protocols, depending on the embodiment. In at least one embodiment, the RISC core may execute a real-time operating system (“RTOS”). In at least one embodiment, the RISC core may be implemented using one or more integrated circuit devices, application-specific integrated circuits (“ASICs”), and / or storage devices. For example, in at least one embodiment, the RISC core may include an instruction cache and / or tightly coupled RAM.

[0232] In at least one embodiment, DMA enables components of the PVA to access system memory independently of one or more CPUs 1106. In at least one embodiment, DMA can support any number of features for providing optimization to the PVA, including but not limited to, support for multidimensional addressing and / or circular addressing. In at least one embodiment, DMA can support up to six or more addressing dimensions, which may include, but are not limited to, block width, block height, block depth, horizontal block step, vertical block step, and / or depth step.

[0233] In at least one embodiment, the vector processor may be a programmable processor designed to efficiently and flexibly execute programming for computer vision algorithms and provide signal processing capabilities. In at least one embodiment, the PVA may include a PVA core and two vector processing subsystem partitions. In at least one embodiment, the PVA core may include a processor subsystem, a DMA engine (e.g., two DMA engines), and / or other peripherals. In at least one embodiment, the vector processing subsystem may serve as the main processing engine of the PVA and may include a vector processing unit (“VPU”), an instruction cache, and / or a vector memory (e.g., “VMEM”). In at least one embodiment, the VPU core may include a digital signal processor, such as a Single Instruction Multiple Data (“SIMD”) or Very Long Instruction Word (“VLIW”) digital signal processor. In at least one embodiment, the combination of SIMD and VLIW can improve throughput and speed.

[0234] In at least one embodiment, each vector processor may include an instruction cache and may be coupled to dedicated memory. As a result, in at least one embodiment, each vector processor may be configured to execute independently of other vector processors. In at least one embodiment, the vector processors included in a particular PVA may be configured to employ data parallelism. For example, in at least one embodiment, multiple vector processors included in a single PVA may execute general-purpose computer vision algorithms, except on different regions of an image. In at least one embodiment, vector processors included in a particular PVA may execute different computer vision algorithms simultaneously on a single image, or even execute different algorithms on a sequence of images or portions of images. In at least one embodiment, among others, any number of PVAs may be included in the hardware-accelerated cluster, and any number of vector processors may be included in each PVA. In at least one embodiment, the PVA may include additional error-correcting code (“ECC”) memory to enhance overall system security.

[0235] In at least one embodiment, one or more accelerators 1114 may include an on-chip computer vision network and static random access memory (“SRAM”) for providing high-bandwidth, low-latency SRAM to one or more accelerators 1114. In at least one embodiment, the on-chip memory may include at least 4 MB of SRAM, comprising, for example, but not limited to, eight field-configurable memory blocks accessible to both the PVA and DLA. In at least one embodiment, each pair of memory blocks may include an Advanced Peripheral Bus (“APB”) interface, configuration circuitry, a controller, and a multiplexer. In at least one embodiment, any type of memory may be used. In at least one embodiment, the PVA and DLA may access the memory via a backbone providing high-speed access to the memory for both the PVA and DLA. In at least one embodiment, the backbone may include an on-chip computer vision network that interconnects the PVA and DLA to the memory (e.g., using an APB).

[0236] In at least one embodiment, the on-chip computer vision network may include an interface that determines that both the PVA and DLA provide ready and valid signals before transmitting any control signals / addresses / data. In at least one embodiment, the interface may provide separate phases and separate channels for transmitting control signals / addresses / data, as well as bursty communication for continuous data transmission. In at least one embodiment, although other standards and protocols may be used, the interface may conform to the International Organization for Standardization (“ISO”) 26262 or the International Electrotechnical Commission (“IEC”) 61508 standard.

[0237] In at least one embodiment, one or more SoCs 1104 may include a real-time eye-tracking hardware accelerator. In at least one embodiment, the real-time eye-tracking hardware accelerator may be used to quickly and efficiently determine the location and extent of an object (e.g., within a world model) to generate real-time visualization simulations for RADAR signal interpretation, for sound propagation synthesis and / or analysis, for simulation of SONAR systems, for general wave propagation simulation, for comparison with LIDAR data for localization and / or other functions, and / or for other purposes.

[0238] In at least one embodiment, one or more accelerators 1114 have broad applications for autonomous driving. In at least one embodiment, PVA can be used in critical processing stages in ADAS and autonomous vehicles. In at least one embodiment, the capabilities of PVA with low power consumption and low latency are well-matched to algorithmic domains requiring predictable processing. In other words, PVA performs well in semi-intensive or intensive conventional computations, even on small datasets that may require predictable runtimes with low latency and low power consumption. In at least one embodiment, such as in vehicle 1100, PVA may be designed to run classical computer vision algorithms, as they are efficient in object detection and integer mathematical operations.

[0239] For example, according to at least one embodiment of the technology, PVA is used to perform computer stereo vision. In at least one embodiment, a semi-global matching-based algorithm may be used in some examples, although this is not intended to be limiting. In at least one embodiment, applications for Level 3-5 autonomous driving use dynamic estimation / stereo matching (e.g., structure recovery from motion, pedestrian recognition, lane detection, etc.) during operation. In at least one embodiment, PVA can perform computer stereo vision functions on input from two monocular cameras.

[0240] In at least one embodiment, the PVA can be used to perform intensive optical flow. For example, in at least one embodiment, the PVA can process raw RADAR data (e.g., using 4D Fast Fourier Transform) to provide processed RADAR data. In at least one embodiment, the PVA is used for time-of-flight depth processing, for example, by processing raw time-of-flight data to provide processed time-of-flight data.

[0241] In at least one embodiment, the DLA can be used to run any type of network to enhance control and driving safety, including, but not limited to, neural networks whose output is used for a confidence score for each object detection. In at least one embodiment, the confidence score can be represented or interpreted as a probability, or as providing a relative “weight” for each detection relative to other detections. In at least one embodiment, the confidence score measurement enables the system to make further decisions about which detections should be considered true positives rather than false positives. In at least one embodiment, the system can set a threshold for the confidence score and only consider detections exceeding the threshold as true positives. In embodiments using an Automatic Emergency Braking (“AEB”) system, false positives would cause the vehicle to automatically perform emergency braking, which is obviously undesirable. In at least one embodiment, a highly confident detection can be considered a trigger for AEB. In at least one embodiment, the DLA can run a neural network for regressing the confidence score value. In at least one embodiment, the neural network may take at least a subset of parameters as its input, such as bounding box size, obtained ground plane estimate (e.g., from another subsystem), and outputs of one or more IMU sensors 1166 related to the vehicle 1100 orientation, distance, and 3D position estimate of the object obtained from the neural network and / or other sensors (e.g., one or more LiDAR sensors 1164 or one or more RADAR sensors 1160).

[0242] In at least one embodiment, one or more SoCs 1104 may include one or more data storage devices 1116 (e.g., memory). In at least one embodiment, one or more data storage devices 1116 may be on-chip memory of one or more SoCs 1104, which may store neural networks to be executed on one or more GPUs 1108 and / or DLAs. In at least one embodiment, one or more data storage devices 1116 may have a sufficiently large capacity to store multiple instances of the neural network for redundancy and security. In at least one embodiment, one or more data storage devices 1116 may include L2 or L3 caches.

[0243] In at least one embodiment, one or more SoCs 1104 may include any number of processors 1110 (e.g., embedded processors). In at least one embodiment, one or more processors 1110 may include a startup and power management processor, which may be a dedicated processor and subsystem for handling startup power and management functions, as well as associated security implementations. In at least one embodiment, the startup and power management processor may be part of a startup sequence of one or more SoCs 1104 and may provide runtime power management services. In at least one embodiment, the startup power and management processor may provide clock and voltage programming, assist system low-power state transitions, thermal and temperature sensor management of one or more SoCs 1104s, and / or power state management of one or more SoCs 1104s. In at least one embodiment, each temperature sensor may be implemented with its output frequency proportional to temperature, and one or more SoCs 1104s may use the ring oscillator to detect the temperature of one or more CPUs 1106, one or more GPUs 1108, and / or one or more accelerators 1114. In at least one embodiment, if it is determined that the temperature exceeds a threshold, the startup and power management processor may enter a temperature fault routine and place one or more SoCs 1104 into a lower power state and / or place the vehicle 1100 into a driver’s safe stopping pattern (e.g., bring the vehicle 1100 to a safe stop).

[0244] In at least one embodiment, one or more processors 1110 may further include a set of embedded processors that can serve as an audio processing engine. The audio processing engine may be an audio subsystem capable of providing full hardware support for multi-channel audio through multiple interfaces and a wide and flexible range of audio I / O interfaces. In at least one embodiment, the audio processing engine is a dedicated processor core with a digital signal processor having dedicated RAM.

[0245] In at least one embodiment, one or more processors 1110 may further include an always-on processor engine that can provide the necessary hardware features to support low-power sensor management and wake-up use cases. In at least one embodiment, the processor on the always-on processor engine may include, but is not limited to, a processor core, tightly coupled RAM, peripheral support devices (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0246] In at least one embodiment, one or more processors 1110 may further include a secure clustering engine, which includes, but is not limited to, a dedicated processor subsystem for handling security management of automotive applications. In at least one embodiment, the secure clustering engine may include, but is not limited to, two or more processor cores, tightly coupled RAM, supporting peripherals (e.g., timers, interrupt controllers, etc.), and / or routing logic. In secure mode, in at least one embodiment, the two or more cores may operate in lockstep mode and may be used as a single core with comparison logic for detecting any differences between their operations. In at least one embodiment, one or more processors 1110 may further include a real-time camera engine, which may include, but is not limited to, a dedicated processor subsystem for handling real-time camera management. In at least one embodiment, one or more processors 1110 may further include a high dynamic range signal processor, which may include, but is not limited to, an image signal processor, which is a hardware engine as part of the camera processing pipeline.

[0247] In at least one embodiment, one or more processors 1110 may include a video image synthesizer, which may be a processing block (e.g., implemented on a microprocessor) that implements video post-processing functions required by the video playback application to produce the final image for the player window. In at least one embodiment, the video image synthesizer may perform lens distortion correction on one or more wide-angle cameras 1170, one or more surround cameras 1174, and / or one or more cabin monitoring camera sensors. In at least one embodiment, preferably, the cabin monitoring camera sensors are monitored by a neural network running on another instance of SoC 1104, the neural network being configured to recognize cabin events and respond accordingly. In at least one embodiment, the cabin system may perform, but is not limited to, lip reading to activate cellular service and make phone calls, instruct emails, change the vehicle's destination, activate or change the vehicle's infotainment system and settings, or provide voice-activated web browsing. In at least one embodiment, certain functions are available to the driver when the vehicle is operating in autonomous mode, and are otherwise disabled.

[0248] In at least one embodiment, the video image synthesizer may include enhanced temporal denoising for simultaneous spatial and temporal denoising. For example, in at least one embodiment, when motion occurs in the video, denoising appropriately weights spatial information, thereby reducing the weight of information provided by adjacent frames. In at least one embodiment, when the image or a portion of the image does not contain motion, temporal denoising performed by the video image synthesizer may use information from previous images to reduce noise in the current image.

[0249] In at least one embodiment, the video image compositor can also be configured to perform stereoscopic correction on the input stereo lens frames. In at least one embodiment, when using an operating system desktop, the video image compositor can also be used for user interface compositing and does not require one or more GPUs 1108 to continuously render new surfaces. In at least one embodiment, when one or more GPUs 1108 are powered and actively performing 3D rendering, the video image compositor can be used to offload one or more GPUs 1108 to improve performance and responsiveness.

[0250] In at least one embodiment, one or more SoCs of SoC 1104 may further include a Mobile Industrial Processor Interface (“MIPI”) camera serial interface, a high-speed interface, and / or a video input block that can be used for receiving video and input from a camera and associated pixel input functions. In at least one embodiment, one or more SoCs of SoC 1104 may further include an input / output controller that can be software controlled and can be used to receive I / O signals not assigned to a specific role.

[0251] In at least one embodiment, one or more SoCs of SoC 1104 may further include extensive peripheral interfaces to enable communication with peripheral devices, audio encoders / decoders (“codecs”), power management and / or other devices. In at least one embodiment, one or more SoCs of SoC 1104 may be used to process data from (e.g., connected via gigabit multimedia serial links and Ethernet channels) cameras, sensors (e.g., one or more LiDAR sensors 1164, one or more RADAR sensors 1160, etc., which may be connected via Ethernet channels), data from bus 1102 (e.g., vehicle 1100 speed, steering wheel position, etc.), data from one or more GNSS sensors 1158 (e.g., connected via Ethernet bus or CAN bus), etc. In at least one embodiment, one or more SoCs of SoC 1104 may further include a dedicated high-performance mass storage controller, which may include its own DMA engine and may be used to free one or more CPUs 1106 from routine data management tasks.

[0252] In at least one embodiment, one or more SoCs 1104 can be an end-to-end platform with a flexible architecture spanning automation levels 3-5, providing a comprehensive functional safety architecture that leverages and effectively utilizes computer vision and ADAS technologies to achieve diversity and redundancy. This provides a platform offering a flexible and reliable driving software stack as well as deep learning tools. In at least one embodiment, one or more SoCs 1104 can be faster, more reliable, and even more energy and space efficient than conventional systems. For example, in at least one embodiment, one or more accelerators 1114, when combined with one or more CPUs 1106, one or more GPUs 1108, and one or more data storage devices 1116, can provide a fast and efficient platform for Level 3-5 autonomous vehicles.

[0253] In at least one embodiment, the computer vision algorithm can be executed on a CPU, which can be configured using a high-level programming language (e.g., C) to execute multiple processing algorithms on a variety of visual data. However, in at least one embodiment, the CPU typically cannot meet the performance requirements of many computer vision applications, such as performance requirements related to execution time and power consumption. In at least one embodiment, many CPUs cannot execute complex object detection algorithms in real time, which are used in automotive ADAS applications and practical Level 3-5 autonomous vehicles.

[0254] The embodiments described herein allow multiple neural networks to be executed simultaneously and / or sequentially, and allow the results to be combined to achieve Level 3-5 autonomous driving capabilities. For example, in at least one embodiment, a CNN executed on a DLA or discrete GPU (e.g., one or more GPUs 1120) may include text and word recognition, thereby allowing a supercomputer to read and understand traffic signs, including signs for which the neural network has not yet been specifically trained. In at least one embodiment, the DLA may also include a neural network capable of recognizing, interpreting, and providing semantic understanding of symbols, and passing this semantic understanding to a path planning module running on a CPU Complex.

[0255] In at least one embodiment, for drives of levels 3, 4, or 5, multiple neural networks can run simultaneously. For example, in at least one embodiment, a warning sign consisting of a light bulb accompanied by the warning sign “Caution: flashing lights indicate icy conditions” can be interpreted independently or jointly by multiple neural networks. In at least one embodiment, the warning sign itself can be recognized as a traffic sign by a first deployed neural network (e.g., a trained neural network), and the text “flashing lights indicate icy conditions” can be interpreted by a second deployed neural network, which informs the vehicle’s path planning software (preferably executed on the CPU Complex) that icing conditions exist when flashing lights are detected. In at least one embodiment, flashing lights can be identified by operating a third deployed neural network across multiple frames, informing the vehicle’s path planning software of the presence (or absence) of flashing lights. In at least one embodiment, all three neural networks can run simultaneously, for example within the DLA and / or on one or more GPUs 1108.

[0256] In at least one embodiment, the CNN for facial recognition and vehicle owner identification can use data from camera sensors to identify the presence of an authorized driver and / or the owner of vehicle 1100. In at least one embodiment, a normally open sensor processor engine can be used to unlock the vehicle when the owner approaches the driver's door and turns on the lights, and, in security mode, can be used to disable the vehicle when the owner leaves it. In this way, one or more SoCs 1104 provide protection against theft and / or carjacking.

[0257] In at least one embodiment, the CNN for emergency vehicle detection and identification can use data from microphone 1196 to detect and identify emergency vehicle sirens. In at least one embodiment, one or more SoCs 1104 use the CNN to classify contextual and urban sounds, as well as visual data. In at least one embodiment, the CNN running on DLA is trained to identify the relative approach speed of emergency vehicles (e.g., by using the Doppler effect). In at least one embodiment, the CNN can also be trained to identify emergency vehicles in the area where the vehicle is operating, as identified by one or more GNSS sensors 1158. In at least one embodiment, when operating in Europe, the CNN will seek to detect European sirens, while in North America, the CNN will seek to identify only North American sirens. In at least one embodiment, once an emergency vehicle is detected, a control program can be used, with the assistance of one or more ultrasonic sensors 1162, to execute emergency vehicle safety routines, slow the vehicle, pull the vehicle to the side of the road, stop, and / or leave the vehicle idle until the emergency vehicle passes.

[0258] In at least one embodiment, vehicle 1100 may include one or more CPUs 1118 (e.g., one or more discrete CPUs or one or more dCPUs) that may be coupled to one or more SoCs 1104 via high-speed interconnects (e.g., PCIe). In at least one embodiment, one or more CPUs 1118 may include x86 processors. For example, one or more CPUs 1118 may be used to perform any of the various functions, such as arbitrating the results of potential inconsistencies between ADAS sensors and one or more SoCs 1104, and / or monitoring the status and health of one or more monitoring controllers 1136 and / or on-chip information systems (“information SoCs”) 1130.

[0259] In at least one embodiment, vehicle 1100 may include one or more GPUs 1120 (e.g., one or more discrete GPUs or one or more dGPUs) coupled to one or more SoCs 1104 via high-speed interconnects (e.g., NVIDIA's NVLINK channels). In at least one embodiment, one or more GPUs 1120 may provide additional artificial intelligence capabilities, such as by executing redundant and / or different neural networks, and may be used to train and / or update the neural networks based at least in part on inputs from sensors of vehicle 1100 (e.g., sensor data).

[0260] In at least one embodiment, vehicle 1100 may further include a network interface 1124, which may include, but is not limited to, one or more wireless antennas 1126 (e.g., one or more wireless antennas for different communication protocols, such as cellular antennas, Bluetooth antennas, etc.). In at least one embodiment, network interface 1124 may be used to enable wireless connectivity with other vehicles and / or computing devices (e.g., passenger client devices) via Internet cloud services (e.g., using servers and / or other network devices). In at least one embodiment, for communication with other vehicles, a direct link and / or an indirect link (e.g., via a network and the Internet) may be established between vehicle 1100 and another vehicle. In at least one embodiment, a vehicle-to-vehicle communication link may be used to provide a direct link. In at least one embodiment, the vehicle-to-vehicle communication link may provide vehicle 1100 with information about vehicles near vehicle 1100 (e.g., vehicles in front, to the side, and / or behind vehicle 1100). In at least one embodiment, the foregoing functionality may be part of a cooperative adaptive cruise control function of vehicle 1100.

[0261] In at least one embodiment, network interface 1124 may include a System-on-Chip (SoC) that provides modulation and demodulation functions and enables one or more controllers 1136 to communicate over a wireless network. In at least one embodiment, network interface 1124 may include a radio frequency (RF) front-end for up-conversion from baseband to RF and down-conversion from RF to baseband. In at least one embodiment, frequency conversion may be performed in any technically feasible manner. For example, frequency conversion may be performed using known processes and / or using a superheterodyne process. In at least one embodiment, the RF front-end functionality may be provided by a separate chip. In at least one embodiment, the network interface may include wireless functions for communication over LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-Wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0262] In at least one embodiment, vehicle 1100 may further include one or more data storage units 1128, which may include, but are not limited to, off-chip (e.g., one or more SoC 1104) storage. In at least one embodiment, one or more data storage units 1128 may include, but are not limited to, one or more storage elements, including RAM, SRAM, dynamic random access memory (“DRAM”), video random access memory (“VRAM”), flash memory, hard disk and / or other components and / or devices capable of storing at least one bit of data.

[0263] In at least one embodiment, vehicle 1100 may further include one or more GNSS sensors 1158 (e.g., GPS and / or auxiliary GPS sensors) to assist in map creation, perception, occupancy raster generation, and / or path planning functions. In at least one embodiment, any number of GNSS sensors 1158 may be used, including, for example, but not limited to, GPS sensors connected to a serial interface (e.g., RS-232) bridge using a USB connector with Ethernet.

[0264] In at least one embodiment, vehicle 1100 may further include one or more RADAR sensors 1160. In at least one embodiment, one or more RADAR sensors 1160 may be used by vehicle 1100 for remote vehicle detection, even in dark and / or inclement weather conditions. In at least one embodiment, the RADAR functional safety level may be ASIL B. In at least one embodiment, one or more RADAR sensors 1160 may use a CAN bus and / or bus 1102 (e.g., to transmit data generated by one or more RADAR sensors 1160) for control and access to object tracking data, and in some examples may access an Ethernet channel to access raw data. In at least one embodiment, a wide variety of RADAR sensor types may be used. For example, but not limited to, one or more of the RADAR sensors 1160 may be suitable for front, rear, and side RADAR use. In at least one embodiment, one or more RADAR sensors 1160 are pulse Doppler RADAR sensors.

[0265] In at least one embodiment, one or more RADAR sensors 1160 may include different configurations, such as long-range with a narrow field of view, short-range with a wide field of view, short-range side coverage, etc. In at least one embodiment, the long-range RADAR can be used for adaptive cruise control functions. In at least one embodiment, the long-range RADAR system can provide a wide field of view achieved through two or more independent scans (e.g., within a 250m range). In at least one embodiment, one or more RADAR sensors 1160 can help distinguish between stationary and moving objects and can be used by the ADAS system 1138 for emergency braking assistance and forward collision warning. In at least one embodiment, one or more sensors 1160 included in the long-range RADAR system may include, but are not limited to, a monostatic multimode RADAR with multiple (e.g., six or more) fixed RADAR antennas and high-speed CAN and FlexRay interfaces. In at least one embodiment, having six antennas, with the four central antennas, can create a focused beammap designed to record the surrounding context of the vehicle 1100 at a high speed while minimizing traffic interference from adjacent lanes. In at least one embodiment, the other two antennas can expand the field of view, thereby enabling rapid detection of vehicles 1100 entering or leaving the lane.

[0266] In at least one embodiment, as an example, a mid-range RADAR system may include, for example, a range of up to 160m (front) or 80m (rear), and a field of view of up to 42 degrees (front) or 150 degrees (rear). In at least one embodiment, a short-range RADAR system may include, but is not limited to, any number of RADAR sensors 1160 designed to be mounted at both ends of the rear bumper. When mounted at both ends of the rear bumper, in at least one embodiment, the RADAR sensor system may generate two beams that continuously monitor the rearward direction of the vehicle and nearby blind spots. In at least one embodiment, the short-range RADAR system may be used in ADAS system 1138 for blind spot detection and / or lane change assistance.

[0267] In at least one embodiment, vehicle 1100 may further include one or more ultrasonic sensors 1162. In at least one embodiment, one or more ultrasonic sensors 1162, which may be positioned at the front, rear, and / or sides of vehicle 1100, may be used for parking assistance and / or creating and updating occupancy detectors. In at least one embodiment, a wide variety of ultrasonic sensors 1162 may be used, and different ultrasonic sensors 1162 may be used for different detection ranges (e.g., 2.5m, 4m). In at least one embodiment, the ultrasonic sensors 1162 may operate at ASIL B functional safety level.

[0268] In at least one embodiment, vehicle 1100 may include one or more LiDAR sensors 1164. In at least one embodiment, one or more LiDAR sensors 1164 may be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. In at least one embodiment, one or more LiDAR sensors 1164 may operate at functional safety level ASIL B. In at least one embodiment, vehicle 1100 may include multiple (e.g., two, four, six, etc.) LiDAR sensors 1164 that can use Ethernet channels (e.g., providing data to a Gigabit Ethernet switch).

[0269] In at least one embodiment, one or more LiDAR sensors 1164 may be able to provide a list of objects and their distances for a 360-degree field of view. In at least one embodiment, one or more commercially available LiDAR sensors 1164 may, for example, have an advertising range of approximately 100m, an accuracy of 2cm-3cm, and support a 100Mbps Ethernet connection. In at least one embodiment, one or more non-protruding LiDAR sensors may be used. In such an embodiment, one or more LiDAR sensors 1164 may include small devices that can be embedded in the front, rear, side, and / or corner locations of vehicle 1100. In at least one embodiment, one or more LiDAR sensors 1164, in such an embodiment, can provide a horizontal field of view of up to 120 degrees and a vertical field of view of 35 degrees, even for objects with low reflectivity, and have a range of 200m. In at least one embodiment, one or more forward-facing LiDAR sensors 1164 may be configured for a horizontal field of view between 45 degrees and 135 degrees.

[0270] In at least one embodiment, LIDAR technology (such as 3D flash LIDAR) may also be used. In at least one embodiment, 3D flash LIDAR uses a laser flash as a transmission source to illuminate approximately 200m around vehicle 1100. In at least one embodiment, the flash LIDAR unit includes, but is not limited to, a receiver that records the laser pulse propagation time and reflected light on each pixel, which in turn corresponds to the range from vehicle 1100 to the object. In at least one embodiment, flash LIDAR can allow the generation of highly accurate and distortion-free images of the surrounding context using each laser flash. In at least one embodiment, four flash LIDAR sensors may be deployed, one on each side of vehicle 1100. In at least one embodiment, the 3D flash LIDAR system includes, but is not limited to, a solid-state 3D line-of-sight array LIDAR camera with no moving parts other than a fan (e.g., a non-scanning LIDAR device). In at least one embodiment, the flash LIDAR device can use a 5-nanosecond Class I (eye-safe) laser pulse per frame and can capture reflected laser light as a 3D ranging point cloud and co-registered intensity data.

[0271] In at least one embodiment, vehicle 1100 may further include one or more IMU sensors 1166. In at least one embodiment, one or more IMU sensors 1166 may be located at the center of the rear axle of vehicle 1100. In at least one embodiment, one or more IMU sensors 1166 may include, for example, but not limited to, one or more accelerometers, one or more magnetometers, one or more gyroscopes, a magnetic compass, multiple magnetic compasses, and / or other sensor types. In at least one embodiment, for example in a six-axis application, one or more IMU sensors 1166 may include, but are not limited to, accelerometers and gyroscopes. In at least one embodiment, for example in a nine-axis application, one or more IMU sensors 1166 may include, but are not limited to, accelerometers, gyroscopes, and magnetometers.

[0272] In at least one embodiment, one or more IMU sensors 1166 may be implemented as a miniature, high-performance GPS-assisted inertial navigation system (“GPS / INS”) combining a microelectromechanical system (“MEMS”) inertial sensor, a high-sensitivity GPS receiver, and an advanced Kalman filtering algorithm to provide position, velocity, and attitude estimations; in at least one embodiment, one or more IMU sensors 1166 may enable vehicle 1100 to estimate heading without input from a magnetic sensor obtained by directly observing and correlating velocity changes from GPS to one or more IMU sensors 1166. In at least one embodiment, one or more IMU sensors 1166 and one or more GNSS sensors 1158 may be combined in a single integrated unit.

[0273] In at least one embodiment, vehicle 1100 may include one or more microphones 1196 placed inside and / or around vehicle 1100. In at least one embodiment, in addition, one or more microphones 1196 may be used for emergency vehicle detection and identification.

[0274] In at least one embodiment, vehicle 1100 may further include any number of camera types, including one or more stereo cameras 1168, one or more wide-angle cameras 1170, one or more infrared cameras 1172, one or more surround cameras 1174, one or more long-range cameras 1198, one or more mid-range cameras 1176, and / or other camera types. In at least one embodiment, the cameras can be used to capture image data around the entire perimeter of vehicle 1100. In at least one embodiment, the type of camera used depends on vehicle 1100. In at least one embodiment, any combination of camera types can be used to provide the necessary coverage around vehicle 1100. In at least one embodiment, the number of cameras deployed may vary depending on the embodiment. For example, in at least one embodiment, vehicle 1100 may include six cameras, seven cameras, ten cameras, twelve cameras, or other numbers of cameras. In at least one embodiment, the cameras may be, by way of example but not limited to, supporting Gigabit Multimedia Serial Link (“GMSL”) and / or Gigabit Ethernet communication. In at least one embodiment, previously referenced herein... Figure 11A and Figure 11B Each camera can be described in more detail.

[0275] In at least one embodiment, vehicle 1100 may further include one or more vibration sensors 1142. In at least one embodiment, one or more vibration sensors 1142 may measure vibrations of components of vehicle 1100 (e.g., axles). For example, in at least one embodiment, changes in vibration may indicate changes in road surface. In at least one embodiment, when two or more vibration sensors 1142 are used, differences between vibrations may be used to determine road surface friction or slippage (e.g., when there is a vibration difference between a power drive axle and a free-rotating axle).

[0276] In at least one embodiment, vehicle 1100 may include ADAS system 1138. In at least one embodiment, ADAS system 1138 may include, but is not limited to, SoC. In at least one embodiment, ADAS system 1138 may include, but is not limited to, any number of autonomous / adaptive / automatic cruise control (“ACC”) systems, cooperative adaptive cruise control (“CACC”) systems, forward collision warning (“FCW”) systems, automatic emergency braking (“AEB”) systems, lane departure warning (“LDW”) systems, lane keeping assist (“LKA”) systems, blind spot warning (“BSW”) systems, rear cross traffic warning (“RCTW”) systems, collision warning (“CW”) systems, lane centering (“LC”) systems, and / or other systems, features, and / or functions, and combinations thereof.

[0277] In at least one embodiment, the ACC system may use one or more RADAR sensors 1160, one or more LIDAR sensors 1164, and / or any number of cameras. In at least one embodiment, the ACC system may include a longitudinal ACC system and / or a lateral ACC system. In at least one embodiment, the longitudinal ACC system monitors and controls the distance to another vehicle adjacent to vehicle 1100 and automatically adjusts the speed of vehicle 1100 to maintain a safe distance from the vehicle ahead. In at least one embodiment, the lateral ACC system performs distance holding and suggests that vehicle 1100 change lanes if necessary. In at least one embodiment, lateral ACC is associated with other ADAS applications, such as LC and CW.

[0278] In at least one embodiment, the CACC system uses information from other vehicles, which may be received from other vehicles via network interface 1124 and / or one or more wireless antennas 1126 via a wireless link or indirectly via a network connection (e.g., via the Internet). In at least one embodiment, the direct link may be provided by a vehicle-to-vehicle (“V2V”) communication link, while the indirect link may be provided by an infrastructure-to-vehicle (“I2V”) communication link. Typically, V2V communication provides information about the vehicle immediately preceding it (e.g., a vehicle immediately in front of vehicle 1100 and in the same lane as it), while I2V communication provides information about traffic further ahead. In at least one embodiment, the CACC system may include one or both of the I2V and V2V information sources. In at least one embodiment, given information about vehicles preceding vehicle 1100, the CACC system can be more reliable and has the potential to improve traffic flow smoothness and reduce road congestion.

[0279] In at least one embodiment, the FCW system is designed to warn the driver of danger so that the driver can take corrective action. In at least one embodiment, the FCW system uses a forward-facing camera and / or one or more RADAR sensors 1160, coupled to a dedicated processor, DSP, FPGA, and / or ASIC, electrically coupled to provide driver feedback, such as a display, speaker, and / or vibration components. In at least one embodiment, the FCW system can provide warnings, for example, in the form of audible, visual warnings, vibrations, and / or rapid braking pulses.

[0280] In at least one embodiment, the AEB system detects an impending forward collision with another vehicle or other object and can automatically apply brakes if the driver does not take corrective action within a specified time or distance parameter. In at least one embodiment, the AEB system may use one or more forward-facing cameras and / or one or more RADAR sensors 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC. In at least one embodiment, when the AEB system detects a hazard, it typically first warns the driver to take corrective action to avoid a collision, and if the driver does not take corrective action, the AEB system may automatically apply brakes to attempt to prevent or at least mitigate the effects of the predicted collision. In at least one embodiment, the AEB system may include techniques such as dynamic braking to support and / or brakes for impending collisions.

[0281] In at least one embodiment, when vehicle 1100 crosses lane markings, the LDW system provides visual, auditory, and / or tactile warnings, such as steering wheel or seat vibrations, to alert the driver. In at least one embodiment, the LDW system is inactive when the driver indicates intentional lane departure, such as by activating turn signals. In at least one embodiment, the LDW system may use a front-facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC, which is electrically coupled to provide driver feedback such as a display, speaker, and / or vibration components. In at least one embodiment, the LKA system is a variant of the LDW system. In at least one embodiment, if vehicle 1100 begins to leave the lane, the LKA system provides steering input or braking to correct vehicle 1100.

[0282] In at least one embodiment, the BSW system detects and warns the driver of a vehicle in the blind spot. In at least one embodiment, the BSW system can provide visual, auditory, and / or tactile alerts to indicate that merging or changing lanes is unsafe. In at least one embodiment, the BSW system can provide additional warnings when the driver uses the turn signal. In at least one embodiment, the BSW system can use one or more rear-facing cameras and / or one or more RADAR sensors 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC, electrically coupled to driver feedback, such as a display, speaker, and / or vibration assembly.

[0283] In at least one embodiment, the RCTW system can provide visual, auditory, and / or tactile notifications when an object is detected outside the range of the rear camera while the vehicle 1100 is reversing. In at least one embodiment, the RCTW system includes an AEB system to ensure the applied vehicle brakes to avoid a collision. In at least one embodiment, the RCTW system may use one or more rear-facing RADAR sensors 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC, which are electrically coupled to provide driver feedback such as displays, speakers, and / or vibration components.

[0284] In at least one embodiment, conventional ADAS systems may be prone to generating false alarms, which can be annoying and distracting to the driver, but are generally not catastrophic because conventional ADAS systems alert the driver and allow the driver to determine whether a safe situation truly exists and take appropriate action. In at least one embodiment, in the event of conflicting results, vehicle 1100 itself decides whether to follow the result of the primary computer or the secondary computer (e.g., the first or second controller of controller 1136). For example, in at least one embodiment, ADAS system 1138 may be a backup and / or auxiliary computer for providing perception information to a backup computer rationality module. In at least one embodiment, the backup computer rationality monitor may run redundant software on hardware components to detect faults in perception and dynamic driving tasks. In at least one embodiment, the output from ADAS system 1138 may be provided to a monitoring MCU. In at least one embodiment, if the output from the primary computer and the output from the auxiliary computer conflict, the monitoring MCU decides how to reconcile the conflict to ensure safe operation.

[0285] In at least one embodiment, the master computer may be configured to provide a confidence score to the supervisory MCU to indicate the master computer's confidence in the selected result. In at least one embodiment, if the confidence score exceeds a threshold, the supervisory MCU may follow the master computer's instructions regardless of whether the auxiliary computer provides conflicting or inconsistent results. In at least one embodiment, if the confidence score does not meet the threshold, and if the master computer and the auxiliary computer indicate different results (e.g., conflict), the supervisory MCU may arbitrate between the computers to determine the appropriate result.

[0286] In at least one embodiment, the supervisory MCU may be configured to run a neural network trained and configured to determine, at least in part, the conditions under which the auxiliary computer provides a false alarm based on outputs from a host computer and an auxiliary computer. In at least one embodiment, the neural network in the supervisory MCU may learn when the outputs of the auxiliary computer can be trusted and when they cannot. For example, in at least one embodiment, when the auxiliary computer is a RADAR-based FCW system, the neural network in the supervisory MCU may learn when the FCW system recognizes a metallic object that is not actually dangerous, such as a drain grating or manhole cover that would trigger an alarm. In at least one embodiment, when the auxiliary computer is a camera-based LDW system, the neural network in the supervisory MCU may learn to override the LDW when a cyclist or pedestrian is present and lane departure is actually the safest operation. In at least one embodiment, the supervisory MCU may include at least one of a DLA or GPU suitable for running a neural network with associated memory. In at least one embodiment, the supervisory MCU may include and / or be included as a component of one or more SoC 1104s.

[0287] In at least one embodiment, the ADAS system 1138 may include an auxiliary computer that performs ADAS functions using conventional computer vision rules. In at least one embodiment, the auxiliary computer may use classic computer vision rules (if-then), and the presence of a neural network in the supervisory MCU can improve reliability, security, and performance. For example, in at least one embodiment, diverse implementations and intentional non-identity make the entire system more fault-tolerant, especially for failures caused by software (or software-hardware interface) functionality. For example, in at least one embodiment, if a software vulnerability or bug exists in the software running on the host computer, and different software code running on the auxiliary computer provides consistent overall results, the supervisory MCU can more confidently assume that the overall result is correct and that the vulnerability in the software or hardware on the host computer will not lead to a significant error.

[0288] In at least one embodiment, the output of the ADAS system 1138 may be input to the perception module and / or the dynamic driving task module of the host computer. For example, in at least one embodiment, if the ADAS system 1138 indicates a forward collision warning due to an object directly ahead, the perception block may use this information when identifying the object. In at least one embodiment, as described herein, the assistance computer may have its own neural network trained to reduce the risk of false alarms.

[0289] In at least one embodiment, vehicle 1100 may further include an infotainment SoC 1130 (e.g., an in-vehicle infotainment system (IVI)). Although shown and described as an SoC, in at least one embodiment, the infotainment system SoC 1130 may not be an SoC and may include, but is not limited to, two or more discrete components. In at least one embodiment, the infotainment SoC 1130 may include, but is not limited to, a combination of hardware and software that can be used to provide audio (e.g., music, personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., television, movies, streaming media, etc.), telephone (e.g., hands-free calling), network connectivity (e.g., LTE, WiFi, etc.) and / or information services (e.g., navigation system, rear parking assist, radio data system, vehicle-related information such as fuel level, total coverage distance, brake fuel level, fuel level, door opening / closing, air filter information, etc.) to vehicle 1100. For example, the infotainment SoC 1130 may include a radio, disk player, navigation system, video player, USB and Bluetooth connectivity, automobile, in-vehicle entertainment system, WiFi, steering wheel audio controls, hands-free voice control, head-up display (“HUD”), HMI display 1134, telematics device, control panel (e.g., for controlling and / or interacting with various components, features and / or systems) and / or other components. In at least one embodiment, the infotainment SoC 1130 may be further used to provide information (e.g., visual and / or auditory) to a user of vehicle 1100, such as information from ADAS system 1138, autonomous driving information (such as planned vehicle maneuvers), trajectory, surrounding context information (e.g., intersection information, vehicle information, road information, etc.) and / or other information.

[0290] In at least one embodiment, the infotainment SoC 1130 may include any number and type of GPU functionality. In at least one embodiment, the infotainment SoC 1130 may communicate with other devices, systems, and / or components of the vehicle 1100 via bus 1102. In at least one embodiment, the infotainment SoC 1130 may be coupled to a monitoring MCU, enabling the GPU of the infotainment system to perform some autonomous driving functions in the event of a failure of the main controller 1136 (e.g., the main computer and / or backup computer of the vehicle 1100). In at least one embodiment, the infotainment SoC 1130 may cause the vehicle 1100 to enter a driver-to-safe-stop mode, as described herein.

[0291] In at least one embodiment, vehicle 1100 may further include instrument panel 1132 (e.g., digital instrument panel, electronic instrument panel, digital instrument control panel, etc.). In at least one embodiment, instrument panel 1132 may include, but is not limited to, controllers and / or supercomputers (e.g., discrete controllers or supercomputers). In at least one embodiment, instrument panel 1132 may include, but is not limited to, any number and combination of a set of instruments, such as speedometer, fuel level, oil pressure, tachometer, odometer, turn indicator, shift position indicator, one or more seatbelt warning lights, one or more parking brake warning lights, one or more engine malfunction lights, auxiliary restraint system (e.g., airbag) information, lighting controls, safety system controls, navigation information, etc. In some examples, information may be displayed and / or shared between infotainment SoC 1130 and instrument panel 1132. In at least one embodiment, instrument panel 1132 may be included as part of infotainment SoC 1130, or vice versa.

[0292] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be implemented in the system. Figure 11C The operation is used to infer or predict the operation based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0293] Figure 11D It is based on at least one embodiment in a cloud-based server and Figure 11AA diagram of a system 1176 for communication between autonomous vehicles 1100. In at least one embodiment, system 1176 may include, but is not limited to, one or more servers 1178, one or more networks 1190, and any number and type of vehicles, including vehicle 1100. In at least one embodiment, one or more servers 1178 may include, but is not limited to, multiple GPUs 1184(A)-1184(H) (collectively referred to herein as GPU 1184), PCIe switches 1182(A)-1182(D) (collectively referred to herein as PCIe switch 1182), and / or CPUs 1180(A)-1180(B) (collectively referred to herein as CPU 1180). GPU 1184, CPU 1180, and PCIe switch 1182 may be interconnected with high-speed interconnects, such as, but not limited to, NVLink interface 1188 developed by NVIDIA and / or PCIe connection 1186. In at least one embodiment, the GPU 1184 is connected via NVLink and / or NVSwitchSoC, and the GPU 1184 and PCIe switch 1182 are connected via PCIe interconnect. Although eight GPUs 1184, two CPUs 1180, and four PCIe switches 1182 are shown, this is not intended to be limiting. In at least one embodiment, each of one or more servers 1178 may include, but is not limited to, any combination of any number of GPUs 1184, CPUs 1180, and / or PCIe switches 1182. For example, in at least one embodiment, one or more servers 1178 may each include eight, sixteen, thirty-two, and / or more GPUs 1184.

[0294] In at least one embodiment, one or more servers 1178 may receive image data representing images from vehicles via one or more networks 1190, the images showing unexpected or changed road conditions, such as recently started roadworks. In at least one embodiment, one or more servers 1178 may transmit updated neural network 1192 and / or map information 1194, including but not limited to information about traffic and road conditions, to vehicles via one or more networks 1190. In at least one embodiment, updating the map information 1194 may include, but is not limited to, updating the HD map 1122, such as information about construction sites, potholes, sidewalks, floods, and / or other obstacles. In at least one embodiment, the neural network 1192 and / or map information 1194 may be generated from new training and / or experience represented by data received from any number of vehicles in the context, and / or at least based on training performed in a data center (e.g., using one or more servers 1178 and / or other servers).

[0295] In at least one embodiment, one or more servers 1178 may be used to train a machine learning model (e.g., a neural network) at least in part based on training data. In at least one embodiment, the training data may be generated by the vehicle, and / or may be generated in a simulation (e.g., using a game engine). In at least one embodiment, any amount of training data is labeled (e.g., where the associated neural network benefits from supervised learning) and / or undergoes other preprocessing. In at least one embodiment, no amount of training data is labeled and / or preprocessed (e.g., where the associated neural network does not require supervised learning). In at least one embodiment, once the machine learning model is trained, the machine learning model may be used by the vehicle (e.g., transmitted to the vehicle via one or more networks 1190), and / or the machine learning model may be used by one or more servers 1178 to remotely monitor the vehicle.

[0296] In at least one embodiment, one or more servers 1178 may receive data from the vehicle and apply the data to state-of-the-art real-time neural networks for real-time intelligent inference. In at least one embodiment, one or more servers 1178 may include a deep learning supercomputer and / or a dedicated AI computer powered by one or more GPUs 1184, such as the DGX and DGX Station machines developed by NVIDIA. However, in at least one embodiment, one or more servers 1178 may include a deep learning infrastructure in a data center using CPU power.

[0297] In at least one embodiment, the deep learning infrastructure of one or more servers 1178 may be capable of fast, real-time inference and may use this capability to assess and verify the health of the processor, software, and / or associated hardware in vehicle 1100. For example, in at least one embodiment, the deep learning infrastructure may receive periodic updates from vehicle 1100, such as image sequences and / or objects located by vehicle 1100 in the image sequence (e.g., via computer vision and / or other machine learning object classification techniques). In at least one embodiment, the deep learning infrastructure may run its own neural network to identify objects and compare them with objects identified by vehicle 1100, and if the results do not match and the deep learning infrastructure determines that the AI ​​in vehicle 1100 is malfunctioning, one or more servers 1178 may signal to vehicle 1100 to instruct the fail-safe computer of vehicle 1100 to take control, notify passengers, and complete a safe stopping operation.

[0298] In at least one embodiment, one or more servers 1178 may include one or more GPUs 1184 and one or more programmable inference accelerators (e.g., NVIDIA's TensorRT 3 devices). In at least one embodiment, the combination of GPU-driven servers and inference acceleration enables real-time response. In at least one embodiment, for example, where performance is less critical, servers driven by CPUs, FPGAs, and other processors may be used for inference. In at least one embodiment, hardware architecture 815 is used to execute one or more embodiments. This document incorporates... Figure 8A and / or Figure 8B Provide details about the 815 hardware architecture.

[0299] Computer System

[0300] Figure 12 This is a block diagram illustrating an exemplary computer system according to at least one embodiment. The exemplary computer system may be a system of interconnected devices and components, a system-on-a-chip (SoC), or some combination thereof formed with a processor, which may include an execution unit to execute instructions. In at least one embodiment, according to this disclosure, such as the embodiments described herein, computer system 1200 may include, but is not limited to, components such as processor 1202, whose execution unit includes logic to execute algorithms for process data. In at least one embodiment, computer system 1200 may include a processor, such as the PENTIUM® processor family or Xeon processor, available from Intel Corporation of Santa Clara, California. TM Itanium®, XScale TM and / or StrongARM TM The system may use an Intel® Core™ or Intel® Nervana™ microprocessor, although other systems (including PCs, engineering workstations, set-top boxes, etc.) with other microprocessors may also be used. In at least one embodiment, the computer system 1200 may execute a version of the Windows operating system available from Microsoft Corporation of Redmond, Wash., although other operating systems (such as UNIX and Linux), embedded software, and / or graphical user interfaces may also be used.

[0301] The embodiments can be used in other devices, such as handheld devices and embedded applications. Some examples of handheld devices include cellular phones, Internet Protocol (IP) devices, digital cameras, personal digital assistants (“PDAs”), and handheld PCs. In at least one embodiment, the embedded application may include a microcontroller, a digital signal processor (“DSP”), a system-on-a-chip (SoC), a network computer (“NetPC”), a set-top box, a network hub, a wide area network (“WAN”) switch, or any other system that can execute one or more instructions according to at least one embodiment.

[0302] In at least one embodiment, the computer system 1200 may include, but is not limited to, a processor 1202, which may include, but is not limited to, one or more execution units 1208, to perform machine learning model training and / or inference according to the techniques described herein. In at least one embodiment, the computer system 1200 is a single-processor desktop or server system, but in another embodiment, the computer system 1200 may be a multiprocessor system. In at least one embodiment, the processor 1202 may include, but is not limited to, a Complex Instruction Set Computer (“CISC”) microprocessor, a Reduced Instruction Set Computing (“RISC”) microprocessor, a Very Long Instruction Word (“VLIW”) microprocessor, a processor implementing instruction set combination, or any other processor device, such as a digital signal processor. In at least one embodiment, the processor 1202 may be coupled to a processor bus 1210, which may transmit data signals between the processor 1202 and other components in the computer system 1200.

[0303] In at least one embodiment, processor 1202 may include, but is not limited to, a Level 1 (“L1”) internal cache memory (“cache”) 1204. In at least one embodiment, processor 1202 may have a single internal cache or multiple levels of internal cache. In at least one embodiment, the cache memory may reside external to processor 1202. Depending on specific implementation and requirements, other embodiments may also include a combination of internal and external caches. In at least one embodiment, register file 1206 may store different types of data in various registers, including but not limited to integer registers, floating-point registers, status registers, and instruction pointer registers.

[0304] In at least one embodiment, an execution unit 1208, including but not limited to logic for performing integer and floating-point operations, is also located within the processor 1202. In at least one embodiment, the processor 1202 may also include a microcode (“ucode”) read-only memory (“ROM”) for storing microcode of certain macro instructions. In at least one embodiment, the execution unit 1208 may include logic for processing a packaged instruction set 1209. In at least one embodiment, by including the packaged instruction set 1209 in the instruction set of a general-purpose processor, along with the associated circuitry for executing the instructions, the packaged data in the processor 1202 can be used to perform operations used by numerous multimedia applications. In one or more embodiments, the execution of numerous multimedia applications can be accelerated and performed more efficiently by using the full width of the processor’s data bus to perform operations on the packaged data, which may eliminate the need to transfer smaller data units on the processor’s data bus to perform one or more operations on one data element at a time.

[0305] In at least one embodiment, execution unit 1208 may also be used in a microcontroller, embedded processor, graphics device, DSP, and other types of logic circuitry. In at least one embodiment, computer system 1200 may include, but is not limited to, memory 1220. In at least one embodiment, memory 1220 may be a dynamic random access memory (“DRAM”) device, a static random access memory (“SRAM”) device, a flash memory device, or another storage device. In at least one embodiment, memory 1220 may store instructions 1219 and / or data 1221 represented by data signals that can be executed by processor 1202.

[0306] In at least one embodiment, the system logic chip may be coupled to processor bus 1210 and memory 1220. In at least one embodiment, the system logic chip may include, but is not limited to, a memory controller hub (“MCH”) 1216, and processor 1202 may communicate with MCH 1216 via processor bus 1210. In at least one embodiment, MCH 1216 may provide a high-bandwidth memory path 1218 to memory 1220 for instruction and data storage, as well as for storage of graphics commands, data, and textures. In at least one embodiment, MCH 1216 may initiate data signals between processor 1202, memory 1220, and other components in computer system 1200, and bridge data signals between processor bus 1210, memory 1220, and system I / O interface 1222. In at least one embodiment, the system logic chip may provide a graphics port for coupling to a graphics controller. In at least one embodiment, MCH 1216 may be coupled to memory 1220 via high-bandwidth memory path 1218, and graphics / video card 1212 may be coupled to MCH 1216 via Accelerated Graphics Port (“AGP”) interconnect 1214.

[0307] In at least one embodiment, the computer system 1200 may use the system I / O interface 1222 as a proprietary hub interface bus to couple the MCH 1216 to the I / O controller hub (“ICH”) 1230. In at least one embodiment, the ICH 1230 may provide direct connectivity to certain I / O devices via a local I / O bus. In at least one embodiment, the local I / O bus may include, but is not limited to, a high-speed I / O bus for connecting peripheral devices to the memory 1220, chipset, and processor 1202. Examples may include, but are not limited to, an audio controller 1229, a firmware hub (“Flash BIOS”) 1228, a wireless transceiver 1226, a data storage 1224, a conventional I / O controller 1223 including a user input and keyboard interface, a serial expansion port 1227 (e.g., a Universal Serial Bus (USB) port), and a network controller 1234. In at least one embodiment, the data storage 1224 may include a hard disk drive, a floppy disk drive, a CD-ROM device, a flash memory device, or other mass storage device.

[0308] In at least one embodiment, Figure 12 A system including interconnected hardware devices or "chips" is shown, while in other embodiments, Figure 12 The SoC can be shown. In at least one embodiment, Figure 12The devices shown can be interconnected with proprietary interconnects, standardized interconnects (e.g., PCIe), or some combination thereof. In at least one embodiment, one or more components of the computer system 1200 are interconnected using a Compute Fast Link (CXL) interconnect.

[0309] The inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be... Figure 12 Used in systems for reasoning or predicting operations based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0310] Figure 13 This is a block diagram illustrating an electronic device 1300 for utilizing a processor 1310 according to at least one embodiment. In at least one embodiment, the electronic device 1300 may be, for example, but not limited to, a laptop computer, tower server, rack server, blade server, laptop computer, desktop computer, tablet computer, mobile device, telephone, embedded computer, or any other suitable electronic device.

[0311] In at least one embodiment, the electronic device 1300 may include, but is not limited to, a processor 1310 communicatively coupled to any suitable number or type of components, peripherals, modules, or devices. In at least one embodiment, the processor 1310 is coupled using a bus or interface, such as I... 2 C-bus, System Management Bus (“SMBus”), Low Pin Count (LPC) bus, Serial Peripheral Interface (“SPI”), High Definition Audio (“HDA”) bus, Serial Advanced Technology Accessory (“SATA”) bus, Universal Serial Bus (“USB”) (versions 1, 2, 3, etc.), or Universal Asynchronous Receiver / Transmitter (“UART”) bus. In at least one embodiment, Figure 13 The system shown includes interconnected hardware devices or "chips," while in other embodiments, Figure 13 An exemplary SoC can be shown. In at least one embodiment, Figure 13 The device shown can be interconnected with proprietary interconnects, standardized interconnects (e.g., PCIe), or some combination thereof. In at least one embodiment, Figure 13 One or more components are interconnected using Computational Fast Link (CXL) interconnects.

[0312] In at least one embodiment, Figure 13This may include a display 1324, a touchscreen 1325, a touchpad 1330, a near-field communication unit (“NFC”) 1345, a sensor hub 1340, a thermal sensor 1346, a fast chipset (“EC”) 1335, a trusted platform module (“TPM”) 1338, a BIOS / firmware / flash (“BIOS, FW Flash”) 1322, a DSP 1360, a drive 1320 (e.g., a solid-state drive (“SSD”) or a hard disk drive (“HDD”)), a wireless local area network unit (“WLAN”) 1350, a Bluetooth unit 1352, a wireless wide area network unit (“WWAN”) 1356, a global positioning system (GPS) unit 1355, a camera (“USB 3.0 camera”) 1354 (e.g., a USB 3.0 camera), and / or a low-power double data rate (“LPDDR”) memory unit (“LPDDR3”) 1315 implemented in, for example, the LPDDR3 standard. These components may each be implemented in any suitable manner.

[0313] In at least one embodiment, other components may be communicatively coupled to processor 1310 via the components described herein. In at least one embodiment, accelerometer 1341, context light sensor (“ALS”) 1342, compass 1343, and gyroscope 1344 may be communicatively coupled to sensor hub 1340. In at least one embodiment, thermal sensor 1339, fan 1337, keyboard 1336, and touchpad 1330 may be communicatively coupled to EC 1335. In at least one embodiment, speaker 1363, earphone 1364, and microphone (“mic”) 1365 may be communicatively coupled to audio unit (“audio codec and Class D amplifier”) 1362, which in turn may be communicatively coupled to DSP 1360. In at least one embodiment, audio unit 1362 may include, for example, but not limited to, audio encoder / decoder (“codec”) and Class D amplifier. In at least one embodiment, SIM card (“SIM”) 1357 may be communicatively coupled to WWAN unit 1356. In at least one embodiment, components such as WLAN unit 1350, Bluetooth unit 1352, and WWAN unit 1356 can be implemented as next-generation form factor (NGFF).

[0314] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be implemented in the system. Figure 13It is used in the context of reasoning or predicting operations based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0315] Figure 14 A computer system 1400 according to at least one embodiment is shown. In at least one embodiment, the computer system 1400 is configured to implement various processes and methods described throughout this disclosure.

[0316] In at least one embodiment, the computer system 1400 includes, but is not limited to, at least one central processing unit (“CPU”) 1402 connected to a communication bus 1410 implemented using any suitable protocol, such as PCI (“Peripheral Device Interconnect”), Peripheral Component Interconnect Express (“PCI-Express”), AGP (“Accelerated Graphics Port”), HyperTransport, or any other bus or point-to-point communication protocol. In at least one embodiment, the computer system 1400 includes, but is not limited to, main memory 1404 and control logic (e.g., implemented in hardware, software, or a combination thereof), and data may be stored in main memory 1404 in the form of random access memory (“RAM”). In at least one embodiment, a network interface subsystem (“Network Interface”) 1422 provides an interface to other computing devices and networks for receiving data using the computer system 1400 and transferring data to other systems.

[0317] In at least one embodiment, the computer system 1400 includes, but is not limited to, an input device 1408, a parallel processing system 1412, and a display device 1406, which may be implemented using conventional cathode ray tube (“CRT”), liquid crystal display (“LCD”), light-emitting diode (“LED”) display, plasma display, or other suitable display technologies. In at least one embodiment, user input is received from the input device 1408 (such as a keyboard, mouse, touchpad, microphone, etc.). In at least one embodiment, each of the modules described herein may reside on a single semiconductor platform to form the processing system.

[0318] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be implemented in the system. Figure 14 It is used to perform inference or prediction operations based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architecture or neural network use cases described herein.

[0319] Figure 15 A computer system 1500 according to at least one embodiment is illustrated. In at least one embodiment, the computer system 1500 includes, but is not limited to, a computer 1510 and a USB flash drive 1520. In at least one embodiment, the computer 1510 may include, but is not limited to, any number and type of processors (not shown) and memory (not shown). In at least one embodiment, the computer 1510 includes, but is not limited to, a server, a cloud instance, a laptop computer, and a desktop computer.

[0320] In at least one embodiment, the USB flash drive 1520 includes, but is not limited to, a processing unit 1530, a USB interface 1540, and USB interface logic 1550. In at least one embodiment, the processing unit 1530 can be any instruction execution system, apparatus, or device capable of executing instructions. In at least one embodiment, the processing unit 1530 can include, but is not limited to, any number and type of processing cores (not shown). In at least one embodiment, the processing unit 1530 includes an application-specific integrated circuit (“ASIC”) optimized to perform any number and type of operations associated with machine learning. For example, in at least one embodiment, the processing unit 1530 is a tensor processing unit (“TPC”) optimized to perform machine learning inference operations. In at least one embodiment, the processing unit 1530 is a vision processing unit (“VPU”) optimized to perform machine vision and machine learning inference operations.

[0321] In at least one embodiment, the USB interface 1540 can be any type of USB connector or USB receptacle. For example, in at least one embodiment, the USB interface 1540 is a USB 3.0 Type-C receptacle for data and power. In at least one embodiment, the USB interface 1540 is a USB 3.0 Type-A connector. In at least one embodiment, the USB interface logic 1550 may include any amount and type of logic enabling the processing unit 1530 to connect to a device (e.g., computer 1510) via the USB interface 1540.

[0322] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be implemented in the system. Figure 15 In use, at least in part, the operation is based on weight parameters, neural network functions and / or architectures computed using neural network training operations, or neural network use cases described herein to infer or predict operations.

[0323] Figure 16A An exemplary architecture is shown in which multiple GPUs 1610(1)-1610(N) are communicatively coupled to multiple multi-core processors 1605(1)-1605(M) via high-speed links 1640(1)-1640(N) (e.g., bus / point-to-point interconnect, etc.). In at least one embodiment, the high-speed links 1640(1)-1640(N) support communication throughput of 4GB / s, 30GB / s, 80GB / s, or higher. In at least one embodiment, various interconnect protocols may be used, including but not limited to PCIe 4.0 or 5.0 and NVLink 2.0. In the various figures, “N” and “M” represent positive integers, the values ​​of which may vary from figure to figure.

[0324] Furthermore, in one embodiment, two or more GPUs 1610 are interconnected via high-speed links 1629(1)-1629(2), which can be implemented using a protocol / link similar to or different from that used for high-speed links 1640(1)-1640(N). Similarly, two or more multi-core processors 1605 can be connected via high-speed link 1628, which can be a symmetric multiprocessor (SMP) bus operating at speeds of 20GB / s, 30GB / s, 120GB / s, or higher. Alternatively, similar protocols / links (e.g., via a common interconnect structure) can be used. Figure 16A This shows all communication between the various system components.

[0325] In one embodiment, each multi-core processor 1605 is communicatively coupled to processor memory 1601(1)-1601(M) via memory interconnects 1626(1)-1626(M), and each GPU 1610(1)-1610(N) is communicatively coupled to GPU memory 1620(1)-1620(N) via GPU memory interconnects 1650(1)-1650(N). In at least one embodiment, memory interconnects 1626 and 1650 may utilize similar or different memory access technologies. By way of example and not limitation, processor memory 1601(1)-1601(M) and GPU memory 1620 may be volatile memory, such as dynamic random access memory (DRAM) (including stacked DRAM), graphics DDR SDRAM (GDDR) (e.g., GDDR5, GDDR6), or high bandwidth memory (HBM), and / or may be non-volatile memory, such as 3D XPoint or Nano-Ram. In at least one embodiment, some portions of the processor memory 1601 may be volatile memory, while other portions may be non-volatile memory (e.g., using a two-level memory (2LM) hierarchy).

[0326] As described herein, although the various multi-core processors 1605 and GPUs 1610 can be physically coupled to specific memories 1601 and 1620 respectively, and / or can implement a unified memory architecture, in which the virtual system address space (also known as the “effective address” space) is distributed among the various physical memories. For example, processor memories 1601(1)-1601(M) can each contain 64 GB of system memory address space, and GPU memories 1620(1)-1620(N) can each contain 32 GB of system memory address space, resulting in a total addressable memory size of 256 GB when M = 2 and N = 4. N and M may also be other values.

[0327] Figure 16B Additional details are shown regarding the interconnection between a multi-core processor 1607 and a graphics acceleration module 1646 according to an exemplary embodiment. In at least one embodiment, the graphics acceleration module 1646 may include one or more GPU chips integrated on a line card coupled to the processor 1607 via a high-speed link 1640 (e.g., PCIe bus, NVLink, etc.). In at least one embodiment, the graphics acceleration module 1646 may optionally be integrated on a package or chip having the processor 1607.

[0328] In at least one embodiment, the processor 1607 includes a plurality of cores 1660A-1660D, each core having a translation back cover (“TLB”) 1661A-1661D and one or more caches 1662A-1662D. In at least one embodiment, the cores 1660A-1660D may include various other components (not shown) for executing instructions and processing data. In at least one embodiment, the caches 1662A-1662D may include level 1 (L1) and level 2 (L2) caches. Furthermore, one or more shared caches 1656 may be included in the caches 1662A-1662D and shared by the respective groups of cores 1660A-1660D. For example, one embodiment of the processor 1607 includes 24 cores, each core having its own L1 cache, twelve shared L2 caches, and twelve shared L3 caches. In this embodiment, two adjacent cores share one or more L2 and L3 caches. In at least one embodiment, the processor 1607 and the graphics acceleration module 1646 are connected to a system memory 1614, which may include... Figure 16A The processor memory in the memory is 1601(1)-1601(M).

[0329] In at least one embodiment, consistency of data and instructions stored in the various caches 1662A-1662D, 1656 and system memory 1614 is maintained via inter-core communication through the consistency bus 1664. In at least one embodiment, for example, each cache may have associated cache consistency logic / circuit to communicate via the consistency bus 1664 in response to the detection of a read or write to a particular cache line. In at least one embodiment, a cache snooping protocol is implemented via the consistency bus 1664 to snoop on cache accesses.

[0330] In at least one embodiment, proxy circuitry 1625 communicatively couples graphics acceleration module 1646 to coherence bus 1664, thereby allowing graphics acceleration module 1646 to participate in cache coherence protocols as a peer of cores 1660A-1660D. Specifically, in at least one embodiment, interface 1635 provides connectivity to proxy circuitry 1625 via high-speed link 1640, and interface 1637 connects graphics acceleration module 1646 to high-speed link 1640.

[0331] In at least one embodiment, the accelerator integrated circuit 1636 represents multiple graphics processing engines 1631(1)-1631(N) of the graphics acceleration module providing cache management, memory access, context management, and interrupt management services. In at least one embodiment, each of the graphics processing engines 1631(1)-1631(N) may include a separate graphics processing unit (GPU). In at least one embodiment, the graphics processing engines 1631(1)-1631(N) may optionally include different types of graphics processing engines within the GPU, such as graphics execution units, media processing engines (e.g., video encoders / decoders), samplers, and blit engines. In at least one embodiment, the graphics acceleration module 1646 may be a GPU having multiple graphics processing engines 1631(1)-1631(N), or the graphics processing engines 1631(1)-1631(N) may be individual GPUs integrated on a general-purpose package, line card, or chip.

[0332] In at least one embodiment, the accelerator integrated circuit 1636 includes a memory management unit (MMU) 1639 for performing various memory management functions, such as virtual-to-physical memory translation (also known as effective-to-real memory translation), and a memory access protocol for accessing system memory 1614. In at least one embodiment, the MMU 1639 may also include a translation back buffer (“TLB”) (not shown) for caching virtual / effective-to-physical / real address translations. In at least one embodiment, cache 1638 may store commands and data for efficient access by graphics processing engines 1631(1)-1631(N). In at least one embodiment, a fetch unit 1644 may be used to keep data stored in cache 1638 and graphics memory 1633(1)-1633(M) consistent with core caches 1662A-1662D, 1656 and system memory 1614. As previously described, this task can be accomplished via proxy circuitry 1625 representing cache 1638 and graphics memory 1633(1)-1633(M) (e.g., sending updates related to the modification / access of cache lines on processor caches 1662A-1662D, 1656 to cache 1638 and receiving updates from cache 1638).

[0333] In at least one embodiment, a set of registers 1645 stores context data of threads executed by graphics processing engines 1631(1)-1631(N), and context management circuitry 1648 manages the thread context. For example, context management circuitry 1648 may perform save and restore operations to save and restore the context of individual threads during context switching (e.g., saving the first thread and storing the second thread so that the second thread can be executed by the graphics processing engine). For example, during context switching, context management circuitry 1648 may store the current register value in a designated area of ​​memory (e.g., identified by a context pointer). The register value can then be restored when returning to the context. In at least one embodiment, interrupt management circuitry 1647 receives and processes interrupts received from system devices.

[0334] In one implementation, MMU 1639 translates virtual / effective addresses from graphics processing engine 1631 into real / physical addresses in system memory 1614. In at least one embodiment, accelerator integrated circuit 1636 supports multiple (e.g., 4, 8, 16) graphics accelerator modules 1646 and / or other accelerator devices. In at least one embodiment, graphics accelerator module 1646 may be dedicated to a single application executing on processor 1607, or may be shared among multiple applications. In at least one embodiment, a virtualized graphics execution context is presented, wherein resources of graphics processing engines 1631(1)-1631(N) are shared with multiple applications or virtual machines (VMs). In at least one embodiment, resources may be subdivided into “slices” based on processing requirements and priorities associated with VMs and / or applications, which are allocated to different VMs and / or applications.

[0335] In at least one embodiment, the accelerator integrated circuit 1636 acts as a bridge to the system of the graphics acceleration module 1646 and provides address translation and system memory caching services. Additionally, in at least one embodiment, the accelerator integrated circuit 1636 can provide virtualization facilities for the host processor to manage the virtualization, interrupt, and memory management of the graphics processing engines 1631(1)-1631(N).

[0336] In at least one embodiment, since the hardware resources of the graphics processing engines 1631(1)-1631(N) are explicitly mapped to the real address space seen by the host processor 1607, any host processor can directly address these resources using valid address values. In at least one embodiment, a function of the accelerator integrated circuit 1636 is to physically separate the graphics processing engines 1631(1)-1631(N) so that they appear as independent units to the system.

[0337] In at least one embodiment, one or more graphics memories 1633(1)-1633(M) are coupled to each graphics processing engine 1631(1)-1631(N), and N = M. In at least one embodiment, the graphics memories 1633(1)-1633(M) store instructions and data processed by each graphics processing engine 1631(1)-1631(N). In at least one embodiment, the graphics memories 1633(1)-1633(M) may be volatile memory, such as DRAM (including stacked DRAM), GDDR memory (e.g., GDDR5, GDDR6), or HBM, and / or may be non-volatile memory, such as 3DXPoint or Nano-RAM.

[0338] In one embodiment, to reduce data traffic on the high-speed link 1640, a biasing technique is used to ensure that the data stored in graphics memory 1633(1)-1633(M) is the data most frequently used by graphics processing engines 1631(1)-1631(N), and preferably data that cores 1660A-1660D do not use (or at least do not use frequently). Similarly, in at least one embodiment, the biasing mechanism attempts to keep the data needed by the cores (and preferably not graphics processing engines 1631(-1)-1631(N)) in caches 1662A-1662D, 1656 and system memory 1614.

[0339] Figure 16C Another exemplary embodiment is shown, wherein the accelerator integrated circuit 1636 is integrated within the processor 1607. In this embodiment, the graphics processing engines 1631(1)-1631(N) communicate directly with the accelerator integrated circuit 1636 via interface 1637 and interface 1635 (which may also be any form of bus or interface protocol) through a high-speed link 1640. In at least one embodiment, the accelerator integrated circuit 1636 can perform operations related to... Figure 16B The described operation is similar. However, due to its close proximity to the coherence bus 1664 and caches 1662A-1662D, 1656, it may have higher throughput. One embodiment supports different programming models, including a dedicated process programming model (without graphics acceleration module virtualization) and a shared programming model (with virtualization), which may include a programming model controlled by accelerator integrated circuit 1636 and a programming model controlled by graphics acceleration module 1646.

[0340] In at least one embodiment, graphics processing engines 1631(1)-1631(N) are dedicated to a single application or process under a single operating system. In at least one embodiment, a single application can funnel requests from other applications to graphics processing engines 1631(1)-1631(N), thereby providing virtualization within a VM / partition.

[0341] In at least one embodiment, graphics processing engines 1631(1)-1631(N) can be shared by multiple VM / application partitions. In at least one embodiment, the shared model can use a hypervisor to virtualize graphics processing engines 1631(1)-1631(N) to allow each operating system to access them. In at least one embodiment, for a single-partition system without a hypervisor, the operating system owns graphics processing engines 1631(1)-1631(N). In at least one embodiment, the operating system can virtualize graphics processing engines 1631(1)-1631(N) to provide access to each process or application.

[0342] In at least one embodiment, the graphics acceleration module 1646 or the individual graphics processing engine 1631(1)-1631(N) uses a process handle to select a process element. In at least one embodiment, the process element is stored in system memory 1614 and can be addressed using the effective address to real address translation techniques described herein. In at least one embodiment, the process handle may be an implementation-specific value provided to the host process when registering its context with the graphics processing engine 1631(1)-1631(N) (i.e., invoking system software to add the process element to the process element linked list). In at least one embodiment, the lower 16 bits of the process handle may be the offset of the process element in the process element linked list.

[0343] Figure 16D An exemplary accelerator integration slice 1690 is illustrated. In at least one embodiment, a "slice" includes a designated portion of the processing resources of an accelerator integrated circuit 1636. In at least one embodiment, the application is an effective address space 1682 in system memory 1614, which stores process element 1683. In at least one embodiment, process element 1683 is stored in response to a GPU call 1681 from an application 1680 executing on processor 1607. In at least one embodiment, process element 1683 contains the process state of the corresponding application 1680. In one embodiment, a job descriptor (WD) 1684 contained in process element 1683 may be a single job requested by the application, or it may contain a pointer to a job queue. In at least one embodiment, WD 1684 is a pointer to a job request queue in the effective address space 1682 of the application.

[0344] In at least one embodiment, the graphics acceleration module 1646 and / or the various graphics processing engines 1631(1)-1631(N) may be shared by all processes or a subset of processes in the system. In at least one embodiment, infrastructure may be included for setting process states and sending WD 1684 to the graphics acceleration module 1646 to start a job in a virtualization context.

[0345] In at least one embodiment, the dedicated process programming model is implementation-specific. In at least one embodiment, in this model, a single process owns either the graphics acceleration module 1646 or an individual graphics processing engine 1631. In at least one embodiment, when the graphics acceleration module 1646 is owned by a single process, the hypervisor initializes the accelerator integrated circuit for the owned partition, and when the graphics acceleration module 1646 is assigned, the operating system initializes the accelerator integrated circuit 1636 for the owned process.

[0346] In at least one embodiment, during operation, the WD acquisition unit 1691 in the accelerator integration slice 1690 acquires the next WD 1684, which includes instructions for work to be performed by one or more graphics processing engines of the graphics acceleration module 1646. In at least one embodiment, data from the WD 1684 may be stored in register 1645 and used by the MMU 1639, interrupt management circuitry 1647, and / or context management circuitry 1648, as shown. For example, one embodiment of the MMU 1639 includes segment / page roaming circuitry for accessing segment / page tables 1686 within the OS virtual address space 1685. In at least one embodiment, the interrupt management circuitry 1647 may process an interrupt event 1692 received from the graphics acceleration module 1646. In at least one embodiment, when performing graphics operations, a valid address 1693 generated by graphics processing engines 1631(1)-1631(N) is translated into a real address by the MMU 1639.

[0347] In one embodiment, register 1645 is copied for each graphics processing engine 1631(1)-1631(N) and / or graphics acceleration module 1646, and register 1645 may be initialized by a hypervisor or operating system. In at least one embodiment, each of these copied registers may be included in accelerator integration slice 1690. Exemplary registers that may be initialized by a hypervisor are shown in Table 1.

[0348]

[0349] Table 2 shows exemplary registers that can be initialized by the operating system.

[0350]

[0351] In at least one embodiment, each WD 1684 is specific to a particular graphics acceleration module 1646 and / or graphics processing engine 1631(1)-1631(N). In at least one embodiment, it contains all the information required for the graphics processing engine 1631(1)-1631(N) to complete its work, or it may be a pointer to a memory location where the application has set up a command queue for the work to be completed.

[0352] Figure 16E Additional details of an exemplary embodiment of the shared model are shown. This embodiment includes a hypervisor real address space 1698, in which a list of process elements 1699 is stored. In at least one embodiment, the hypervisor real address space 1698 can be accessed via a hypervisor 1696, which virtualizes the graphics acceleration module engine for an operating system 1695.

[0353] In at least one embodiment, the shared programming model allows all processes or subsets of processes from all partitions or subsets of partitions in the system to use the graphics acceleration module 1646. In at least one embodiment, there are two programming models in which the graphics acceleration module 1646 is shared by multiple processes and partitions, namely, time-slice sharing and graphics-oriented sharing.

[0354] In at least one embodiment, in this model, the hypervisor 1696 owns the graphics acceleration module 1646 and makes its functionality available to all operating systems 1695. In at least one embodiment, for the graphics acceleration module 1646 to support virtualization through the hypervisor 1696, the graphics acceleration module 1646 may comply with certain requirements, such as (1) the job requests of the application must be autonomous (i.e., no need to maintain state between jobs), or the graphics acceleration module 1646 must provide a context saving and restoring mechanism, (2) the graphics acceleration module 1646 guarantees that the job requests of the application are completed within a specified amount of time, including any conversion errors, or the graphics acceleration module 1646 provides the ability to preempt job processing, and (3) when operating in a directed shared programming model, fairness between the processes of the graphics acceleration module 1646 must be ensured.

[0355] In at least one embodiment, application 1680 needs to make operating system 1695 system calls using the graphics acceleration module type, working descriptor (WD), authority mask register (AMR) value, and context save / restore region pointer (CSRP). In at least one embodiment, the graphics acceleration module type describes the target acceleration function for the system call. In at least one embodiment, the graphics acceleration module type can be a system-specific value. In at least one embodiment, the WD is specifically formatted for graphics acceleration module 1646 and can take the form of graphics acceleration module 1646 commands, valid address pointers to user-defined structures, valid address pointers to command queues, or any other data structure describing the work to be performed by graphics acceleration module 1646.

[0356] In at least one embodiment, the AMR value is the AMR state for the current process. In at least one embodiment, the value passed to the operating system is similar to that of the application that sets the AMR. In at least one embodiment, if the implementation of the accelerator integrated circuit 1636 (not shown) and the graphics acceleration module 1646 does not support the User Rights Mask Overwrite Register (UAMOR), the operating system may apply the current UAMOR value to the AMR value before passing the AMR in the hypervisor call. In at least one embodiment, the hypervisor 1696 may selectively apply the current Rights Mask Overwrite Register (AMOR) value before placing the AMR into the process element 1683. In at least one embodiment, CSRP is one of the registers 1645 that contains the effective address of a region in the effective address space 1682 of the application for the graphics acceleration module 1646 to save and restore the context state. In at least one embodiment, this pointer is optional if it is not necessary to save the state between jobs or when a job is preempted. In at least one embodiment, the context save / restore region may be fixed system memory.

[0357] Upon receiving a system call, the operating system 1695 can verify that the application 1680 has been registered and granted permission to use the graphics acceleration module 1646. Then, in at least one embodiment, the operating system 1695 uses the information shown in Table 3 to invoke the hypervisor 1696.

[0358]

[0359] In at least one embodiment, upon receiving a hypervisor call, hypervisor 1696 verifies that operating system 1695 has been registered and granted permission to use graphics acceleration module 1646. Then, in at least one embodiment, hypervisor 1696 adds process element 1683 to a linked list of process elements of the corresponding graphics acceleration module 1646 type. In at least one embodiment, the process element may include the information shown in Table 4.

[0360]

[0361] In at least one embodiment, the hypervisor initializes multiple accelerator integration slice 1690 registers 1645.

[0362] like Figure 16FAs shown, in at least one embodiment, a unified memory is used, which is addressable via a common virtual memory address space for accessing physical processor memories 1601(1)-1601(N) and GPU memories 1620(1)-1620(N). In this implementation, operations performed on GPUs 1610(1)-1610(N) utilize the same virtual / effective memory address space to access processor memories 1601(1)-1601(M) and vice versa, thereby simplifying programmability. In at least one embodiment, a first portion of the virtual / effective address space is allocated to processor memory 1601(1), a second portion to second processor memory 1601(N), a third portion to GPU memory 1620(1), and so on. In at least one embodiment, the entire virtual / effective memory space (sometimes referred to as the effective address space) is thus distributed across each of processor memory 1601 and GPU memory 1620, thereby allowing any processor or GPU to access that memory using a virtual address mapped to any physical memory.

[0363] In one embodiment, the bias / coherence management circuitry 1694A-1694E within one or more MMUs 1639A-1639E ensures cache coherence between one or more host processors (e.g., 1605) and the cache of the GPU 1610, and implements biasing techniques to indicate the physical memory in which certain types of data should be stored. In at least one embodiment, although in Figure 16F Several instances of the bias / coherence management circuitry 1694A-1694E are shown, but the bias / coherence circuitry can be implemented within the MMU of one or more host processors 1605 and / or within the accelerator integrated circuit 1636.

[0364] One embodiment allows GPU memory 1620 to be mapped as part of system memory and accessed using shared virtual memory (SVM) technology without suffering the performance drawbacks associated with full system cache coherence. In at least one embodiment, the ability to access GPU memory 1620 as system memory without the heavy overhead of cache coherence provides a favorable operational context for GPU offloading. In at least one embodiment, this arrangement allows the host processor 1605 to software-set operands and access computation results without the overhead of conventional I / O DMA data copying. In at least one embodiment, such conventional copying includes driver calls, interrupts, and memory-mapped I / O (MMIO) accesses, all of which are less efficient than simple memory accesses. In at least one embodiment, the ability to access GPU memory 1620 without cache coherence overhead may be critical to the execution time of offloaded computations. In at least one embodiment, for example, in cases with high streaming write memory traffic, cache coherence overhead can significantly reduce the effective write bandwidth seen by GPU 1610. In at least one embodiment, the efficiency of operand setting, the efficiency of result access, and the efficiency of GPU computation may play a role in determining the effectiveness of GPU offloading.

[0365] In at least one embodiment, the selection of GPU bias and host processor bias is driven by a bias tracker data structure. In at least one embodiment, for example, a bias table can be used, which may be a page-granular structure (e.g., controlled at the memory page level) comprising one or two bits of memory pages attached to each GPU. In at least one embodiment, with or without a bias cache (e.g., for caching frequently / recently used entries in the bias table) in GPU 1610, the bias table can be implemented across one or more stolen memory ranges of GPU memory 1620. Alternatively, in at least one embodiment, the entire bias table can be maintained within the GPU.

[0366] In at least one embodiment, prior to actual access to GPU memory, an access to the bias table entry associated with each access to GPU-attached memory 1620 is performed, resulting in the following operations: In at least one embodiment, a local request from GPU 1610 to find its page in the GPU bias is forwarded directly to the corresponding GPU memory 1620. In at least one embodiment, a local request from the GPU to find its page in the host bias is forwarded to processor 1605 (e.g., via the high-speed link described herein). In at least one embodiment, a request from processor 1605 to find the requested page in the host processor bias completes a request similar to a normal memory read. Alternatively, a request for a page pointing to the GPU bias can be forwarded to GPU 1610. In at least one embodiment, if the GPU is not currently using the page, the GPU may subsequently migrate the page to the host processor bias. In at least one embodiment, the page bias state can be changed through a software-based mechanism, a hardware-assisted software mechanism, or, in limited cases, a purely hardware-based mechanism.

[0367] In at least one embodiment, a mechanism for changing the bias state employs an API call (e.g., OpenCL), which subsequently invokes the GPU's device driver. The device driver then sends a message (or enqueues a command descriptor) to the GPU, instructing the GPU to change the bias state and, in some migration, performs a cache refresh operation on the host. In at least one embodiment, the cache refresh operation is used for migration from the host processor 1605 bias to the GPU bias, but not for the reverse migration.

[0368] In one embodiment, cache coherence is maintained by temporarily rendering GPU bias pages that the host processor 1605 cannot cache. In at least one embodiment, to access these pages, the processor 1605 may request access from the GPU 1610, which may or may not immediately grant access. Therefore, in at least one embodiment, to reduce communication between the processor 1605 and the GPU 1610, it is beneficial to ensure that the GPU bias pages are pages needed by the GPU, not those needed by the host processor 1605, and vice versa.

[0369] One or more hardware structures 815 are used to execute one or more embodiments. This document may combine... Figure 8A and / or Figure 8B Provide details about one or more hardware architectures 815.

[0370] Figure 17Exemplary integrated circuits and associated graphics processors according to various embodiments described herein are illustrated, which may be manufactured using one or more IP cores. In addition to the illustrations, at least one embodiment may include other logic and circuitry, including additional graphics processors / cores, peripheral interface controllers, or general-purpose processor cores.

[0371] Figure 17 This is a block diagram illustrating an exemplary system on a chip integrated circuit 1700 that can be fabricated using one or more IP cores according to at least one embodiment. In at least one embodiment, the integrated circuit 1700 includes one or more application processors 1705 (e.g., CPUs), at least one graphics processor 1710, and may additionally include an image processor 1715 and / or a video processor 1720, any of which may be a modular IP core. In at least one embodiment, the integrated circuit 1700 includes peripheral or bus logic, which includes a USB controller 1725, a UART controller 1730, an SPI / SDIO controller 1735, and an I... 2 2S / I 2 2C controller 1740. In at least one embodiment, integrated circuit 1700 may include display device 1745 coupled to one or more of High Definition Multimedia Interface (HDMI) controller 1750 and Mobile Industrial Processor Interface (MIPI) display interface 1755. In at least one embodiment, storage may be provided by flash memory subsystem 1760, including flash memory and flash memory controller. In at least one embodiment, a memory interface may be provided via memory controller 1765 for accessing SDRAM or SRAM memory devices. In at least one embodiment, some integrated circuits also include embedded security engine 1770.

[0372] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be used in integrated circuit 1700 to infer or predict operations based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0373] Figures 18A-18B Exemplary integrated circuits and associated graphics processors according to various embodiments described herein are illustrated, which may be manufactured using one or more IP cores. In addition to the illustrations, at least one embodiment may include other logic and circuitry, including additional graphics processors / cores, peripheral interface controllers, or general-purpose processor cores.

[0374] Figures 18A-18B This is a block diagram illustrating an exemplary graphics processor used within a SoC according to embodiments described herein. Figure 18A An exemplary graphics processor 1810 of a system-on-a-chip according to at least one embodiment is shown, which can be manufactured using one or more IP cores. Figure 18B Further exemplary graphics processor 1840 of a system-on-a-chip according to at least one embodiment is shown, which can be manufactured using one or more IP cores. In at least one embodiment, Figure 18A The graphics processor 1810 is a low-power graphics processor core. In at least one embodiment, Figure 18B The graphics processor 1840 is a higher-performance graphics processor core. In at least one embodiment, each graphics processor 1810, 1840 may be... Figure 17 A variant of the 1710 graphics processor.

[0375] In at least one embodiment, the graphics processor 1810 includes a vertex processor 1805 and one or more fragment processors 1815A-1815N (e.g., 1815A, 1815B, 1815C, 1815D to 1815N-1 and 1815N). In at least one embodiment, the graphics processor 1810 can execute different shader programs via separate logic, such that the vertex processor 1805 is optimized to perform operations for the vertex shader program, while one or more fragment processors 1815A-1815N perform fragment (e.g., pixel) shading operations for fragments or pixels or shader programs. In at least one embodiment, the vertex processor 1805 performs the vertex processing stage of the 3D graphics pipeline and generates primitive and vertex data. In at least one embodiment, one or more fragment processors 1815A-1815N use the primitive and vertex data generated by the vertex processor 1805 to generate a framebuffer for display on a display device. In at least one embodiment, one or more fragment processors 1815A-1815N are optimized to execute fragment shader programs as provided in the OpenGL API, which can be used to perform operations similar to those of pixel shader programs provided in the Direct 3D API.

[0376] In at least one embodiment, the graphics processor 1810 additionally includes one or more memory management units (MMUs) 1820A-1820B, one or more caches 1825A-1825B, and one or more circuit interconnects 1830A-1830B. In at least one embodiment, one or more MMUs 1820A-1820B provide a virtual-to-physical address mapping for the graphics processor 1810, including for the vertex processor 1805 and / or fragment processors 1815A-1815N, which can reference vertex or image / texture data stored in memory, in addition to vertex or image / texture data stored in one or more caches 1825A-1825B. In at least one embodiment, one or more MMUs 1820A-1820B can be synchronized with other MMUs within the system, including with... Figure 17 One or more application processors 1705, graphics processors 1715, and / or video processors 1720 are associated with one or more MMUs, enabling each processor 1705-1720 to participate in a shared or unified virtual memory system. In at least one embodiment, one or more circuit interconnects 1830A-1830B enable the graphics processor 1810 to connect to other IP cores within the SoC via the SoC's internal bus or via a direct connection.

[0377] In at least one embodiment, the graphics processor 1840 includes one or more shader cores 1855A-1855N (e.g., 1855A, 1855B, 1855C, 1855D, 1855E, 1855F to 1855N-1 and 1855N), such as Figure 18B As shown, it provides a unified shader core architecture, where a single core or type or core can execute all types of programmable shader code, including shader program code for implementing vertex shaders, fragment shaders, and / or compute shaders. In at least one embodiment, the number of shader cores can vary. In at least one embodiment, the graphics processor 1840 includes an inter-core task manager 1845, which acts as a thread dispatcher to assign execution threads to one or more shader cores 1855A-1855N and a tile unit 1858 to accelerate tile-based rendering operations, where scene rendering operations are subdivided in image space, for example, to take advantage of local spatial consistency within the scene or optimize the use of internal caches.

[0378] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8BDetails are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be integrated into an integrated circuit. Figure 18A and / or Figure 18B The above is used for inference or prediction operations based at least in part on weight parameters calculated using neural network training operations, neural network functions or architectures, or neural network use cases described herein.

[0379] Figures 19A-19B Additional exemplary graphics processor logic according to embodiments described herein is illustrated. In at least one embodiment, Figure 19A It shows that it can be included in Figure 17 The graphics core 1900 within the graphics processor 1710, and in at least one embodiment, may be as follows: Figure 18B The Unified Shader Core 1855A-1855N is shown. Figure 19B A highly parallel general-purpose graphics processing unit (“GPGPU”) 1930 suitable for deployment on a multi-chip module is shown in at least one embodiment.

[0380] In at least one embodiment, the graphics core 1900 includes a shared instruction cache 1902, texture units 1918, and cache / shared memory 1920, which are common to the execution resources within the graphics core 1900. In at least one embodiment, the graphics core 1900 may include multiple slices 1901A-1901N or partitions of each core, and the graphics processor may include multiple instances of the graphics core 1900. In at least one embodiment, slices 1901A-1901N may include supporting logic including local instruction caches 1904A-1904N, thread schedulers 1906A-1906N, thread dispatchers 1908A-1908N, and a set of registers 1910A-1910N. In at least one embodiment, slices 1901A-1901N may include a set of additional functional units (AFU 1912A-1912N), floating-point units (FPU 1914A-1914N), integer arithmetic logic units (ALU 1916A-1916N), address calculation units (ACU 1913A-1913N), double-precision floating-point units (DPFPU 1915A-1915N), and matrix processing units (MPU 1917A-1917N).

[0381] In at least one embodiment, the FPU 1914A-1914N can perform single-precision (32-bit) and half-precision (16-bit) floating-point operations, while the DPFPU 1915A-1915N performs double-precision (64-bit) floating-point operations. In at least one embodiment, the ALU 1916A-1916N can perform variable-precision integer operations with 8-bit, 16-bit, and 32-bit precision, and can be configured for mixed-precision operations. In at least one embodiment, the MPU 1917A-1917N can also be configured for mixed-precision matrix operations, including half-precision floating-point operations and 8-bit integer operations. In at least one embodiment, the MPU 1917-1917N can perform various matrix operations to accelerate machine learning application frameworks, including enabling support for accelerated generalized matrix-to-matrix multiplication (GEMM). In at least one embodiment, the AFU 1912A-1912N can perform additional logical operations not supported by floating-point or integer units, including trigonometric operations (e.g., sine, cosine, etc.).

[0382] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be used in the graphics core 1900 to infer or predict operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0383] Figure 19BA general-purpose processing unit (GPGPU) 1930 is illustrated in at least one embodiment, which can be configured to enable highly parallel computational operations to be performed by a set of graphics processing units. In at least one embodiment, the GPGPU 1930 can be directly linked to other instances of the GPGPU 1930 to create a multi-GPU cluster to improve the training speed for deep neural networks. In at least one embodiment, the GPGPU 1930 includes a host interface 1932 for connection to a host processor. In at least one embodiment, the host interface 1932 is a PCI Express interface. In at least one embodiment, the host interface 1932 can be a vendor-specific communication interface or communication structure. In at least one embodiment, the GPGPU 1930 receives commands from the host processor and uses a global scheduler 1934 to allocate execution threads associated with those commands to a set of compute clusters 1936A-1936H. In at least one embodiment, compute clusters 1936A-1936H share a cache memory 1938. In at least one embodiment, cache memory 1938 can be used as a higher-level cache than cache memory within computing clusters 1936A-1936H.

[0384] In at least one embodiment, the GPGPU 1930 includes memories 1944A-1944B, which are coupled to computing clusters 1936A-1936H via a set of memory controllers 1942A-1942B. In at least one embodiment, memories 1944A-1944B may include various types of memory devices, including dynamic random access memory (DRAM) or graphics random access memory, such as synchronous graphics random access memory (SGRAM), which includes graphics double data rate (GDDR) memory.

[0385] In at least one embodiment, each of the computing clusters 1936A-1936H includes a set of graphics cores, for example... Figure 19A The graphics core 1900 may include various types of integer and floating-point logic units that can perform computational operations across a range of precisions, including precisions suitable for machine learning computations. For example, in at least one embodiment, at least a subset of the floating-point units in each computing cluster 1936A-1936H may be configured to perform 16-bit or 32-bit floating-point operations, while different subsets of the floating-point units may be configured to perform 64-bit floating-point operations.

[0386] In at least one embodiment, multiple instances of the GPGPU 1930 can be configured as a computing cluster. In at least one embodiment, the communication used for synchronization and data exchange by the computing clusters 1936A-1936H varies between embodiments. In at least one embodiment, the multiple instances of the GPGPU 1930 communicate via a host interface 1932. In at least one embodiment, the GPGPU 1930 includes an I / O hub 1939 that couples the GPGPU 1930 to a GPU link 1940, enabling direct connection to other instances of the GPGPU 1930. In at least one embodiment, the GPU link 1940 is coupled to a dedicated GPU-to-GPU bridge, which enables communication and synchronization between the multiple instances of the GPGPU 1930. In at least one embodiment, the GPU link 1940 is coupled to a high-speed interconnect for sending and receiving data to and from other GPGPUs or parallel processors. In at least one embodiment, the multiple instances of the GPGPU 1930 reside in a separate data processing system and communicate via network devices accessible through the host interface 1932. In at least one embodiment, GPU link 1940 may be configured to connect to a host processor other than or as a replacement for host interface 1932.

[0387] In at least one embodiment, the GPGPU 1930 can be configured to train a neural network. In at least one embodiment, the GPGPU 1930 can be used within an inference platform. In at least one embodiment, where the GPGPU 1930 is used for inference, the GPGPU 1930 may include fewer compute clusters 1936A-1936H compared to when the GPGPU 1930 is used to train a neural network. In at least one embodiment, the memory technology associated with the memories 1944A-1944B can differ between inference and training configurations, with higher bandwidth memory technology dedicated to the training configuration. In at least one embodiment, the inference configuration of the GPGPU 1930 can support inference-specific instructions. For example, in at least one embodiment, the inference configuration can provide support for one or more 8-bit integer dot product instructions, which can be used during the inference operation of the deployed neural network.

[0388] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8BDetails are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be used in the GPGPU 1930 for inferring or predicting operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architecture, or neural network use cases described herein.

[0389] Figure 20 A block diagram of a computer system 2000 according to at least one embodiment is shown. In at least one embodiment, the computer system 2000 includes a processing subsystem 2001 having one or more processors 2002 and a system memory 2004 communicating via an interconnect path that may include a memory hub 2005. In at least one embodiment, the memory hub 2005 may be a separate component within a chipset component or may be integrated within one or more processors 2002. In at least one embodiment, the memory hub 2005 is coupled to an I / O subsystem 2011 via a communication link 2006. In one embodiment, the I / O subsystem 2011 includes an I / O hub 2007 that enables the computer system 2000 to receive input from one or more input devices 2008. In at least one embodiment, the I / O hub 2007 enables a display controller to provide output to one or more display devices 2010A, the display controller being included in one or more processors 2002. In at least one embodiment, one or more display devices 2010A coupled to the I / O hub 2007 may include local, internal or embedded display devices.

[0390] In at least one embodiment, the processing subsystem 2001 includes one or more parallel processors 2012 coupled to the memory hub 2005 via a bus or other communication link 2013. In at least one embodiment, the communication link 2013 may use any of many standards-based communication link technologies or protocols, such as, but not limited to, PCI Express, or may be a vendor-specific communication interface or communication architecture. In at least one embodiment, one or more parallel processors 2012 form a computationally concentrated parallel or vector processing system, which may include a large number of processing cores and / or processing clusters, such as a multi-core integrated (MIC) processor. In at least one embodiment, one or more parallel processors 2012 form a graphics processing subsystem that can output pixels to one or more display devices 2010A coupled via an I / O hub 2007. In at least one embodiment, the parallel processors 2012 may also include a display controller and a display interface (not shown) to enable direct connection to one or more display devices 2010B.

[0391] In at least one embodiment, system storage unit 2014 may be connected to I / O hub 2007 to provide a storage mechanism for computer system 2000. In at least one embodiment, I / O switch 2016 may be used to provide an interface mechanism to enable connectivity between I / O hub 2007 and other components, such as network adapter 2018 and / or wireless network adapter 2019 which may be integrated into the platform, and various other devices that can be added via one or more additional devices 2020. In at least one embodiment, network adapter 2018 may be an Ethernet adapter or another wired network adapter. In at least one embodiment, wireless network adapter 2019 may include one or more of Wi-Fi, Bluetooth, near field communication (NFC), or other network devices including one or more wireless devices.

[0392] In at least one embodiment, the computer system 2000 may include other components not explicitly shown, such as USB or other port connections, optical storage drives, video capture devices, etc., which may also be connected to the I / O hub 2007. In at least one embodiment, the interconnection can be implemented using any suitable protocol, such as PCI (Peripheral Component Interconnect) based protocols (e.g., PCI-Express) or other bus or point-to-point communication interfaces and / or protocols. Figure 20 The communication paths of the various components, such as NV-Link high-speed interconnect or interconnect protocols.

[0393] In at least one embodiment, one or more parallel processors 2012 include circuitry optimized for graphics and video processing, including, for example, video output circuitry, and constituting a graphics processing unit (GPU). In at least one embodiment, the parallel processor 2012 includes circuitry optimized for general-purpose processing. In at least one embodiment, components of the computer system 2000 may be integrated with one or more other system elements on a single integrated circuit. For example, in at least one embodiment, the parallel processor 2012, memory hub 2005, processor 2002, and I / O hub 2007 may be integrated into a system-on-a-chip (SoC) integrated circuit. In at least one embodiment, components of the computer system 2000 may be integrated into a single package to form a system-in-package (SIP) configuration. In at least one embodiment, at least a portion of the components of the computer system 2000 may be integrated into a multi-chip module (MCM) that can interconnect with other MCMs to a modular computer system.

[0394] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 can be... Figure 20 The system is used in 2000 for reasoning or predicting operations based at least in part on weight parameters calculated using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0395] processor

[0396] Figure 21A A parallel processor 2100 according to at least one embodiment is illustrated. In at least one embodiment, various components of the parallel processor 2100 may be implemented using one or more integrated circuit devices, such as programmable processors, application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). In at least one embodiment, the illustrated parallel processor 2100 is according to an exemplary embodiment. Figure 20 The variant of the 2012 with one or more parallel processors is shown.

[0397] In at least one embodiment, the parallel processor 2100 includes a parallel processing unit 2102. In at least one embodiment, the parallel processing unit 2102 includes an I / O unit 2104 that enables communication with other devices, including other instances of the parallel processing unit 2102. In at least one embodiment, the I / O unit 2104 can be directly connected to other devices. In at least one embodiment, the I / O unit 2104 is connected to other devices using a hub or switch interface (e.g., a memory hub 2105). In at least one embodiment, the connection between the memory hub 2105 and the I / O unit 2104 forms a communication link 2113. In at least one embodiment, the I / O unit 2104 is connected to a host interface 2106 and a memory crossbar switch 2116, wherein the host interface 2106 receives commands for performing processing operations, and the memory crossbar switch 2116 receives commands for performing memory operations.

[0398] In at least one embodiment, when host interface 2106 receives a command buffer via I / O unit 2104, host interface 2106 can direct work operations to execute those commands to front end 2108. In at least one embodiment, front end 2108 is coupled to scheduler 2110, which is configured to assign commands or other work items to processing cluster array 2112. In at least one embodiment, scheduler 2110 ensures that processing cluster array 2112 is correctly configured and in an active state before assigning tasks to processing cluster array 2112. In at least one embodiment, scheduler 2110 is implemented via firmware logic executed on a microcontroller. In at least one embodiment, the microcontroller-implemented scheduler 2110 can be configured to perform complex scheduling and work assignment operations at both coarse and fine granular levels, thereby enabling fast preemption and context switching of threads executing on processing array 2112. In at least one embodiment, host software can demonstrate workloads for scheduling on processing array 2112 via one of multiple graphics processing paths. In at least one embodiment, the workload can then be automatically distributed on the processing array 2112 by the scheduler 2110 logic within the microcontroller, which includes the scheduler 2110.

[0399] In at least one embodiment, the processing cluster array 2112 may include up to "N" processing clusters (e.g., clusters 2114A, 2114B to 2114N), where "N" represents a positive integer (which may be an integer different from the integer "N" used in other diagrams). In at least one embodiment, each cluster 2114A-2114N of the processing cluster array 2112 can execute a large number of concurrent threads. In at least one embodiment, the scheduler 2110 may use various scheduling and / or work allocation algorithms to allocate work to the clusters 2114A-2114N of the processing cluster array 2112, which may vary depending on the workload generated by each type of program or computation. In at least one embodiment, scheduling may be handled dynamically by the scheduler 2110, or may be partially assisted by compiler logic during the compilation of program logic configured to be executed by the processing cluster array 2112. In at least one embodiment, the different clusters 2114A-2114N of the processing cluster array 2112 may be assigned to process different types of programs or to perform different types of computations.

[0400] In at least one embodiment, the processing cluster array 2112 can be configured to perform various types of parallel processing operations. In at least one embodiment, the processing cluster array 2112 is configured to perform general-purpose parallel computing operations. For example, in at least one embodiment, the processing cluster array 2112 may include logic for performing processing tasks, including filtering video and / or audio data, performing modeling operations, including physical operations, and performing data transformations.

[0401] In at least one embodiment, the processing cluster array 2112 is configured to perform parallel graphics processing operations. In at least one embodiment, the processing cluster array 2112 may include additional logic to support the execution of such graphics processing operations, including but not limited to texture sampling logic for performing texture operations, as well as tessellation logic and other vertex processing logic. In at least one embodiment, the processing cluster array 2112 may be configured to execute shader programs related to graphics processing, such as, but not limited to, vertex shaders, tessellation shaders, geometry shaders, and pixel shaders. In at least one embodiment, the parallel processing unit 2102 may transfer data from system memory via I / O unit 2104 for processing. In at least one embodiment, during processing, the transferred data may be stored in on-chip memory (e.g., parallel processor memory 2122) and then written back to system memory.

[0402] In at least one embodiment, when the parallel processing unit 2102 is used to perform graphics processing, the scheduler 2110 may be configured to divide the processing workload into tasks of approximately equal size to better distribute graphics processing operations among the multiple clusters 2114A-2114N of the processing cluster array 2112. In at least one embodiment, portions of the processing cluster array 2112 may be configured to perform different types of processing. For example, in at least one embodiment, a first portion may be configured to perform vertex shading and topology generation, a second portion may be configured to perform tessellation and geometry shading, and a third portion may be configured to perform pixel shading or other screen-space operations to generate a rendered image for display. In at least one embodiment, intermediate data generated by one or more of the clusters 2114A-2114N may be stored in a buffer to allow intermediate data to be transferred between the clusters 2114A-2114N for further processing.

[0403] In at least one embodiment, the processing cluster array 2112 may receive processing tasks to be executed via a scheduler 2110, which receives commands defining the processing tasks from a front end 2108. In at least one embodiment, the processing task may include an index of data to be processed, such as surface (patch) data, raw data, vertex data, and / or pixel data, as well as state parameters and commands defining how the data is processed (e.g., what program to execute). In at least one embodiment, the scheduler 2110 may be configured to acquire an index corresponding to a task, or may receive an index from the front end 2108. In at least one embodiment, the front end 2108 may be configured to ensure that the processing cluster array 2112 is configured to be active before initiating the workload specified by an incoming command buffer (e.g., a batch buffer, push buffer, etc.).

[0404] In at least one embodiment, each of one or more instances of the parallel processing unit 2102 may be coupled to the parallel processor memory 2122. In at least one embodiment, the parallel processor memory 2122 may be accessed via a memory crossbar switch 2116, which may receive memory requests from the processing cluster array 2112 and the I / O unit 2104. In at least one embodiment, the memory crossbar switch 2116 may be accessed via a memory interface 2118. In at least one embodiment, the memory interface 2118 may include a plurality of partition units (e.g., partition units 2120A, 2120B to 2120N), each of which may be coupled to a portion (e.g., a memory cell) of the parallel processor memory 2122. In at least one embodiment, the plurality of partition units 2120A-2120N are configured to be equal to the number of memory units, such that the first partition unit 2120A has a corresponding first memory unit 2124A, the second partition unit 2120B has a corresponding memory unit 2124B, and the Nth partition unit 2120N has a corresponding Nth memory unit 2124N. In at least one embodiment, the number of partition units 2120A-2120N may not be equal to the number of memory units.

[0405] In at least one embodiment, memory cells 2124A-2124N may include various types of memory devices, including dynamic random access memory (DRAM) or graphics random access memory, such as synchronous graphics random access memory (SGRAM), including graphics double data rate (GDDR) memory. In at least one embodiment, memory cells 2124A-2124N may also include 3D stacked memory, including but not limited to high bandwidth memory (HBM). In at least one embodiment, rendering targets such as frame buffers or texture maps may be stored across memory cells 2124A-2124N, allowing partitioning cells 2120A-2120N to write portions of each rendering target in parallel, to efficiently utilize the available bandwidth of the parallel processor memory 2122. In at least one embodiment, local instances of the parallel processor memory 2122 may be excluded to facilitate a unified memory design that combines system memory with local cache memory.

[0406] In at least one embodiment, any of the clusters 2114A-2114N of the processing cluster array 2112 can process data to be written to any memory cell 2124A-2124N within the parallel processor memory 2122. In at least one embodiment, the memory crossbar switch 2116 can be configured to transfer the output of each cluster 2114A-2114N to any partition cell 2120A-2120N or another cluster 2114A-2114N, and the clusters 2114A-2114N can perform further processing operations on the output. In at least one embodiment, each cluster 2114A-2114N can communicate with the memory interface 2118 via the memory crossbar switch 2116 to read from or write to various external storage devices. In at least one embodiment, the memory crossbar switch 2116 has a connection to the memory interface 2118 for communication with the I / O unit 2104, and a connection to a local instance of the parallel processor memory 2122, thereby enabling processing units within different processing clusters 2114A-2114N to communicate with system memory or other memory not local to the parallel processing unit 2102. In at least one embodiment, the memory crossbar switch 2116 may use virtual channels to separate traffic flows between clusters 2114A-2114N and partition units 2120A-2120N.

[0407] In at least one embodiment, multiple instances of the parallel processing unit 2102 may be provided on a single insert card, or multiple insert cards may be interconnected. In at least one embodiment, different instances of the parallel processing unit 2102 may be configured to interoperate, even if the different instances have different numbers of processing cores, different numbers of local parallel processor memories, and / or other configuration differences. For example, in at least one embodiment, some instances of the parallel processing unit 2102 may include higher-precision floating-point units relative to other instances. In at least one embodiment, a system combining one or more instances of the parallel processing unit 2102 or the parallel processor 2100 may be implemented in various configurations and form factors, including but not limited to desktop, laptop, or handheld personal computers, servers, workstations, game consoles, and / or embedded systems.

[0408] Figure 21B This is a block diagram of a partitioning unit 2120 according to at least one embodiment. In at least one embodiment, the partitioning unit 2120 is... Figure 21A This is an example of one of the partitioning units 2120A-2120N. In at least one embodiment, the partitioning unit 2120 includes an L2 cache 2121, a frame buffer interface 2125, and a ROP 2126 (raster operation unit). In at least one embodiment, the L2 cache 2121 is a read / write cache configured to perform load and store operations received from the memory crossbar switch 2116 and the ROP 2126. In at least one embodiment, the L2 cache 2121 outputs read misses and urgent write-back requests to the frame buffer interface 2125 for processing. In at least one embodiment, updates can also be sent to the frame buffer for processing via the frame buffer interface 2125. In at least one embodiment, the frame buffer interface 2125 communicates with memory cells in the parallel processor memory (such as...). Figure 21A The memory cells 2124A-2124N (e.g., within the parallel processor memory 2122) interact with one of them.

[0409] In at least one embodiment, ROP 2126 is a processing unit that performs raster operations such as stenciling, z-testing, blending, etc. In at least one embodiment, ROP 2126 then outputs processed graphics data stored in graphics memory. In at least one embodiment, ROP 2126 includes compression logic to compress depth or color data written to memory and decompress depth or color data read from memory. In at least one embodiment, the compression logic may be lossless compression logic utilizing one or more of a variety of compression algorithms. In at least one embodiment, the type of compression performed by ROP 2126 may vary based on the statistical characteristics of the data to be compressed. For example, in at least one embodiment, incremental color compression is performed based on depth and color data on a per-tile basis.

[0410] In at least one embodiment, ROP 2126 is included within each processing cluster (e.g., Figure 21A Clusters 2114A-2114N are used instead of partition units 2120. In at least one embodiment, read and write requests for pixel data are made via memory crossbar switch 2116 instead of pixel fragment data transfer. In at least one embodiment, the processed graphics data can be displayed on a display device (such as...). Figure 22 Displayed by one or more display devices 2210, routed by processor 2202 for further processing, or by... Figure 21A One of the processing entities within the parallel processor 2100 is routed for further processing.

[0411] Figure 21C This is a block diagram of a processing cluster 2114 within a parallel processing unit according to at least one embodiment. In at least one embodiment, the processing cluster is... Figure 21A An instance of one of the processing clusters 2114A-2114N. In at least one embodiment, the processing cluster 2114 can be configured to execute a number of threads in parallel, where a "thread" refers to an instance of a specific program executing on a particular set of input data. In at least one embodiment, a Single Instruction Multiple Data (SIMD) instruction issuing technique is used to support the parallel execution of a large number of threads without providing multiple independent instruction units. In at least one embodiment, a Single Instruction Multiple Threading (SIMT) technique is used to support the parallel execution of a large number of generally synchronous threads, which uses a common instruction unit configured to issue instructions to a set of processing engines within each processing cluster.

[0412] In at least one embodiment, the operation of the processing cluster 2114 can be controlled by a pipeline manager 2132 that assigns processing tasks to the SIMT parallel processors. In at least one embodiment, the pipeline manager 2132... Figure 21AThe scheduler 2110 receives instructions and manages the execution of these instructions via the graphics multiprocessor 2134 and / or texture unit 2136. In at least one embodiment, the graphics multiprocessor 2134 is an exemplary instance of a SIMT parallel processor. However, in at least one embodiment, the processing cluster 2114 may include various types of SIMT parallel processors with different architectures. In at least one embodiment, the processing cluster 2114 may include one or more instances of the graphics multiprocessor 2134. In at least one embodiment, the graphics multiprocessor 2134 can process data, and the data crossover switch 2140 can be used to distribute the processed data to one of a number of possible destinations, including other shader units. In at least one embodiment, the pipeline manager 2132 can facilitate the distribution of processed data by specifying the destination of the processed data to be distributed via the data crossover switch 2140.

[0413] In at least one embodiment, each graphics multiprocessor 2134 within the processing cluster 2114 may include the same set of functional execution logic (e.g., arithmetic logic units, load-memory units, etc.). In at least one embodiment, the functional execution logic may be configured in a pipelined manner, wherein new instructions may be issued before previous instructions complete. In at least one embodiment, the functional execution logic supports a variety of operations, including integer and floating-point arithmetic, comparison operations, Boolean operations, shift operations, and computation of various algebraic functions. In at least one embodiment, the same functional unit hardware may be used to perform different operations, and any combination of functional units may exist.

[0414] In at least one embodiment, instructions sent to the processing cluster 2114 constitute threads. In at least one embodiment, a group of threads executed across a set of parallel processing engines is a thread group. In at least one embodiment, the thread group executes a general program on different input data. In at least one embodiment, each thread within the thread group may be assigned to a different processing engine within the graphics multiprocessor 2134. In at least one embodiment, the thread group may include fewer threads than the number of processing engines within the graphics multiprocessor 2134. In at least one embodiment, when the number of threads included in the thread group is less than the number of processing engines, one or more processing engines may be idle during a loop that is processing the thread group. In at least one embodiment, the thread group may also include more threads than the number of processing engines within the graphics multiprocessor 2134. In at least one embodiment, when the thread group includes more threads than the number of processing engines within the graphics multiprocessor 2134, processing can be performed in consecutive clock cycles. In at least one embodiment, multiple thread groups can be executed simultaneously on the graphics multiprocessor 2134.

[0415] In at least one embodiment, the graphics multiprocessor 2134 includes an internal cache memory for performing load and store operations. In at least one embodiment, the graphics multiprocessor 2134 may forgo the internal cache and use a cache memory within the processing cluster 2114 (e.g., L1 cache 2148). In at least one embodiment, each graphics multiprocessor 2134 may also access partition units (e.g., Figure 21A The L2 cache is located within partition units 2120A-2120N, which are shared across all processing clusters 2114 and can be used to transfer data between threads. In at least one embodiment, the graphics multiprocessor 2134 can also access off-chip global memory, which may include one or more of local parallel processor memory and / or system memory. In at least one embodiment, any memory outside of the parallel processing unit 2102 can be used as global memory. In at least one embodiment, the processing cluster 2114 includes multiple instances of the graphics multiprocessor 2134, which can share common instructions and data that can be stored in the L1 cache 2148.

[0416] In at least one embodiment, each processing cluster 2114 may include a memory management unit (“MMU”) 2145 configured to map virtual addresses to physical addresses. In at least one embodiment, one or more instances of the MMU 2145 may reside in Figure 21A The memory interface 2118 is located within the MMU 2145. In at least one embodiment, the MMU 2145 includes a set of page table entries (PTEs) for mapping virtual addresses to physical addresses of tiles and optionally to cache line indices. In at least one embodiment, the MMU 2145 may include an address translation lookup buffer (TLB) or a cache that may reside within the graphics multiprocessor 2134, the L1 cache 2148, or the processing cluster 2114. In at least one embodiment, physical addresses are processed to allocate surface data access locality for efficient request interleaving between partition units. In at least one embodiment, cache line indices may be used to determine whether a request for a cache line is a hit or a miss.

[0417] In at least one embodiment, the processing cluster 2114 can be configured such that each graphics multiprocessor 2134 is coupled to a texture unit 2136 to perform texture mapping operations that determine texture sample locations, read texture data, and filter texture data. In at least one embodiment, texture data is read as needed from an internal texture L1 cache (not shown) or from an L1 cache within the graphics multiprocessor 2134, and texture data is also retrieved from an L2 cache, local parallel processor memory, or system memory. In at least one embodiment, each graphics multiprocessor 2134 outputs a processed task to a data crossbar switch 2140 to provide the processed task to another processing cluster 2114 for further processing or to store the processed task in an L2 cache, local parallel processor memory, or system memory via a memory crossbar switch 2116. In at least one embodiment, a preROP 2142 (pre-raster operation unit) is configured to receive data from the graphics multiprocessor 2134 and direct the data to a ROP unit, which may be associated with a partitioning unit (e.g., [missing information]). Figure 21A The PreROP 2142 unit is located together with the partitioning units 2120A-2120N. In at least one embodiment, the PreROP 2142 unit can perform optimizations for color blending, organize pixel color data, and perform address translation.

[0418] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be used in the graphics processing cluster 2114 to perform inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0419] Figure 21D A graphics multiprocessor 2134 according to at least one embodiment is illustrated. In at least one embodiment, the graphics multiprocessor 2134 is coupled to a pipeline manager 2132 of a processing cluster 2114. In at least one embodiment, the graphics multiprocessor 2134 has an execution pipeline including, but not limited to, an instruction cache 2152, an instruction unit 2154, an address mapping unit 2156, a register file 2158, one or more general-purpose graphics processing unit (GPGPU) cores 2162, and one or more load / store units 2166. In at least one embodiment, the GPGPU cores 2162 and the load / store units 2166 are coupled to a cache memory 2172 and a shared memory 2170 via a memory and cache interconnect 2168.

[0420] In at least one embodiment, instruction cache 2152 receives a stream of instructions to be executed from pipeline manager 2132. In at least one embodiment, instructions are cached in instruction cache 2152 and dispatched to instruction unit 2154 for execution. In one embodiment, instruction unit 2154 may dispatch instructions as thread groups (e.g., thread bundles), assigning each thread of the thread group to a different execution unit within GPGPU core 2162. In at least one embodiment, instructions can access any local, shared, or global address space by specifying an address within a unified address space. In at least one embodiment, address mapping unit 2156 may be used to translate addresses in the unified address space into different memory addresses that can be accessed by load / store unit 2166.

[0421] In at least one embodiment, register file 2158 provides a set of registers for functional units of graphics multiprocessor 2134. In at least one embodiment, register file 2158 provides temporary storage for operands of data paths connected to functional units of graphics multiprocessor 2134 (e.g., GPGPU core 2162, load / store unit 2166). In at least one embodiment, register file 2158 is partitioned between each functional unit, such that a dedicated portion of register file 2158 is allocated to each functional unit. In at least one embodiment, register file 2158 is partitioned between different thread bundles being executed by graphics multiprocessor 2134.

[0422] In at least one embodiment, each of the GPGPU cores 2162 may include a floating-point unit (FPU) and / or an integer arithmetic logic unit (ALU) for executing instructions of the graphics multiprocessor 2134. In at least one embodiment, the GPGPU cores 2162 may be architecturally similar or may differ in architecture. In at least one embodiment, a first portion of the GPGPU core 2162 includes a single-precision FPU and an integer ALU, while a second portion of the GPGPU core includes a double-precision FPU. In at least one embodiment, the FPU may implement the IEEE 754-2008 standard for floating-point algorithms or enable variable-precision floating-point algorithms. In at least one embodiment, the graphics multiprocessor 2134 may additionally include one or more fixed-function or special-function units to perform specific functions, such as copying rectangles or pixel blending operations. In at least one embodiment, one or more of the GPGPU cores 2162 may also include fixed-function or special-function logic.

[0423] In at least one embodiment, the GPGPU core 2162 includes SIMD logic capable of executing a single instruction on multiple sets of data. In one embodiment, the GPGPU core 2162 can physically execute SIMD4, SIMD8, and SIMD16 instructions and logically execute SIMD1, SIMD2, and SIMD32 instructions. In at least one embodiment, the SIMD instructions for the GPGPU core can be generated by a shader compiler at compile time or automatically generated when executing a program written and compiled for a Single Program Multiple Data (SPMD) or SIMT architecture. In at least one embodiment, multiple threads of a program configured for a SIMT execution model can be executed using a single SIMD instruction. For example, in at least one embodiment, eight SIMD threads performing the same or similar operations can be executed in parallel using a single SIMD8 logic unit.

[0424] In at least one embodiment, the memory and cache interconnect 2168 is an interconnect network connecting each functional unit of the graphics multiprocessor 2134 to the register file 2158 and the shared memory 2170. In at least one embodiment, the memory and cache interconnect 2168 is a cross-switch interconnect that allows the load / store unit 2166 to perform load and store operations between the shared memory 2170 and the register file 2158. In at least one embodiment, the register file 2158 can operate at the same frequency as the GPGPU core 2162, resulting in very low latency for data transfer between the GPGPU core 2162 and the register file 2158. In at least one embodiment, the shared memory 2170 can be used to enable communication between threads executing on functional units within the graphics multiprocessor 2134. In at least one embodiment, the cache memory 2172 can be used, for example, as a data cache to cache texture data communicated between functional units and texture units 2136. In at least one embodiment, the shared memory 2170 can also be used as a program-managed cache. In at least one embodiment, in addition to the data automatically cached in cache memory 2172, the thread executing on GPGPU core 2162 can also programmatically store data in shared memory.

[0425] In at least one embodiment, a parallel processor or GPGPU, as described herein, is communicatively coupled to a host / processor core to accelerate graphics operations, machine learning operations, pattern analysis operations, and various general-purpose GPU (GPGPU) functions. In at least one embodiment, the GPU may be communicatively coupled to the host processor / core via a bus or other interconnect (e.g., high-speed interconnects such as PCIe or NVLink). In at least one embodiment, the GPU may be integrated with the core on a package or chip and communicatively coupled to the core via an internal processor bus / interconnect (i.e., within the package or chip). In at least one embodiment, regardless of how the GPU is connected, the processor core may assign work to the GPU in the form of a sequence of commands / instructions contained in a job descriptor. In at least one embodiment, the GPU then uses dedicated circuitry / logic to efficiently process these commands / instructions.

[0426] The inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. The following is in conjunction with... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be used in the graphics multiprocessor 2134 to perform inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0427] Figure 22A multi-GPU computing system 2200 according to at least one embodiment is illustrated. In at least one embodiment, the multi-GPU computing system 2200 may include a processor 2202 coupled to a plurality of general-purpose graphics processing units (GPGPUs) 2206A-D via a host interface switch 2204. In at least one embodiment, the host interface switch 2204 is a PCI Express switch device that couples the processor 2202 to a PCI Express bus, through which the processor 2202 can communicate with the GPGPUs 2206A-D. In at least one embodiment, the GPGPUs 2206A-D may be interconnected via a set of high-speed P2P GPU-to-GPU links 2216. In at least one embodiment, the GPU-to-GPU links 2216 are connected to each of the GPGPUs 2206A-D via dedicated GPU links. In at least one embodiment, the P2P GPU links 2216 enable direct communication between each GPGPU 2206A-D without communication via the host interface bus 2204 to which the processor 2202 is connected. In at least one embodiment, when GPU-to-GPU traffic is directed to the P2P GPU link 2216, the host interface bus 2204 remains available for system memory access or, for example, communication with other instances of the multi-GPU computing system 2200 via one or more network devices. While in at least one embodiment, the GPGPUs 2206A-D are connected to the processor 2202 via the host interface switch 2204, in at least one embodiment, the processor 2202 includes direct support for the P2P GPU link 2216 and can be directly connected to the GPGPUs 2206A-D.

[0428] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding inference and / or training logic 815. In at least one embodiment, inference and / or training logic 815 may be used in a multi-GPU computing system 2200 for performing inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0429] Figure 23This is a block diagram of a graphics processor 2300 according to at least one embodiment. In at least one embodiment, the graphics processor 2300 includes a ring interconnect 2302, a pipeline front end 2304, a media engine 2337, and graphics cores 2380A-2380N. In at least one embodiment, the ring interconnect 2302 couples the graphics processor 2300 to other processing units, said processing units including other graphics processors or one or more general-purpose processor cores. In at least one embodiment, the graphics processor 2300 is one of many processors integrated within a multi-core processing system.

[0430] In at least one embodiment, the graphics processor 2300 receives multiple batches of commands via a ring interconnect 2302. In at least one embodiment, the input commands are interpreted by a command streamer 2303 in a pipeline front-end 2304. In at least one embodiment, the graphics processor 2300 includes scalable execution logic for performing 3D geometry processing and media processing via graphics cores 2380A-2380N. In at least one embodiment, for 3D geometry processing commands, the command streamer 2303 provides the commands to the geometry pipeline 2336. In at least one embodiment, for at least some media processing commands, the command streamer 2303 provides the commands to a video front-end 2334, which is coupled to a media engine 2337. In at least one embodiment, the media engine 2337 includes a video quality engine (VQE) 2330 for video and image post-processing, and a multi-format encoding / decoding (MFX) engine 2333 for providing hardware-accelerated media data encoding and decoding. In at least one embodiment, the geometry pipeline 2336 and the media engine 2337 each generate an execution thread for thread execution resources provided by at least one graphics core 2380.

[0431] In at least one embodiment, the graphics processor 2300 includes scalable thread execution resources featuring graphics cores 2380A-2380N (which may be modular and sometimes referred to as core slices), each graphics core having multiple sub-cores 2350A-2350N, 2360A-2360N (sometimes referred to as core sub-slices). In at least one embodiment, the graphics processor 2300 may have any number of graphics cores 2380A. In at least one embodiment, the graphics processor 2300 includes graphics cores 2380A having at least a first sub-core 2350A and a second sub-core 2360A. In at least one embodiment, the graphics processor 2300 is a low-power processor having a single sub-core (e.g., 2350A). In at least one embodiment, the graphics processor 2300 includes multiple graphics cores 2380A-2380N, each graphics core including a set of first sub-cores 2350A-2350N and a set of second sub-cores 2360A-2360N. In at least one embodiment, each of the first sub-cores 2350A-2350N includes at least a first set of execution units 2352A-2352N and media / texture samplers 2354A-2354N. In at least one embodiment, each of the second sub-cores 2360A-2360N includes at least a second set of execution units 2362A-2362N and samplers 2364A-2364N. In at least one embodiment, each of the sub-cores 2350A-2350N and 2360A-2360N shares a set of shared resources 2370A-2370N. In at least one embodiment, the shared resources include a shared cache memory and pixel operation logic.

[0432] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details are provided regarding the inference and / or training logic 815. In at least one embodiment, the inference and / or training logic 815 may be used in the graphics processor 2300 to perform inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures or neural network use cases described herein.

[0433] Figure 24This is a block diagram illustrating a microarchitecture for a processor 2400 according to at least one embodiment, the processor 2400 including logic circuitry for executing instructions. In at least one embodiment, the processor 2400 can execute instructions, including x86 instructions, ARM instructions, and special-purpose instructions for application-specific integrated circuits (ASICs). In at least one embodiment, the processor 2400 may include registers for storing packaged data, such as the 64-bit wide MMX registers used in Intel Corporation's Santa Clara, California-enabled microprocessors employing MMX technology. TM Registers. In at least one embodiment, MMX registers available in integer and floating-point forms can operate alongside packaged data elements accompanied by Single Instruction Multiple Data (“SIMD”) and Streaming SIMD Extensions (“SSE”) instructions. In at least one embodiment, a 128-bit wide XMM register associated with SSE2, SSE3, SSE4, AVX, or later (generally referred to as “SSEx”) technologies can hold such packaged data operands. In at least one embodiment, processor 2400 can execute instructions to accelerate machine learning or deep learning algorithms, training, or inference.

[0434] In at least one embodiment, processor 2400 includes an ordered front end (“front end”) 2401 to fetch instructions to be executed and prepare instructions for later use in the processor pipeline. In at least one embodiment, front end 2401 may include several units. In at least one embodiment, instruction prefetcher 2426 fetches instructions from memory and provides the instructions to instruction decoder 2428, which in turn decodes or interprets the instructions. For example, in at least one embodiment, instruction decoder 2428 decodes the received instructions into one or more machine-executable so-called “micro-instructions” or “micro-operations” (also referred to as “micro-operations” or “micro-instructions”). In at least one embodiment, instruction decoder 2428 parses the instructions into opcodes and corresponding data and control fields, which can be used by the microarchitecture to perform operations according to at least one embodiment. In at least one embodiment, trace cache 2430 may assemble the decoded micro-instructions into a program-ordered sequence or trace in micro-instruction queue 2434 for execution. In at least one embodiment, when trace cache 2430 encounters complex instructions, microcode ROM 2432 provides the micro-instructions required to complete the operation.

[0435] In at least one embodiment, some instructions may be converted into a single micro-operation, while others require several micro-operations to complete the entire operation. In at least one embodiment, if more than four micro-instructions are required to complete an instruction, the instruction decoder 2428 may access the microcode ROM 2432 to execute the instruction. In at least one embodiment, an instruction may be decoded into a small number of micro-instructions for processing at the instruction decoder 2428. In at least one embodiment, if multiple micro-instructions are required to complete the operation, the instructions may be stored in the microcode ROM 2432. In at least one embodiment, the trace cache 2430 references an entry point programmable logic array (“PLA”) to determine the correct micro-instruction pointer for reading a microcode sequence from the microcode ROM 2432 to complete one or more instructions, according to at least one embodiment. In at least one embodiment, after the microcode ROM 2432 has completed the micro-operation ordering of the instructions, the machine front end 2401 may resume fetching micro-operations from the trace cache 2430.

[0436] In at least one embodiment, the out-of-order execution engine (“out-of-order engine”) 2403 can prepare instructions for execution. In at least one embodiment, the out-of-order execution logic has multiple buffers to smooth and reorder the instruction flow to optimize performance as instructions descend the pipeline and are scheduled for execution. In at least one embodiment, the out-of-order execution engine 2403 includes, but is not limited to, an allocator / register renamer 2440, a memory microinstruction queue 2442, an integer / floating-point microinstruction queue 2444, a memory scheduler 2446, a fast scheduler 2402, a slow / general-purpose floating-point scheduler (“slow / general-purpose FP scheduler”) 2404, and a simple floating-point scheduler (“simple FP scheduler”) 2406. In at least one embodiment, the fast scheduler 2402, the slow / general-purpose floating-point scheduler 2404, and the simple floating-point scheduler 2406 are also collectively referred to as “microinstruction schedulers 2402, 2404, 2406”. In at least one embodiment, the allocator / register renamer 2440 allocates the machine buffers and resources required for the sequential execution of each microinstruction. In at least one embodiment, the allocator / register renamer 2440 renames logical registers to entries in a register file. In at least one embodiment, the allocator / register renamer 2440 also allocates entries for each microinstruction in one of two microinstruction queues, a memory microinstruction queue 2442 for memory operations and an integer / floating-point microinstruction queue 2444 for non-memory operations, preceding the memory scheduler 2446 and microinstruction schedulers 2402, 2404, and 2406. In at least one embodiment, the microinstruction schedulers 2402, 2404, and 2406 determine when they are ready to execute a microinstruction based on the readiness of their dependent input register operand sources and the availability of the execution resource microinstructions that need to be completed. In at least one embodiment, the fast scheduler 2402 can schedule on each half of the master clock cycle, while the slow / general-purpose floating-point scheduler 2404 and the simple floating-point scheduler 2406 can schedule once per master processor clock cycle. In at least one embodiment, microinstruction schedulers 2402, 2404, and 2406 arbitrate the scheduling port to schedule microinstructions for execution.

[0437] In at least one embodiment, execution block 2411 includes, but is not limited to, integer register file / branch network 2408, floating-point register file / branch network (“FP register file / branch network”) 2410, address generation units (“AGU”) 2412 and 2414, fast arithmetic logic units (“fast ALU”) 2416 and 2418, slow arithmetic logic unit (“slow ALU”) 2420, floating-point ALU (“FP”) 2422, and floating-point movement unit (“FP movement”) 2424. In at least one embodiment, integer register file / branch network 2408 and floating-point register file / bypass network 2410 are also referred to herein as “register files 2408, 2410”. In at least one embodiment, AGUs 2412 and 2414, fast ALUs 2416 and 2418, slow ALU 2420, floating-point ALU 2422, and floating-point movement unit 2424 are also referred to herein as "execution units 2412, 2414, 2416, 2418, 2420, 2422, and 2424". In at least one embodiment, execution block 2411 may include, but is not limited to, any number (including zero) and type of register files, branch networks, address generation units, and execution units (in any combination).

[0438] In at least one embodiment, register networks 2408, 2410 may be arranged between microinstruction schedulers 2402, 2404, 2406 and execution units 2412, 2414, 2416, 2418, 2420, 2422, and 2424. In at least one embodiment, integer register file / tribute network 2408 performs integer operations. In at least one embodiment, floating-point register file / tribute network 2410 performs floating-point operations. In at least one embodiment, each of register networks 2408, 2410 may include, but is not limited to, a tribute network that can bypass or forward recently completed results not yet written to a register file to a new dependent object. In at least one embodiment, register networks 2408, 2410 may communicate data with each other. In at least one embodiment, integer register file / tribute network 2408 may include, but is not limited to, two separate register files, one register file for low-order 32-bit data and a second register file for high-order 32-bit data. In at least one embodiment, the floating-point register file / branch network 2410 may include, but is not limited to, entries with a width of 128 bits, since floating-point instructions typically have operands with a width of 64 to 128 bits.

[0439] In at least one embodiment, execution units 2412, 2414, 2416, 2418, 2420, 2422, and 2424 can execute instructions. In at least one embodiment, register networks 2408 and 2410 store integer and floating-point data operation values ​​that the microinstructions need to execute. In at least one embodiment, processor 2400 may include, but is not limited to, any number of execution units 2412, 2414, 2416, 2418, 2420, 2422, and 2424, and combinations thereof. In at least one embodiment, floating-point ALU 2422 and floating-point move unit 2424 can perform floating-point, MMX, SIMD, AVX, and SSE or other operations, including specialized machine learning instructions. In at least one embodiment, floating-point ALU 2422 may include, but is not limited to, a 64-bit multiplication-64-bit floating-point divider to perform division, square root, and remainder micro-operations. In at least one embodiment, floating-point hardware can be used to process instructions involving floating-point values. In at least one embodiment, ALU operations can be passed to fast ALUs 2416 and 2418. In at least one embodiment, fast ALUs 2416 and 2418 can perform fast operations with an effective delay of half a clock cycle. In at least one embodiment, most complex integer operations are routed to slow ALU 2420, because slow ALU 2420 can include, but is not limited to, integer execution hardware for long-latency type operations, such as multipliers, shifters, flag logic, and branching. In at least one embodiment, memory load / store operations can be performed by AGUs 2412 and 2414. In at least one embodiment, fast ALU 2416, fast ALU 2418, and slow ALU 2420 can perform integer operations on 64-bit data operands. In at least one embodiment, fast ALU 2416, fast ALU 2418, and slow ALU 2420 can be implemented to support various data bit sizes, including sixteen, thirty-two, 128, 256, etc. In at least one embodiment, the floating-point ALU 2422 and the floating-point moving unit 2424 can be implemented to support a range of operands with various bit widths, for example, they can be combined with SIMD and multimedia instructions to operate on 128-bit wide packaged data operands.

[0440] In at least one embodiment, microinstruction schedulers 2402, 2404, and 2406 schedule dependent operations before the parent load completes execution. In at least one embodiment, since microinstructions can be speculatively scheduled and executed within processor 2400, processor 2400 may also include logic for handling memory misses. In at least one embodiment, if a data load miss occurs in the data cache, there may be a dependent operation running in the pipeline that temporarily deprives the scheduler of the correct data. In at least one embodiment, a replay mechanism tracks and re-executes instructions that use incorrect data. In at least one embodiment, it may be necessary to replay dependent operations and may allow independent operations to be completed. In at least one embodiment, the scheduler and replay mechanism of at least one embodiment of the processor may also be designed to capture instruction sequences used for text string comparison operations.

[0441] In at least one embodiment, "register" can refer to an onboard processor storage location that can be used as part of an instruction that identifies operands. In at least one embodiment, a register can be one that can be used externally to the processor (from a programmer's perspective). In at least one embodiment, a register may not be limited to a particular type of circuit. Rather, in at least one embodiment, a register can store data, provide data, and perform the functions described herein. In at least one embodiment, the registers described herein can be implemented using a variety of different techniques via circuitry within the processor, such as dedicated physical registers, dynamically allocated physical registers renamed using register renaming, a combination of dedicated and dynamically allocated physical registers, etc. In at least one embodiment, an integer register stores 32-bit integer data. The register file of at least one embodiment also includes eight multimedia SIMD registers for encapsulating data.

[0442] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details regarding inference and / or training logic 815 are provided. In at least one embodiment, some or all of the inference and / or training logic 815 may be incorporated into execution block 2411 and other memories or registers shown or not shown. For example, in at least one embodiment, the training and / or inference techniques described herein may use one or more ALUs shown in execution block 2411. Furthermore, weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALUs of execution block 2411 to execute one or more machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0443] Figure 25A deep learning application processor 2500 according to at least one embodiment is illustrated. In at least one embodiment, the deep learning application processor 2500 uses instructions, which, if executed by the deep learning application processor 2500, cause the deep learning application processor 2500 to perform some or all of the processes and techniques described herein. In at least one embodiment, the deep learning application processor 2500 is an application-specific integrated circuit (ASIC). In at least one embodiment, the application processor 2500 performs matrix multiplication operations or is "hardwired" into hardware as a result of executing one or more instructions or both. In at least one embodiment, the deep learning application processor 2500 includes, but is not limited to, processing clusters 2510(1)-2510(12), inter-chip links (“ICL”) 2520(1)-2520(12), inter-chip controllers (“ICC”) 2530(1)-2530(2), second-generation high-bandwidth memory (“HBM2”) 2540(1)-2540(4), memory controllers (“MemCtrlr”) 2542(1)-2542(4), high-bandwidth memory physical layers (“HBM PHY”) 2544(1)-2544(4), management controller central processing unit (“management controller CPU”) 2550, serial peripheral interface, internal integrated circuits and general purpose input / output blocks (“SPI, I2C, GPIO”) 2560, peripheral component interconnect fast controller and direct memory access block (“PCIe controller and DMA”) 2570, and sixteen-channel peripheral component interconnect fast port (“PCI Express x”). 16”2580.

[0444] In at least one embodiment, processing cluster 2510 can perform deep learning operations, including inference or prediction operations based on weight parameters computed using one or more training techniques, including those described herein. In at least one embodiment, each processing cluster 2510 can include, but is not limited to, any number and type of processors. In at least one embodiment, deep learning application processor 2500 can include any number and type of processing cluster 2510. In at least one embodiment, the inter-chip link 2520 is bidirectional. In at least one embodiment, the inter-chip link 2520 and the inter-chip controller 2530 enable multiple deep learning application processors 2500 to exchange information, including activation information generated from executing one or more machine learning algorithms embodied in one or more neural networks. In at least one embodiment, deep learning application processor 2500 can include any number (including zero) and type of ICL 2520 and ICC 2530.

[0445] In at least one embodiment, the HBM2 2540 provides a total of 32 GB of memory. In at least one embodiment, the HBM2 2540(i) is associated with both the memory controller 2542(i) and the HBM PHY 2544(i), where “i” is any integer. In at least one embodiment, any number of HBM2 2540s can provide any type and total amount of high-bandwidth memory and can be associated with any number (including zero) and type of memory controller 2542 and HBM PHY 2544. In at least one embodiment, any number and type of blocks can replace SPI, I2C, GPIO 3360, PCIe controller 2560 and DMA2570 and / or PCIe2580 to implement any number and type of communication standards in any technically feasible manner.

[0446] Inference and / or training logic 815 is used to perform inference and / or training operations associated with one or more embodiments. This document combines... Figure 8A and / or Figure 8B Details regarding the inference and / or training logic 815 are provided. In at least one embodiment, the deep learning application processor is used to train a machine learning model (e.g., a neural network) to predict or infer information provided to the deep learning application processor 2500. In at least one embodiment, the deep learning application processor 2500 is used to infer or predict information based on a trained machine learning model (e.g., a neural network) that has been trained by another processor or system or by the deep learning application processor 2500. In at least one embodiment, the processor 2500 may be used to perform one or more neural network use cases described herein.

[0447] Figure 26This is a block diagram of a neuromorphic processor 2600 according to at least one embodiment. In at least one embodiment, the neuromorphic processor 2600 may receive one or more inputs from a source external to the neuromorphic processor 2600. In at least one embodiment, these inputs may be transmitted to one or more neurons 2602 within the neuromorphic processor 2600. In at least one embodiment, the neurons 2602 and their components may be implemented using circuitry or logic including one or more arithmetic logic units (ALUs). In at least one embodiment, the neuromorphic processor 2600 may include, but is not limited to, thousands upon thousands of instances of neurons 2602, but any suitable number of neurons 2602 may be used. In at least one embodiment, each instance of a neuron 2602 may include a neuron input 2604 and a neuron output 2606. In at least one embodiment, a neuron 2602 may generate an output that can be transmitted to the inputs of other instances of the neuron 2602. In at least one embodiment, the neuron input 2604 and the neuron output 2606 may be interconnected via synapses 2608.

[0448] In at least one embodiment, neuron 2602 and synapse 2608 may be interconnected, causing neuromorphic processor 2600 to operate to process or analyze information received by neuromorphic processor 2600. In at least one embodiment, neuron 2602 may send an output pulse (or “trigger” or “peak”) when the input received through neuron input 2604 exceeds a threshold. In at least one embodiment, neuron 2602 may sum or integrate the signal received at neuron input 2604. For example, in at least one embodiment, neuron 2602 may be implemented as a leaky integral-triggered neuron, wherein if the summation (referred to as a “membrane potential”) exceeds a threshold, neuron 2602 may use a transfer function such as a sigmoid or threshold function to generate an output (or “trigger”). In at least one embodiment, the leaky integral-triggered neuron may sum the signal received at neuron input 2604 to a membrane potential and may apply an attenuation factor (or leak) to reduce the membrane potential. In at least one embodiment, a leaking integral-triggered neuron may trigger if multiple input signals are received at neuron input 2604 quickly enough to exceed a threshold (i.e., before the membrane potential decays too low to trigger). In at least one embodiment, neuron 2602 may be implemented using circuitry or logic that receives input, integrates the input to the membrane potential, and decays the membrane potential. In at least one embodiment, the input may be averaged, or any other suitable transfer function may be used. Furthermore, in at least one embodiment, neuron 2602 may include, but is not limited to, comparator circuitry or logic that generates an output spike at neuron output 2606 when the result of applying the transfer function to neuron input 2604 exceeds a threshold. In at least one embodiment, once neuron 2602 is triggered, it can ignore previously received input information by, for example, resetting the membrane potential to 0 or another suitable default value. In at least one embodiment, once the membrane potential is reset to 0, neuron 2602 may resume normal operation after a suitable period of time (or recovery period).

[0449] In at least one embodiment, neurons 2602 can be interconnected via synapses 2608. In at least one embodiment, synapses 2608 can be operated to transmit signals from the output of a first neuron 2602 to the input of a second neuron 2602. In at least one embodiment, neurons 2602 can transmit information on more than one instance of synapse 2608. In at least one embodiment, one or more instances of neuron output 2606 can be connected via instances of synapses 2608 to instances of neuron input 2604 in the same neuron 2602. In at least one embodiment, an instance of neuron 2602 that produces an output to be transmitted on the instance of synapse 2608 can be referred to as a "presynaptic neuron". In at least one embodiment, an instance of neuron 2602 that receives input transmitted via an instance of synapse 2608 can be referred to as a "postsynaptic neuron". In at least one embodiment, regarding various instances of synapse 2608, since an instance of neuron 2602 can receive input from one or more instances of synapse 2608 and can also transmit output through one or more instances of synapse 2608, a single instance of neuron 2602 can be both a "presynaptic neuron" and a "postsynaptic neuron".

[0450] In at least one embodiment, neurons 2602 may be organized into one or more layers. In at least one embodiment, each instance of neuron 2602 may have a neuron output 2606 that fans out to one or more neuron inputs 2604 via one or more synapses 2608. In at least one embodiment, the neuron output 2606 of neuron 2602 in the first layer 2610 may be connected to the neuron input 2604 of neuron 2602 in the second layer 2612. In at least one embodiment, layer 2610 may be referred to as a “feedforward layer.” In at least one embodiment, each instance of neuron 2602 in an instance of the first layer 2610 may fan out to each instance of neuron 2602 in the second layer 2612. In at least one embodiment, the first layer 2610 may be referred to as a “fully connected feedforward layer.” In at least one embodiment, each instance of neuron 2602 in an instance of the second layer 2612 fans out to fewer than all instances of neuron 2602 in the third layer 2614. In at least one embodiment, the second layer 2612 may be referred to as a “sparsely connected feedforward layer.” In at least one embodiment, neurons 2602 in the second layer 2612 may fan out to neurons 2602 in multiple other layers, including neurons 2602 fan out to the second layer 2612. In at least one embodiment, the second layer 2612 may be referred to as a “recurrent layer.” In at least one embodiment, the neuromorphic processor 2600 may be any suitable combination of recurrent layers and feedforward layers, including but not limited to sparsely connected feedforward layers and fully connected feedforward layers.

[0451] In at least one embodiment, the neuromorphic processor 2600 may include, but is not limited to, a reconfigurable interconnect architecture or dedicated hardwired interconnects to connect synapses 2608 to neurons 2602. In at least one embodiment, the neuromorphic processor 2600 may include, but is not limited to, circuitry or logic that allows synapses to be assigned to different neurons 2602 as needed, depending on the neural network topology and neuron fan-in / fan-out. For example, in at least one embodiment, synapses 2608 may be connected to neurons 2602 using interconnect structures such as on-chip networks or via dedicated connections. In at least one embodiment, synaptic interconnects and their components may be implemented using circuitry or logic.

[0452] Figure 27A processing system according to at least one embodiment is illustrated. In at least one embodiment, system 2700 includes one or more processors 2702 and one or more graphics processors 2708, and may be a single-processor desktop system, a multi-processor workstation system, or a server system having a large number of processors 2702 or processor cores 2707. In at least one embodiment, system 2700 is a processing platform incorporated within a system-on-a-chip (SoC) integrated circuit for use in mobile, handheld, or embedded devices.

[0453] In at least one embodiment, system 2700 may include or be integrated into a server-based gaming platform, including a game console, mobile game console, handheld game console, or online game console, which are game and media consoles. In at least one embodiment, system 2700 is a mobile phone, smartphone, tablet computing device, or mobile internet device. In at least one embodiment, processing system 2700 may also include components coupled to or integrated into a wearable device, such as a smartwatch wearable device, smart glasses device, augmented reality device, or virtual reality device. In at least one embodiment, processing system 2700 is a television or set-top box device having one or more processors 2702 and a graphical interface generated by one or more graphics processors 2708.

[0454] In at least one embodiment, each of the one or more processors 2702 includes one or more processor cores 2707 for processing instructions that, when executed, perform operations against the system and user software. In at least one embodiment, each of the one or more processor cores 2707 is configured to process a specific instruction sequence 2709. In at least one embodiment, the instruction sequence 2709 may facilitate Complex Instruction Set Computing (CISC), Reduced Instruction Set Computing (RISC), or computation via Very Long Instruction Word (VLIW). In at least one embodiment, each processor core 2707 may process a different instruction sequence 2709, which may include instructions that facilitate the emulation of other instruction sequences. In at least one embodiment, the processor core 2707 may also include other processing devices, such as a digital signal processor (DSP).

[0455] In at least one embodiment, processor 2702 includes cache memory 2704. In at least one embodiment, processor 2702 may have a single internal cache or multiple levels of internal caches. In at least one embodiment, the cache memory is shared among various components of processor 2702. In at least one embodiment, processor 2702 also uses an external cache (e.g., a Level 3 (L3) cache or a last-level cache (LLC)) (not shown), which can be shared among processor cores 2707 using known cache coherence techniques. In at least one embodiment, processor 2702 further includes a register file 2706, which may include different types of registers (e.g., integer registers, floating-point registers, status registers, and instruction pointer registers) for storing different types of data. In at least one embodiment, register file 2706 may include general-purpose registers or other registers.

[0456] In at least one embodiment, one or more processors 2702 are coupled to one or more interface buses 2710 to transmit communication signals, such as address, data, or control signals, between the processors 2702 and other components in the system 2700. In at least one embodiment, the interface bus 2710 may be a processor bus, such as a version of the Direct Media Interface (DMI) bus. In at least one embodiment, the interface bus 2710 is not limited to the DMI bus and may include one or more peripheral component interconnect buses (e.g., PCI, PCI Express), memory buses, or other types of interface buses. In at least one embodiment, the processor 2702 includes an integrated memory controller 2716 and a platform controller hub 2730. In at least one embodiment, the memory controller 2716 facilitates communication between memory devices and other components of the processing system 2700, while the platform controller hub (PCH) 2730 provides connectivity to input / output (I / O) devices via a local I / O bus.

[0457] In at least one embodiment, memory device 2720 may be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, a flash memory device, a phase-change memory device, or a device with suitable performance for use as processor memory. In at least one embodiment, memory device 2720 may be used as system memory of processing system 2700 to store data 2722 and instructions 2721 for use when one or more processors 2702 execute an application or process. In at least one embodiment, memory controller 2716 is also coupled to an optional external graphics processor 2712, which may communicate with one or more graphics processors 2708 of processor 2702 to perform graphics and media operations. In at least one embodiment, display device 2711 may be connected to processor 2702. In at least one embodiment, display device 2711 may include one or more internal display devices, such as in mobile electronic devices or laptop devices, or external display devices connected via a display interface (e.g., DisplayPort). In at least one embodiment, the display device 2711 may include a head-mounted display (HMD), such as a stereoscopic display device for virtual reality (VR) or augmented reality (AR) applications.

[0458] In at least one embodiment, the platform controller hub 2730 enables peripheral devices to connect to the storage device 2720 and the processor 2702 via a high-speed I / O bus. In at least one embodiment, the I / O peripheral devices include, but are not limited to, an audio controller 2746, a network controller 2734, a firmware interface 2728, a wireless transceiver 2726, a touch sensor 2725, and a data storage device 2724 (e.g., a hard disk drive, flash memory, etc.). In at least one embodiment, the data storage device 2724 may be connected via a storage interface (e.g., SATA) or via a peripheral bus, such as a peripheral component interconnect bus (e.g., PCI, PCIe). In at least one embodiment, the touch sensor 2725 may include a touchscreen sensor, a pressure sensor, or a fingerprint sensor. In at least one embodiment, the wireless transceiver 2726 may be a Wi-Fi transceiver, a Bluetooth transceiver, or a mobile network transceiver, such as a 3G, 4G, or LTE transceiver. In at least one embodiment, the firmware interface 2728 enables communication with the system firmware and may be, for example, a Unified Extensible Firmware Interface (UEFI). In at least one embodiment, network controller 2734 may enable network connectivity to a wired network. In at least one embodiment, a high-performance network controller (not shown) is coupled to interface bus 2710. In at least one embodiment, audio controller 2746 is a multi-channel high-definition audio controller. In at least one embodiment, processing system 2700 includes an optional legacy I / O controller 2740 for coupling legacy (e.g., Personal System 2 (PS / 2)) devices to system 2700. In at least one embodiment, platform controller hub 2730 may also be connected to one or more Universal Serial Bus (USB) controllers 2742 that connect input devices, such as a keyboard and mouse combination 2743, a camera 2744, or other USB input devices.

[0459] In at least one embodiment, instances of the memory controller 2716 and platform controller hub 2730 may be integrated into a discrete external graphics processor, such as external graphics processor 2712. In at least one embodiment, the platform controller hub 2730 and / or the memory controller 2716 may be external to one or more processors 2702. For example, in at least one embodiment, system 2700 may include an external memory controller 2716 and platform controller hub 2730...

Claims

1. A method comprising: The first dataset was generated using a machine learning model; The calculation is to be applied to a set of variations in the first dataset so that a set of features in the first dataset more closely matches a corresponding set of features in the second dataset, wherein the second dataset includes a sample of the privacy dataset; The set of changes is modified, at least in part, based on noise information corresponding to one or more constraints applied to the machine learning model, to generate a modified set of changes; and The machine learning model is updated using the modified set of changes. The set of changes is calculated by computing the Sinkhorn loss based at least in part on the first dataset and the second dataset.

2. The method of claim 1, wherein computing the set of changes further comprises: The gradient applied to the data in the first dataset is calculated at least in part based on the Sinkhorn loss and the cost matrix corresponding to the first dataset and the second dataset.

3. The method of claim 2, wherein modifying the set of changes further comprises: The gradient is clipped to reduce its vector magnitude.

4. The method of claim 2, wherein updating the machine learning model using the modified set of changes further comprises: Apply the function to the gradient.

5. The method of claim 1, wherein the noise information further includes Gaussian noise.

6. The method of claim 1, wherein using the machine learning model to generate the first dataset further comprises: Select the category for the first dataset.

7. The method of claim 1, wherein the privacy dataset is maintained by a data administrator and protected by a privacy barrier.

8. The method of claim 1, wherein the method further comprises: The machine learning model generates training data that simulates the privacy dataset; as well as The training data is provided to train one or more generative adversarial networks.

9. The method of claim 1, wherein updating the machine learning model using the modified set of changes further comprises: Modify a set of parameters of the machine learning model.

10. The method of claim 9, wherein the one or more constraints include privacy constraints, and wherein modifying the set of parameters of the machine learning model further includes: The gradient is modified with a value determined at least in part based on the privacy constraints.

11. The method of claim 1, wherein the first dataset further comprises a plurality of images.

12. The method of claim 1, wherein the first dataset further includes one or more category labels corresponding to the first dataset.

13. The method of claim 1, wherein the set of changes is calculated at least by computing a semi-debiased Sinkhorn loss based at least in part on the first dataset and the second dataset.

14. A processor including one or more circuits for generating a machine learning model, said machine learning model being trained at least in part by: A second set of changes to the machine learning model is determined at least in part based on noise information and a first set of changes, wherein the first set of changes is applied to a first dataset generated by the machine learning model to make the first dataset more closely match the second dataset; and The machine learning model is updated at least in part based on the second set of changes, wherein the second dataset includes data sampled from the privacy dataset. The first set of changes also includes calculating the optimal transmission plan in its original form based at least in part on samples from the first dataset and the second dataset. The calculation of the optimal transmission plan in the original form also includes calculating the Sinkhorn loss.

15. The processor of claim 14, wherein calculating the original optimal transfer plan further includes calculating the semi-debiased Sinkhorn loss.

16. The processor of claim 14, wherein determining the second set of changes further includes performing backpropagation using gradient descent to generate gradients for a set of parameters of the machine learning model.

17. A machine-readable medium having an instruction set stored thereon, the instruction set, if executed by one or more processors, causing the one or more processors to at least: A first set of changes is determined to make to a set of features in a first dataset to more closely match a corresponding set of features in a second dataset, the first dataset being at least partially based on a model, wherein the second dataset includes data sampled from a privacy dataset; The second set of changes to the model is determined at least in part based on the first set of changes; and The model is updated using the second set of changes. The instruction set that enables the one or more processors to determine the first set of variations further includes instructions that, if executed by the one or more processors, enable the one or more processors to determine the optimal transfer plan in the original form at least in part by calculating the Sinkhorn loss based on the first dataset and the second dataset.

18. The machine-readable medium of claim 17, wherein the instruction set further includes instructions that, if executed by the one or more processors, cause the one or more processors to compute gradients of the first set of changes.

19. The machine-readable medium of claim 18, wherein the instruction set that causes the one or more processors to compute the gradient of the first set of variations further includes instructions that, if executed by the one or more processors, cause the one or more processors to prune the gradient to reduce the maximum value included in the gradient.

20. The machine-readable medium of claim 19, wherein the instruction set further comprises instructions that, if executed by the one or more processors, cause the one or more processors to perform at least one of the following: Add noise to the gradient; The category of the first dataset is selected before it is generated by the model; Calculate the gradient of the second set of changes; and Modify the gradient by value.

21. The machine-readable medium of claim 17, wherein the first dataset further comprises a set of images.

22. The machine-readable medium of claim 21, wherein the set of images further comprises images of multiple categories.

23. The machine-readable medium of claim 17, wherein the model further comprises a parametric model that generates training data for training one or more neural networks; and The second set of changes also includes changes to the set of parameters determined at least in part based on the gradients of a set of parameters of the parametric model, the gradients of which are calculated at least in part based on the Sinkhorn loss between the first dataset and the second dataset, wherein the second dataset is sampled from a privacy dataset.

Images