Object detection method and system using neural network for weak supervision

Abstract

The invention discloses object detection with weak supervision using one or more neural networks, and specifically discloses devices, systems, and techniques for detecting objects in an image, including digital representations of these objects. In at least one embodiment, one or more objects in the image are detected based at least in part on one or more dummy tags corresponding to the one or more objects.

Description

[0001] This patent application is a divisional application of an invention patent application with an application date of June 10, 2020, application number 202010524624.9, and invention name “Object detection method and system using neural network for weak supervision”. Background Art

[0002] Advances in computer technology have led to an increased ability to identify and analyze objects. For the purposes of this analysis, machine learning has been used as a tool to detect objects in image data. In order to train machine learning, large amounts of labeled training data are required. Creating this training data can be a long and complex process. BRIEF DESCRIPTION OF THE DRAWINGS

[0003] The present disclosure will be described with reference to the accompanying drawings, in which:

[0004] Figure 1A , Figure 1B and Figure 1C illustrates an example image that may be processed or generated using one or more neural networks in at least one embodiment;

[0005] Figure 2 Components of an example multi-branch network in at least one embodiment are shown;

[0006] Figure 3A , Figure 3B , Figure 3C and Figure 3D illustrates components of an example network in at least one embodiment;

[0007] Figure 4A and Figure 4B An example drop block method in at least one embodiment is shown;

[0008] Figure 5 illustrates example object detection that can be reasoned about in at least one embodiment;

[0009] Fig. 6A and Figure 6B shows a portion of an example process for training a neural network to infer object representations and to use the network for inference in at least one embodiment;

[0010] Figure 7 illustrates an example environment in which aspects of at least one embodiment may be implemented;

[0011] Figure 8 An example system for training an image synthesis network that may be utilized in at least one embodiment is shown;

[0012] Fig. 9illustrates various layers of an example statistical model that may be utilized in at least one embodiment;

[0013] Fig.10 Inference and / or training logic in at least one embodiment is shown;

[0014] Fig.11 Inference and / or training logic in at least one embodiment is shown;

[0015] Fig.12 The training and deployment of a deep neural network in at least one embodiment is shown;

[0016] Fig.13 An example data center system in at least one embodiment is shown;

[0017] Fig.14 is a block diagram illustrating a computer system in at least one embodiment;

[0018] Fig.15 is a block diagram illustrating a computer system in at least one embodiment;

[0019] Fig.16 A computer system in at least one embodiment is shown;

[0020] Fig.17 A computer system according to at least one embodiment is shown;

[0021] Fig.18 An exemplary integrated circuit and associated graphics processor that may be fabricated using one or more IP cores in at least one embodiment described herein are shown;

[0022] Figure 19A-19B An exemplary integrated circuit and associated graphics processor that may be fabricated using one or more IP cores in at least one embodiment are shown;

[0023] Figure 20A-20B Additional exemplary graphics processor logic in at least one embodiment is shown;

[0024] Fig.21 A computer system in at least one embodiment is shown;

[0025] Fig. 22 A parallel processor in at least one embodiment is shown;

[0026] Fig.23 A partition unit in at least one embodiment is shown;

[0027] Fig.24 illustrates a processing cluster in at least one embodiment;

[0028] Fig.25 A graphics multiprocessor in at least one embodiment is shown;

[0029] Fig.26 is a block diagram illustrating a processor microarchitecture for a processor in at least one embodiment;

[0030] Fig. 27 A deep learning application processor in at least one embodiment is shown;

[0031] Fig.28 is a block diagram illustrating an example neuromorphic processor in at least one embodiment;

[0032] Fig.29 illustrates at least a portion of a graphics processor according to one or more embodiments;

[0033] Fig.30 is a block diagram of at least a portion of a graphics processor core in at least one embodiment;

[0034] Fig.31A and Fig.31B Thread execution logic in at least one embodiment is shown;

[0035] Fig.32 A parallel processing unit ("PPU") in at least one embodiment is shown.

[0036] Fig.33 illustrates a general processing cluster ("GPC") in at least one embodiment;

[0037] Fig.34 A memory partitioning unit of a parallel processing unit ("PPU") in at least one embodiment is shown; and

[0038] Fig.35 A streaming multiprocessor in at least one embodiment is shown. DETAILED DESCRIPTION

[0039] Figure 1A An example image 100 is shown that can be analyzed in at least one embodiment. The example image includes representations of multiple objects 102, including a man and a dog trying to herd a flock of sheep. In at least one embodiment, such images used for training are accompanied by specific labels, which include the type of object in the image and the location of the object. In at least one embodiment, the location information may include coordinates for a bounding box 142 that indicates the location of the object (e.g., Figure 1BHowever, as described above, in at least one embodiment, such annotations are not required for each training image, and the locations of various objects can be determined without specific labels. Figure 1B As shown, at least one embodiment can determine candidates (proposals) for different categories of objects and then use one or more refinements to reach the location of each object. In at least one embodiment, as Figure 1C As shown in the example image 180 of , the classification and position information can be used to detect and / or segment the object 182 represented in the image. It should be understood that in at least one embodiment, position information can also be used instead of bounding box coordinates.

[0040] In at least one embodiment, a framework is provided for instance-aware and context-aware weakly supervised object detection that can accept image-level labels without requiring additional spatial information. In at least one embodiment, such a framework can distinguish between multiple instances of the same type of object, rather than obtaining a single combined detection. In at least one embodiment, the example framework can also obtain accurate results for classes with large intra-class variance and can be memory-sustainable. Figure 2 An example framework 200 in at least one embodiment is shown in , where the framework includes an image branch 206 and at least one refinement branch 208, 210, 212. In at least one embodiment, the image branch can accept image-level labels and an image as input, and can supervise a detection model that produces outputs and a region of interest (ROI) score. In at least one embodiment, the ROI score can be an indication of the confidence that an object is represented in the region in at least one embodiment. In at least one embodiment, one or more refinement branches can take at least some of these outputs, such as the ROI score, and attempt to refine or improve the accuracy of the score in order to make a more accurate object detection determination. In at least one embodiment, the refinement branch can help resolve instance ambiguity and partial dominance.

[0041] In at least one embodiment, the detection of objects of interest can be performed using a multi-label image classification framework. In one embodiment, the object categories of interest can be collected in a set C. Given an input image I and a pre-computed set of region candidates R, in at least one embodiment, a convolutional layer ( Figure 2 The “Base” 202 in ( ) can be used to extract features, which can then be fed into the ROI pooling component to produce regional features for each region in the set R. In at least one embodiment, these features can be passed through a set of intermediate layers ( Figure 2204), and then passed through two different fully connected layers (such as via Figure 2 Box FC cls , FC det As shown), to generate classification logits for each object category c∈C and each region r∈R and number of detection pairs In this example, the vector w contains all trainable parameters, including Base, Neck, and Head. In at least one embodiment, when using softmax on the classification logit, the score s(c|r) of region r classified as category c is obtained, which can be given as follows:

[0042]

[0043] Similarly, in at least one embodiment, when using softmax on the detection logits, a score s(c|r) for class c for detection region r may be obtained:

[0044]

[0045] In at least one embodiment, the region of interest (ROI) score s used to assign a category c∈C to a region r∈R w The (c,r) set can be computed via element-wise product, for example, corresponding to s w (c,r)=s w (c|r)s w (r|c)∈[0,1]. For inference, in at least one embodiment, these ROI scores can be used to independently rank all regions of each category. In at least one embodiment, non-maximum suppression and thresholding can be applied to remove redundant boxes in at least one embodiment.

[0046] Since only image-level labels are available, in at least one embodiment, the ROI scores s for all regions r∈R can be calculated by w The loss function is constructed by summing (c, r) to obtain the image evidence of category c in the image (as follows):

[0047]

[0048] In at least one embodiment, a ground truth class label y(c)∈{0,1} may be used to indicate the image-level presence of class c in a given image, and the parameter w may be trained using a multi-label binary cross entropy loss such as may be given by:

[0049]

[0050] In at least one embodiment, the calculation corresponds to Figure 2 Image branch 206 is shown.

[0051] Instead of s from these ROI scores w (c, r) computes detections, and in at least one embodiment, the scores can be used to generate instance-level self-supervision for refinement branches. In at least one embodiment, the online refinement module can be used to encourage highly overlapping candidates to be assigned the same classification. For example, in at least one embodiment, the ROI score s w (c,r) can be used to generate pseudo labels Each region r can be assigned only one category.

[0052] With pseudo labels As self-supervision, in at least one embodiment, the refinement branch can solve instance-level classification by:

[0053]

[0054] where λ r is a hyperparameter and is the output of the first refinement branch, in this example, Figure 2 FC in ref The output is then passed through a class-wise softmax. In at least one embodiment, the hyperparameter λ r Can be used to adaptively balance and When the pseudo label As you gradually improve during training, you can In at least one embodiment, Figure 2 As shown, multiple refinement branches Can be stacked, and the current ROI score To calculate pseudo labels In at least one embodiment, all (excluding s w (c,r)) are averaged to obtain the final score.

[0055] In at least one embodiment, such a framework can produce convincing results for instance-level object detection. However, it is possible to obtain more accurate instance-level pseudo-labels, for example by paying sufficient attention to the context. In at least one embodiment, to obtain pseudo-labels (c, r), the highest scoring candidate can be the focus point. In at least one embodiment, for each category c, if region r and region r *,k (c) The intersection over union (“IoU”) between them is greater than 0.5, and the top ranking can be located by, for example, setting:

[0056]

[0057] It should be noted that in this example,

[0058] In order to find pseudo labels that are more suitable for instance-level WSOD In at least one embodiment, the spatial relationship of multiple instances in the image can be considered. For example, in at least one embodiment, all candidates r can be selected according to their ROI scores for each category c. to classify, rather than just using the highest ROI r *,k (c) operation. In at least one embodiment, the top p% of the highest ranking ROIs for each category c can be selected, and R'(c) can be used to refer to the result set. Thereafter, in at least one embodiment, a greedy algorithm can be used to select instances from R'(c), which is at least partially driven by the motivation that two non-overlapping and confident candidates have a high probability of corresponding to different objects. In at least one embodiment, the highest candidate can be added to the pseudo-label pool. Then, in at least one embodiment, starting from the second highest to the last, if the ROI does not overlap with another in the pool, the ROI can be added. In at least one embodiment, since the two instances can be relatively close, a very small threshold τ can be used to reject the ROI instead of strictly enforcing the zero overlap rule. Similarly, in at least one embodiment, candidates with an IoU greater than 0.5 and with any pseudo-label in the pool can be considered as positive examples during training.

[0059] In at least one embodiment, memory-efficient sequential batch forward and reverse computations can be used, which can be customized for WSOD models. In at least one embodiment, basic training via backpropagation can store all intermediate activations during the forward pass, which can be reused when calculating the gradients of network parameters. In at least one embodiment, such a method can be computationally efficient in part due to memory, but requires memory for the same reason. In at least one embodiment, only a subset of intermediate activations is saved during the forward pass of the key layer. In at least one embodiment, the entire model can be divided into smaller subnets at these key layers. In at least one embodiment, when calculating the gradient of a subnet, a forward pass can be applied starting from the activation stored at the input key layer of the subnet to obtain an intermediate representation of the subnet. Combined with the gradient propagated from an earlier subnet, in at least one embodiment, the gradient of the subnet weight can be calculated and the gradient can be propagated to the output of the earlier subnet.

[0060] In at least one embodiment, this is particularly advantageous for very deep networks in at least one embodiment, where memory costs can be distributed roughly evenly along the layers. However, in at least one embodiment, when these DNNs are adapted for detection, the activations (after ROI pooling) can grow from 1x CHW (image features) to N x CHW (ROI features), where N is in the thousands, for weakly supervised models. In at least one embodiment, these candidates can be kept to maintain good recall and therefore good performance in the absence of ground truth boxes.

[0061] In at least one embodiment, this training challenge can be addressed at least in part by using sequential batch backpropagation. In at least one embodiment, this can be accomplished in a manner such as Figure 3A In the Head submodule shown in Example 300, Figure 3B and Figure 3C In the Neck submodule shown in Examples 320 and 340, Figure 3D The Base submodule shown in example 360 ​​of is executed. The activations, gradients, and modules being updated in this step are shown in these figures (which illustrate sequential batch backpropagation in one iteration). In at least one embodiment, during the forward pass, the input image is first passed through Base and Neck, and the activations A are stored only after Base. b As shown, in at least one embodiment, the output of Neck then enters Head for its first forward and backward passes to update the weights and gradients G of Head. nIn at least one embodiment, to update the parameters of Neck, the ROI features can be divided into multiple batches, and back propagation can be run sequentially on the small batches. This approach can avoid storing memory-consuming feature maps and their gradients in Neck. In at least one embodiment, an example of this sequential approach is as follows Figure 3C As shown, 2000 candidates are divided into two batches, each batch of 1000 candidates. In at least one embodiment, as Figure 3D As shown, the cumulative gradient G b And it is used to update the parameters of the Base network via conventional back propagation. In at least one embodiment, for testing, if the number of ROIs or the size of Neck is too large, a similar strategy can be applied.

[0062] In at least one embodiment, due at least in part to intra-class variation, existing WSOD methods often incorrectly predict only the discriminative parts of an object rather than its full extent. In at least one embodiment, the DNN can be encouraged to focus on context which can be achieved by discarding some of the most discriminative parts (referred to herein as performing spatial dropout). Since the discriminative parts of an object vary in position and size, in at least one embodiment, object detection can utilize a structured dropout block approach for each region of interest, such as in Figure 4A and Figure 4B In at least one embodiment, during training of points on the ROI feature map, they may be sampled, for example, with a Bernoulli distribution with probability γ as blob centers. In at least one embodiment, the object region may be divided into an array of cells 402, such as Figure 4A 402 as shown in the example of . In at least one embodiment, as Figure 4B As shown in example 450 of , square or rectangular regions 452, 454 around these centers of size wxw can be zeroed, for example by setting the relevant pixel values ​​to zero or another set value across all channels on the feature map. For example, in at least one embodiment, the feature values ​​can then be rescaled by a factor of the ROI area over the area of ​​the non-discarded area so that normalization does not have to be applied to reason that all features are retained. In at least one embodiment, distinctive parts such as faces, car heads, and heads are zeroed.

[0063] In at least one embodiment, models for object detection and / or segmentation may be trained using data that does not contain very detailed labels, and in some cases may simply indicate the class of the object represented in the image or video frame, without providing any location or occurrence information. In at least one embodiment, the label may simply state that at least one sheep and at least one dog are represented in the image, but without indicating how many sheep or dogs are represented, or where these representations may be located in the image. In at least one embodiment, these types of labels are referred to herein as image-level labels. In at least one embodiment, a fully supervised approach will require bounding box coordinates or other location information to determine the size of the object boundaries and the location of the object representation in the image in order to accurately train the model.

[0064] In at least one embodiment, there may be different numbers and / or selections of refinement branches and functions. In at least one embodiment, the image detection branch may receive and / or provide candidates for multiple bounding boxes, each of which corresponds to a potential object representation, or a representation of at least one type of object or an object that meets certain criteria, etc. In at least one embodiment, a separate object region candidate network or algorithm will generate an initial set of candidates. In at least one embodiment, the image detection branch outputs a score or set of scores for the object candidates and one or more classifications, and in at least one embodiment, the number of candidates may be reduced based at least in part on scores and / or redundant candidates. In at least one embodiment, the output of the image branch for a given table is a matrix of dimension n (for multiple candidates) by c (for multiple identified categories). In at least one embodiment, each element of the matrix is ​​a confidence score (or other such score) indicating the confidence that a given object candidate belongs to a particular category. In at least one embodiment, these results may include FCs corresponding to post-processing results from the table. cls and FC det Value. FC cls and FC det Each may be a different nxc matrix derived from the original nxc matrix for the initial image candidate. In at least one embodiment, each matrix may be normalized, for example, by utilizing a softmax function along both the candidate dimension and the classification dimension. In at least one embodiment, normalization may provide easier comparison. In at least one embodiment, a higher element score may indicate a higher likelihood of a correctly identified object and classification. Then, in at least one embodiment, as Figure 2The tables can be multiplied backwards as shown. In at least one embodiment, the result is a new n×c matrix output by the network. In at least one embodiment, the result is a new n×c matrix output by the network. In at least one embodiment, global, adaptive, and / or other such thresholds can be applied, and any score below the threshold can be determined to not correspond to a correct actual object representation. In at least one embodiment, elements with scores equal to or above the threshold in at least one embodiment can be determined to be correct object representations in at least one embodiment.

[0065] Since image-level labels are available, in at least one embodiment, summation can be performed across individual instances along the n-dimension of the matrix. In at least one embodiment, the result is a c-dimensional pool vector supervised by the image-level labels. In at least one embodiment, this approach provides a multi-label loss. In at least one embodiment, the loss can supervise the vector, which can backpropagate the gradients used to train the network. In at least one embodiment, the sum pooling process can help reduce a large initial candidate table to a form that is more like an image-level representation, which can then be more easily refined. In at least one embodiment, the refinement branch follows a teacher / student model. In at least one embodiment, it treats the labels from the initial prediction as a set of candidates with corresponding detection scores. In at least one embodiment, for each object category from the image-level labels, the candidates can be analyzed and their scores are ranked. In at least one embodiment, all scores can be normalized to values ​​between 0 and 1. In at least one embodiment, a global threshold can be applied, so that the value of any candidate with a value equal to or exceeding the threshold is set to 1, and the value of any candidate with a value less than the threshold can be set to 0. In at least one embodiment, this process results in quantization of all scores. In at least one embodiment, the input is the score from the previous branch, but due to this quantization, a set of pseudo labels is generated for each refinement branch. An exemplary refinement branch can produce an nxc matrix with all element values ​​0 or 1. In at least one embodiment, the pseudo labels can be used to define the loss, which can also help refine the branch.

[0066] In at least one embodiment, the positions of the initial seeds or individual object candidates are prevented from becoming very close to each other. In at least one embodiment, a sorting of top windows can be performed, where a "top window" can correspond to a bounding box of a seed or candidate with a high confidence score. In at least one embodiment, many of these top windows may be very close in the image because multiple windows may correspond to different views or approximations of an object represented in the image. In at least one embodiment, a top window that is determined to be very close to another top window or similar in size and position can be excluded from consideration. In at least one embodiment, a greedy method with maximum suppression can be used to exclude redundant windows or candidates from consideration because they are likely to correspond to the same object representation. In at least one embodiment, quantization of the nxc matrix can be performed again, where the initial (highest value similar) window is set to a value of 1 and the redundant or similar window is set to a value of 0. In at least one embodiment, a large set of different seeds that can be used as a starting point can then be selected or identified. In at least one embodiment, the elements of the matrix or pseudo-label for each candidate can be adjusted so that any window that significantly overlaps with another window will be excluded from consideration. In at least one embodiment, the process is a form of label propagation, which can help prevent partial dominance problems.

[0067] In at least one embodiment, the thinning method is referred to herein as drop-block thinning. In at least one embodiment, this thinning method may remove one or more sub-regions or portions of an object candidate region from consideration. Figure 4A and Figure 4B An example discard block method that may be utilized in accordance with at least one embodiment is shown.In at least one embodiment, a set of "blocks" (where a block represents a pixel or group of pixels in an image) are discarded or removed from consideration for candidate regions of an image. Figure 4A The exemplary case 400 in FIG. 4 shows segmenting the object candidate region into blocks. Figure 4BA situation 450 is shown where a subset of these blocks are removed from consideration in at least one embodiment. In at least one embodiment, the blocks and block sizes may be selected randomly or according to a block selection algorithm. In at least one embodiment, the remaining portion of the candidate data may be analyzed to determine whether a similar confidence score is obtained, indicating that the region may represent an object of that class. In at least one embodiment, this may also help train the network to reason about the correct classification of objects for a variety of different views or parts of the object. In at least one embodiment, drops may be selected from regions (e.g., faces) to be discarded, while other portions will provide high confidence regions for one type of object. In at least one embodiment, the process of randomly discarding blocks may also effectively introduce some noise into image regions, which helps improve the ability of the network to recognize different object representations after training.

[0068] In at least one embodiment, further advantages can be obtained that are at least partially independent of the selected refinement method. In at least one embodiment, the memory consumption problem can be solved by utilizing batch backpropagation techniques. In at least one embodiment, the data can be divided into smaller batches that can be processed individually and / or in parallel, so that for any given backpropagation task, less data needs to be stored in memory. In at least one embodiment, the processing of the various batches can also be interleaved so that all data does not need to be processed simultaneously and / or all data is stored in the processor memory at the same time. In at least one embodiment, this reduction in memory consumption also enables object detection and / or segmentation to be performed on video or other large data sets, streams or sources.

[0069] In at least one embodiment, the refinement branches all operate in parallel, taking as input the candidate data from the detection branch. In at least one embodiment, the results from the various refinement branches can then be aggregated with the results from the object detection branch, such as by averaging or using a selected aggregation function that computes a weighted or other function of the various values ​​of the candidates. In at least one embodiment, candidates that do not meet the various refinement requirements can have their scores significantly reduced, which can result in candidate values ​​being below a threshold and therefore considered not to correspond to an object of the relevant class.

[0070] In at least one embodiment, the refinement branches can each take the outputs (e.g., ROI scores) from the previous branches in a specified order. In at least one embodiment, such an approach can save additional processing steps because the ROI scores from the image branch will be refined as they pass through the various refinement branches. In at least one embodiment, some precise object representations may have corresponding ROI scores that are inappropriately lowered by the refinement branches, to the extent that the candidates are removed from consideration. In at least one embodiment, performing refinement alone allows object candidates to remain in consideration if, for example, all but one refinement branch indicate a high confidence determination for an object of a given class.

[0071] In at least one embodiment, such a process is adaptive in that it can be performed for multiple instances of one type of object as well as multiple types of objects represented in a single image or video frame. In at least one embodiment, the generated pseudo-labels are used as a type of ground truth data to provide at least some constraints for learning. In at least one embodiment, multiple pseudo-labels with scores above a threshold are determined and used to refine the scores of object candidates from the initial object detection branch or process. In at least one embodiment, the initial candidate scores are averaged with the refined scores to generate a final score, which can then be compared to a threshold or otherwise processed to determine which object candidates are accepted as accurate.

[0072] Figure 5 A collection of example images 500 including object detection regions that can be inferred according to at least one embodiment is shown. In at least one embodiment, for each pair of images, the results of different methods are shown on the left and the results of the method according to at least one embodiment are shown on the right. In at least one embodiment, each column in a pair also illustrates that the method according to at least one embodiment improves the results for a specific problem. For example, in at least one embodiment, the left column shows that a method does not miss instances of a certain type of object in an image. In at least one embodiment, the middle column shows that the method according to at least one embodiment does not have problems detecting objects when object candidate regions are initially grouped in the image. In addition, in at least one embodiment, the right column shows that the method according to at least one embodiment does not determine object regions in an image based on regions of a face or grill that partially control, control detection, and may miss other parts of the object.

[0073] Fig. 6AAn example process 600 for training a neural network (e.g., a generative adversarial network (GAN)) to infer object detection and / or segmentation in at least one embodiment is shown. For this process and other processes discussed herein, it should be understood that, unless otherwise stated, additional, alternative, or fewer steps may be performed in a similar or alternative order or in parallel in at least one embodiment. In addition, this example discusses training a generative adversarial network (GAN) using alphanumeric semantic description data, but as discussed elsewhere herein, various types of models may be trained using various different types of data within the scope of at least one embodiment. In at least one embodiment, an image (or video frame, etc.) including at least one representation of an object type of interest is obtained 602, wherein the image includes one or more image-level labels indicating one or more object categories represented in the image. In at least one embodiment, the labels may not include any location or instance data, etc. In at least one embodiment, a set of object candidates may be determined, whether external or internal to a neural network, such as the GAN in this example. In at least one embodiment, the object candidate regions may be input to the detection branch of the GAN, which may determine a region of interest (ROI) score or similar such value for each candidate, as discussed elsewhere herein, and in at least one embodiment, may reduce the number of candidates based at least in part on the determined score. In at least one embodiment, the ROI scores for each candidate may be passed to one or more refinement branches of the network, which may determine pseudo labels 606 for each candidate. In at least one embodiment, these pseudo labels may be associated with new or updated ROI scores for each candidate determined using the criteria or methods of the associated refinement branch. In at least one embodiment, the ROI scores from the object detection branch are combined 608 with the pseudo label scores from the refinement branch to generate a set of final ROI scores for each candidate and class. In at least one embodiment, the set of final ROI scores may be used to determine a loss function 610. The network parameters of the GAN may then be updated 612 based at least in part on the determined loss function.

[0074] Figure 6B An example process 650 for inferring object representations in an image using such a trained model in at least one embodiment is shown. In at least one embodiment, an image or video frame for inference is obtained 652. In at least one embodiment, the image may be provided as input to the trained model 654. In at least one embodiment, the trained model may process the data and infer one or more classes and locations of objects represented in the image 656. In at least one embodiment, the inferred locations may correspond to segmentations of the objects in at least one embodiment.

[0075] More and more industries and applications are leveraging machine learning. For example, deep neural networks (DNNs) developed on processors have been used in a wide variety of use cases, from self-driving cars to faster drug development, from automatic image analysis for security systems to intelligent real-time language translation in video chat applications. Deep learning is a technology that models the neural learning process of the human brain and continuously learns over time, getting smarter and providing more accurate results faster. Children are initially taught by adults to correctly identify and classify various shapes, and eventually are able to recognize shapes without any guidance. Similarly, a deep learning or neural learning system designed to accomplish similar tasks needs to be trained so that it becomes smarter and more efficient in recognizing basic objects, occluded objects, etc., while also assigning context to these objects.

[0076] At the simplest level, neurons in the human brain look at the various inputs they receive, assign a level of importance to each of these inputs, and pass outputs to other neurons for processing. An artificial neuron, or perceptron, is the most basic model of a neural network. In one example, a perceptron can receive one or more inputs representing various features of an object that the perceptron is trained to recognize and classify, and assign a certain weight to each of these features based on the importance of that feature in defining the shape of the object.

[0077] Deep neural network (DNN) models include multiple layers of many connected perceptrons (e.g., nodes) that can be trained with large amounts of input data to solve complex problems quickly and with high accuracy. In one example, the first layer of a DNN model breaks down an input image of a car into different parts and looks for basic patterns, such as lines and angles. The second layer assembles lines to look for higher-level patterns, such as wheels, windshields, and mirrors. The next layer identifies the type of vehicle, and the final layers generate labels for the input image, identifying models of specific car brands. Once a DNN is trained, it can be deployed and used to recognize and classify objects or patterns in a process known as reasoning. Examples of reasoning (the process by which a DNN extracts useful information from a given input) include recognizing handwritten numbers on checks deposited into an ATM machine, identifying images of friends in photos, providing movie recommendations, identifying and classifying different types of cars, pedestrians, and road hazards in self-driving cars, or translating human language in near real time.

[0078] During training, data flows through the DNN in the forward propagation phase until a prediction is produced indicating the label corresponding to the input. If the neural network does not label the input correctly, the error between the correct label and the predicted label is analyzed, and the weights of each feature are adjusted in the backpropagation phase until the DNN correctly labels this and other inputs in the training dataset. Training complex neural networks requires a large amount of parallel computing performance, including supporting floating-point multiplication and addition. Inference, which is less computationally intensive than training, is a latency-sensitive process in which the trained neural network is applied to new inputs it has never seen before to classify images, translate speech, and reason about new information.

[0079] Neural networks rely heavily on matrix math operations, and complex multi-layer networks require a lot of floating-point performance and bandwidth for both efficiency and speed. Computing platforms with thousands of processing cores optimized for matrix math operations and delivering tens to hundreds of TFLOPS of performance can deliver the performance required for deep neural network-based artificial intelligence and machine learning applications.

[0080] Figure 7 Components of an example system 700 that can be used to train and utilize machine learning in accordance with at least one embodiment are shown. As will be discussed, various components can be provided by various combinations of computing devices and resources or a single computing system that can be under the control of a single entity or multiple entities. In addition, various aspects can be triggered, initiated, or requested by different entities. For example, in at least one embodiment, the training of a neural network can be directed by a vendor associated with a vendor environment 706, while in at least one embodiment, the training of a neural network can be requested by a customer or other user who can access the vendor environment through a client device 702 or other such resource. Training data (or data to be analyzed by the trained neural network) can be provided by a vendor, a user, or a third-party content provider 724. In at least one embodiment, a client device 702 can be a vehicle or an object that can navigate on behalf of a user, for example, the user can submit requests and / or receive instructions to facilitate navigation of the device.

[0081] In this example, the request can be submitted through at least one network 704 to be received into the provider environment 706. The client device can be any suitable electronic and / or computing device that enables a user to generate and send such a request, such as a desktop computer, a laptop computer, a computer server, a smart phone, a tablet computer, a game console (portable or otherwise), a computer processor, a computing logic, and a set-top box. The one or more networks 704 may include any suitable network for sending requests or other such data, such as the Internet, an intranet, an Ethernet, a cellular network, a local area network (LAN), a network with direct wireless connections between nodes, and the like.

[0082] The request may be received to the interface layer 708, which in this example may forward the data to the training and reasoning manager 710. In at least one embodiment, this manager may be a system or service including hardware and software for managing services and requests consistent with data or content. This manager may receive a request to train a neural network and may provide the data of this request to the training manager 712. If this request is not specified, the training manager 712 may select an appropriate model or network to be used and may use the relevant training data to train this model. In at least one embodiment, the training data may be a batch of data received from the client device 702 or obtained from a third-party supplier 724 and stored in the training data repository 714. The training manager 712 may be responsible for training this data, for example, by using the LARC-based method discussed herein. The network may be any appropriate network, such as a recurrent neural network (RNN) or a convolutional neural network (CNN). Once the network is trained and successfully evaluated, the trained network may be stored in the model repository 716, for example, which may store different models or networks for users, applications, services, etc. As described above, in at least one embodiment, multiple models may exist for a single application or entity, such that multiple models may be utilized based on a number of different factors.

[0083] At a subsequent point in time, a request for content (e.g., path determination) or data that is at least partially determined or influenced by a trained neural network may be received from the client device 702 (or another such device). The request may include, for example, input data to be processed using the neural network to obtain one or more inferences or other output values, classifications, or predictions. Although different systems or services may also be used in at least one embodiment, the input data may be received to the interface layer 708 and directed to the inference module 718. If not already stored locally to the inference module 718, the inference module 718 may obtain an appropriate training network from the model repository 716, such as a trained deep neural network (DNN) described herein. The inference module 718 may provide data as input to the trained network, and may then generate one or more inferences as output. For example, this may include a classification of an instance of the input data. The inference may then be sent to the client device 702 for display to the user or other communication with the user. The user's environmental data may also be stored in a user environmental data repository 722, which may include data about the user that may be used as a network input to generate inferences or determine data returned to the user after obtaining an instance. Related data, including at least a portion of the input or inference data, may also be stored in a local database 720 for use in processing future requests. In at least one embodiment, a user may use an account or other information to access resources or functionality of a provider environment. If permitted and available, user data may also be collected and used to further train the model to provide more accurate inferences for future requests. In at least one embodiment, a request for a machine learning application 726 executed on a client device 702 may be received through a user interface, and the results may be displayed through the same interface. The client device may include resources such as a processor 728 and a memory 730 for generating this request and processing results or responses, and at least one data storage element 732 for storing data for the machine learning application 726.

[0084] In at least one embodiment, the processor 728 (or the processor of the training manager 712 or the inference module 718) will be a central processing unit (CPU). However, as described above, resources in such an environment may utilize GPUs to process data for at least some types of requests. GPUs have thousands of cores and are designed to handle massively parallel workloads, and have therefore become popular in deep learning for training neural networks and generating predictions. While using GPUs for offline construction can train larger and more complex models faster, generating predictions offline means that request-time input features cannot be used, or predictions must be generated for all feature permutations and stored in lookup tables to serve real-time requests. If the deep learning framework supports CPU mode, and the model is small and simple enough that feedforward can be performed on the CPU with reasonable latency, a service on a CPU instance can host the model. In this case, training can be performed offline on the GPU and inference can be performed in real time on this CPU. If the CPU approach is not a viable option, the service can run on a GPU instance. However, because GPUs have different performance and cost characteristics than CPUs, running a service that offloads runtime algorithms to GPUs may require that it be designed differently than a CPU-based service.

[0085] Figure 8An example system 800 that can be used to classify data or generate inferences according to at least one embodiment is shown. Based on the teachings and suggestions contained herein, it should be apparent that various types of predictions, labels, or other outputs can also be generated for input data. In addition, supervised training and unsupervised training can be used in at least one embodiment discussed herein. In this example, a set of training data 802 (e.g., classified or labeled data) is provided as input to be used as training data. The training data may include instances of at least one type of object for which the neural network is to be trained, and information identifying the type of object. For example, the training data may include a set of images, each image containing a representation of an object type, wherein each image also contains or is associated with a label, metadata, classification, or other information identifying the type of object represented in the respective image. Various other types of data may also be used as training data, and may include text data, audio data, video data, etc. In this example, the training data 802 is provided as training input to a training manager 804. The training manager 804 can be a system or service including hardware and software, such as one or more computing devices that execute a training application for training a neural network (or other model or algorithm, etc.). In this example, the training manager 804 receives an instruction or request indicating the type of model to be used for training. The model can be any suitable statistical model, network or algorithm that can be used for such purposes, and may include, for example, an artificial neural network, a deep learning algorithm, a learning classifier, a Bayesian network, etc. The training manager 804 can select an initial model or other untrained model from an appropriate repository 806, and use the training data 802 to train this model to generate a trained model 808 (e.g., a trained deep neural network) that can be used to classify similar types of data, or generate other such inferences. In at least one embodiment where training data is not used, the input data can still be trained based on the selection of an appropriate initial model by the training manager 804.

[0086] Models can be trained in a variety of different ways, which may depend in part on the type of model selected. For example, in at least one embodiment, a set of training data can be provided to a machine learning algorithm, where this model is a model artifact created by a training process. Each instance of the training data contains a correct answer (e.g., a classification), which can be referred to as a target or target attribute. The learning algorithm finds patterns in the training data that map the input data attributes to the target, the answer to be predicted, and outputs a machine learning model that captures these patterns. The machine learning model can then be used to obtain predictions for new data for which a target is not specified.

[0087] In one example, the training manager 804 can select from a set of machine learning models, including binary classification, multi-classification, and regression models. The type of model to be used can depend at least in part on the type of target to be predicted. Machine learning models for binary classification problems can predict binary outcomes, such as one of two possible classes. Learning algorithms (such as logistic regression) can be used to train binary classification models. Machine learning models for multi-class classification problems allow predictions to be generated for multiple classes, such as predicting one of more than two outcomes. Multinomial logistic regression can be useful for training multi-class models. Machine learning models for regression problems can predict numerical values. Linear regression is useful for training regression models.

[0088] In order to train a machine learning model according to at least one embodiment, the training manager must determine the input training data source and other information, such as the name of the data attribute containing the target to be predicted, the required data transformation instructions, and training parameters to control the learning algorithm. During the training process, in at least one embodiment, the training manager 804 can automatically select an appropriate learning algorithm based on the target type specified in the training data source. The machine learning algorithm can accept parameters for controlling the training process and certain properties of the resulting machine learning model. These are referred to as training parameters in this article. If no training parameters are specified, the training manager can use known default values ​​to handle a wide range of machine learning tasks well. Examples of training parameters for which values ​​can be specified include the maximum model size, the maximum number of times through the training data, the type of random play, the type of regularization, the learning rate, and the amount of regularization. Default settings can be specified, with options for adjusting values ​​to fine-tune performance.

[0089] The maximum model size is the total size in bytes of the patterns created during training the model. By default, a model of the specified size is created, for example, a 100MB model. If the training manager cannot determine that there are enough patterns to fill the model size, a smaller model may be created. If the training manager finds that there are more patterns than can fit within the specified size, it may enforce a maximum cutoff by trimming the patterns that have the least impact on the quality of the learned model. Choosing a model size provides control over the tradeoff between the predictive quality of the model and the cost of using it. Smaller models may cause the training manager to remove many patterns to fit within the maximum size limit, affecting the quality of the predictions. Larger models may be more expensive to query for live predictions. Larger input datasets do not necessarily result in larger models, because the model stores the patterns rather than the input data. If the patterns are few and simple, the resulting model will be small. Input data with a large number of raw attributes (input columns) or derived features (output of data transformations) may find and store more patterns during the training process.

[0090] In at least one embodiment, the training manager 804 may make multiple passes or iterations over the training data to try to discover patterns. There may be a default number of passes, such as ten, and in at least one embodiment, a maximum number of passes may be set, such as up to one hundred passes. In at least one embodiment, there may be no maximum set, or there may be a convergence criterion or other set of factors that trigger the end of the training process. In at least one embodiment, the training manager 804 may monitor the quality of the pattern during training and may automatically stop training when there are no more data points or patterns to discover. Data sets with only a small number of observations may require more data passes to obtain a sufficiently high model quality. Larger data sets may contain many similar data points, which can reduce the need for a large number of passes. The potential impact of selecting more passes through the data is that model training may take longer and cost more in terms of resources and system utilization.

[0091] In at least one embodiment, the training data is shuffled before training or between passes of training. In at least one embodiment, shuffling is a random or pseudo-random shuffle for generating a truly random ordering, although there may be some constraints to ensure that certain types of data are not grouped, or if such grouping exists, the shuffled data can be reshuffled, etc. Shuffling changes the sequence or arrangement in which the data is used for training so that the training algorithm does not encounter groupings of similar types of data or single-type data that is observed too many times in a row. For example, a model can be trained to predict objects. Before uploading, the data may be sorted by object type. Then, the algorithm can process the data in alphabetical sequence by object type, first encountering only data of a specific object type. The model will begin to learn the pattern of that type of object. Then, the model will only encounter data for the second object type, and will try to adjust the model to fit that object type, which may degrade the pattern that fits the first object type. This sudden switch between object types may produce a model that cannot learn how to accurately predict the object type. Before dividing the training data set into a training subset and an evaluation subset, in at least one embodiment, shuffling can be performed to utilize a relatively uniform distribution of data types for both stages. In at least one embodiment, training manager 804 may automatically shuffle the data using, for example, a pseudo-random shuffling technique.

[0092] In at least one embodiment, when creating a machine learning model, the training manager 804 can enable the user to specify settings. For example, the user can specify one or more evaluation settings to indicate a portion of the input data to be retained for evaluating the predictive quality of the machine learning model. The user can specify a policy that indicates which attributes and attribute transformations can be used for model training. The user can also specify various training parameters that control the training process and certain properties of the resulting model.

[0093] Once the training manager determines that the training of the model is completed, such as by using at least one final criterion discussed herein, the trained model 808 can be provided for the classifier 814 to classify (or otherwise generate reasoning) verification data 812. As shown, this involves a logical conversion between the training mode of the model and the reasoning mode of the model. However, in at least one embodiment, the trained model 808 will first be passed to the evaluator 810, which may include an application, process or service executed on at least one computing resource (e.g., a CPU or GPU of at least one server) for evaluating the quality (or other aspects) of the trained model. The model is evaluated to determine whether the model will at least provide a minimum acceptable or threshold level of performance when predicting the target of new data and future data. If not, the training manager 804 can continue to train the model. Since future data instances will generally have unknown target values, it may be desirable to check the accuracy metric of machine learning on data with known target answers and use the evaluation as a proxy for the prediction accuracy of future data.

[0094] In at least one embodiment, a subset of the training data 802 provided for training is used to evaluate the model. The subset can be determined using the shuffling and splitting method described above. This evaluation data subset will be marked with a target, so it can serve as a resource for evaluating ground truth. It is useless to use the same data used for training to evaluate the prediction accuracy of the machine learning model, because a positive evaluation may be generated for a model that remembers the training data instead of generalizing it. Once the training is completed, the evaluation data subset is processed using the trained model 808, and the evaluator 810 can determine the accuracy of the model by comparing the ground truth data with the corresponding output (or prediction / observation) of the model. In at least one embodiment, the evaluator 810 can provide a summary or performance metric that indicates the degree of match between the predicted value and the true value. If the trained model does not meet at least the minimum performance standard or other such accuracy threshold, the training manager 804 can be instructed to perform further training, or in some cases, try to train a new or different model. If the trained model 808 meets the relevant standards, the trained model can be provided for use by the classifier 814.

[0095] When creating and training a machine learning model, in at least one embodiment, it may be desirable to specify model settings or training parameters that will result in a model that is capable of making the most accurate predictions. Example parameters include the number of passes to be performed (forward and / or backward), regularization, model size, and shuffle type. However, as described above, selecting the model parameter settings that produce the best predictive performance on the evaluation data may result in overfitting the model. Overfitting occurs when the model stores patterns that appear in the training and evaluation data sources, but fails to generalize patterns in the data. Overfitting often occurs when the training data includes all the data used in the evaluation. An overfitted model may perform well during evaluation, but may not make accurate predictions on new data or other validation data. To avoid selecting an overfitted model as the best model, the training manager may retain additional data to validate the performance of this model. For example, the training data set may be divided into 60% for training and 40% for evaluation or validation, which may be divided into two or more stages. After the model parameters that best fit the evaluation data are selected, resulting in convergence to a subset of the validation data (e.g., half of the validation data), a second validation may be performed using the remaining validation data to ensure the performance of this model. If the model meets the expectations on the validation data, then the model is not overfitting the data. Optionally, a test set or holdout set can be used to test parameters. Using a second validation or testing step can help choose appropriate model parameters to prevent overfitting. However, taking more data out of the training process for validation leaves less data available for training. This can be problematic for smaller datasets because there may not be enough data available for training. One approach in this case is to perform cross-validation, as described elsewhere in this article.

[0096] There are many metrics or insights that can be used to review and evaluate the predictive accuracy of a given model. An example evaluation result contains a predictive accuracy metric to report the overall success of the model, as well as visualizations to help explore the accuracy of this model beyond the predictive accuracy metric. The results can also provide the ability to see the impact of setting score thresholds (such as binary classification) and can generate alerts on criteria used to check the effectiveness of the evaluation. The choice of metrics and visualizations can depend at least in part on the type of model being evaluated.

[0097] After satisfactory training and evaluation, the trained machine learning model can be used to build or support machine learning applications. In at least one embodiment, building a machine learning application is an iterative process involving a series of steps. The core machine learning problem can be constructed based on what is observed and the answer that this model is to predict. Data can then be collected, cleaned, and prepared to make this data suitable for use by the machine learning model training algorithm. This data can be visualized and analyzed for integrity checks to verify the quality of this data and understand this data. The original data (such as input variables) and the answer data (such as targets) may not be represented in a way that can be used to train a highly predictive model. Therefore, it may be desirable to build a more predictive input representation or feature from the original variables. The resulting features can be input into the learning algorithm to build the model and evaluate the quality of the model based on the data retained from the model building. The model can then be used to generate predictions of the target answer for new data instances.

[0098] exist Figure 8 In the exemplary system 800 of , after providing the evaluation, this trained model 810 is provided to or made available to a classifier 814, which is capable of using the trained model to process validation data. For example, this may include data received from a user or an unclassified third party, such as a query image that is being queried for information about the content represented in these images. The validation data can be processed by the classifier using this trained model, and the resulting results 816 can be sent back to the corresponding source or otherwise processed or stored. In at least one embodiment, and where such use is permitted, these currently classified data instances can be stored in a training data repository, which can be used by a training manager for further training of the trained model 808. In at least one embodiment, the models will be continuously trained as new data becomes available, but in at least one embodiment, these models will be trained periodically, such as once a day or once a week, depending on factors such as the size of the data set or the complexity of the model.

[0099] The classifier 814 may include appropriate hardware and software for processing the validation data 812 using the trained model. In at least one embodiment, the classifier will include one or more computer servers, each having one or more graphics processing units (GPUs) capable of processing data. The configuration and design of the GPUs may make them more suitable for processing machine learning data than CPUs or other such components. In at least one embodiment, the trained model can be loaded into the GPU memory, and the received data instances can be provided to the GPU for processing. The GPU can have many more cores than the CPU, and the GPU core can be less complex. Therefore, a given GPU may be able to process thousands of data instances simultaneously through different hardware threads. The GPU can also be configured to maximize floating point throughput, which can provide significant additional processing advantages for large data sets.

[0100] Even when GPUs, accelerators, and other such hardware are used to accelerate tasks such as model training or data classification using such models, such tasks may still require significant time, resource allocation, and cost. For example, if a machine learning model is to be trained using 800 passes, and the data set includes 1,000,000 data instances to be used for training, each pass will need to process all million instances. Different parts of the architecture may also be supported by different types of devices. For example, training may be performed using a set of servers at a logically centralized location, such as may be provided as a service, while classification of the raw data may be performed by such a service or on a client device. In at least one embodiment, these devices may also be owned, operated, or controlled by the same entity or entities.

[0101] Fig. 9 An example neural network 900 that can be trained or otherwise utilized according to at least one embodiment is shown. In this example, the statistical model is an artificial neural network (ANN) that includes multiple node layers, including an input layer 902, an output layer 906, and multiple layers 904 of intermediate nodes, typically referred to as "hidden" layers, because the internal layers and nodes are typically not visible or accessible in various neural networks. Although several intermediate layers are shown for explanation purposes only, it should be understood that there is no limit to the number of intermediate layers that can be utilized, and any limit on layers will generally be a factor of the resources or time required to process using this model. As discussed elsewhere herein, other types of models, networks, algorithms, or processes may also be used, which may include other numbers or selections of nodes and layers. Validation data may be processed by each layer of this network to generate a set of inference or reasoning scores, which may then be fed to a loss function 908.

[0102] In this example network 900, all nodes of a given layer are interconnected to all nodes of adjacent layers. As shown, the nodes of the middle layer will then be connected to the nodes of two adjacent layers respectively. In some models, nodes are also called neurons or connected units, and the connections between nodes are called edges. Each node can perform a function for the received input, such as by using a specified function. Nodes and edges can obtain different weights during the training process, and the various layers of nodes can perform specific types of transformations on the received input, and these transformations can also be learned or adjusted during the training process. Learning can be supervised learning or unsupervised learning, which may depend at least in part on the type of information contained in the training data set. Various types of neural networks can be used, such as convolutional neural networks (CNNs), which include many convolutional layers and a set of pooling layers, and have been shown to be beneficial for applications such as image recognition. CNNs are also easier to train than other networks due to the relatively small number of parameters to be determined.

[0103] In at least one embodiment, various tuning parameters can be used to train such complex machine learning models. Selecting parameters, fitting the model, and evaluating this model are part of the model tuning process, often referred to as hyperparameter optimization. In at least one embodiment, such tuning can include introspection of the underlying model or data. In a training or production setting, a robust workflow is very important to avoid overfitting of hyperparameters, as described elsewhere herein. Cross-validation and adding Gaussian noise to the training data set are useful techniques to avoid overfitting any one data set. For hyperparameter optimization, in at least one embodiment, it may be necessary to keep the training set and validation set fixed. In at least one embodiment, hyperparameters can be adjusted in certain categories, which may include, for example, data preprocessing, CNN architecture definition (e.g., filter size, number of filters), stochastic gradient descent (SGD) parameters (e.g., learning rate), regularization (e.g., drop probability).

[0104] In the example preprocessing step, the instances of the data set can be embedded into a lower dimensional space of a specific size. The size of this space is a parameter to be adjusted. The architecture of this CNN contains many adjustable parameters. The parameters of the filter size can represent the interpretation of information corresponding to the size of the instance to be analyzed. In computational linguistics, this is called n-gram size. The example CNN uses three different filter sizes, which represent possibly different n-gram sizes. The number of filters for each filter size can correspond to the depth of this filter. Each filter attempts to learn something different from the instance structure, such as the sentence structure of text data. In the convolutional layer, the activation function can be a rectified linear unit, and the pool type is set to a maximum pool. The results can then be concatenated into a one-dimensional vector, and the last layer is fully connected to the two-dimensional output. This corresponds to a binary classification to which an optimization function can be applied. One such function is an implementation of the root mean square (RMS) propagation method of gradient descent, where example hyperparameters can include learning rate, batch size, maximum gradient normal, and epoch. Neural network, regularization may be a very important consideration. As described, in at least one embodiment, the input data can be relatively sparse. In this case, the main hyperparameters can be dropped at the penultimate layer, which means that a certain proportion of nodes will not "fire" at each training cycle. The example training process can suggest different hyperparameter configurations based on feedback on the performance of previous configurations. The model can be trained using the suggested configuration, can be evaluated on a specified validation set, and performance can be reported. This process can be repeated, such as to balance exploration (learn more about different configurations) and development (leverage previous knowledge to get better results).

[0105] Since training CNNs can be parallelized and can take advantage of GPU-powered computing resources, multiple optimization strategies can be tried for different scenarios. Complex scenarios allow tuning of the model architecture and preprocessing and stochastic gradient descent parameters. This expands the model configuration space. In the base case, only preprocessing and stochastic gradient descent parameters are tuned. There can be many more configuration parameters in this complex scheme than in the base scheme. The tuning of the joint space can be performed with linear or exponential steps and iterated through the optimization loop of the model. Such a tuning process can be significantly less expensive than tuning processes such as random search and grid search without any noticeable performance loss.

[0106] In at least one embodiment, back propagation can be used to calculate the gradient for determining the weight of the neural network. Back propagation is a form of differentiation, and as described above, a gradient descent optimization algorithm can be used to adjust the weights applied to various nodes or neurons. In at least one embodiment, the gradient of the relevant loss function can be used to determine the weight. Back propagation can utilize the derivative of the output generated by the loss function to the statistical model. As described above, each node can have an associated activation function that defines the output of each node. Various activation functions can be used appropriately, such as radial basis functions (RBF) and sigmoid functions, which can be used for data conversion by various support vector machines (SVM). The activation function of the middle layer of the node is referred to as the inner product core in this article. These functions can include, for example, recognition functions, step functions, sigmoid functions, ramp functions, etc. The activation function can also be linear or nonlinear.

[0107] Reasoning and training logic

[0108] Fig.10 Inference and / or training logic 1015 is shown for performing reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Provide details about the inference and / or training logic 1015.

[0109] In at least one embodiment, the inference and / or training logic 1015 may include, but is not limited to, a data store 1001 to store forward and / or output weights and / or input / output data corresponding to neurons or layers of a neural network trained and / or used for inference in aspects of one or more embodiments. In at least one embodiment, the data store 1001 stores input / output data during training and / or inference using aspects of one or more embodiments and / or weight parameters and / or input / output data for each layer of a neural network trained or used in conjunction with one or more embodiments during forward propagation of weight parameters. In at least one embodiment, any portion of the data store 1001 may be included in other on-chip or off-chip data stores, including the processor's L1, L2, or L3 cache or system memory.

[0110] In at least one embodiment, any portion of data storage 1001 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, data storage 1001 may be 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 data storage 1001 is internal or external to a processor, for example, or composed of DRAM, SRAM, flash memory, or some other storage type, may depend on the available storage space 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.

[0111] In at least one embodiment, the inference and / or training logic 1015 may include, but is not limited to, a data store 1005 to store the reverse and / or output weights and / or input / output data corresponding to the neurons or layers of the neural network trained and / or used for inference in aspects of one or more embodiments. In at least one embodiment, during training and / or inference using aspects of one or more embodiments, the data store 1005 stores the weight parameters and / or input / output data for each layer of the neural network trained or used in conjunction with one or more embodiments during the back propagation of the input / output data and / or weight parameters. In at least one embodiment, any portion of the data store 1005 may be included with other on-chip or off-chip data stores, including the L1, L2, or L3 cache or system memory of a processor. In at least one embodiment, any portion of the data store 1005 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, the data store 1005 may be a cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice of whether data storage 1005 is internal or external to the processor, e.g., consisting of DRAM, SRAM, flash memory, or some other storage type, 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 sizes used in inference and / or training of the neural network, or some combination of these factors.

[0112] In at least one embodiment, data store 1001 and data store 1005 may be separate storage structures. In at least one embodiment, data store 1001 and data store 1005 may be the same storage structure. In at least one embodiment, data store 1001 and data store 1005 may be partially the same storage structure and partially separate storage structures. In at least one embodiment, any portion of data store 1001 and data store 1005 may be included with other on-chip or off-chip data stores, including the processor's L1, L2, or L3 cache or system memory.

[0113] In at least one embodiment, the inference and / or training logic 1015 may include, but is not limited to, one or more arithmetic logic units ("ALUs") 1010 for performing logic and / or mathematical operations based at least in part on or directed by training and / or inference code, the results of which may result in activations stored in activation storage 1020 (e.g., output values ​​from a layer or neuron within a neural network), which are functions of input / output and / or weight parameter data stored in data store 1001 and / or data store 1005. In at least one embodiment, the activations stored in activation storage 1020 are generated by linear algebra and / or matrix-based mathematics performed by ALU 1010 in response to executing instructions or other code, wherein weight values ​​stored in data store 1005 and / or data store 1001 are used as operands with other values, such as bias values, gradient information, momentum values, or other parameters or hyperparameters, any or all of which may be stored in data store 1005 and / or data store 1001 or other on-chip or off-chip storage. In at least one embodiment, one or more ALUs 1010 are included in one or more processors or other hardware logic devices or circuits, while in another embodiment, one or more ALUs 1010 may be outside a processor or other hardware logic device or circuit that uses them (e.g., a coprocessor). In at least one embodiment, one or more ALUs 1010 may be included within an execution unit of a processor, or otherwise included in a group of ALUs accessible by an execution unit of a processor, which may be within the same processor or distributed between 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, data storage 1001, data storage 1005, and activation storage 1020 may be on the same processor or other hardware logic device or circuit, while in another embodiment, they may be in different processors or other hardware logic devices or circuits or some combination of the same and different processors or other hardware logic devices or circuits. In at least one embodiment, any portion of activation storage 1020 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. Additionally, inference and / or training code may be stored with other code accessible to a 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.

[0114] In at least one embodiment, activation storage 1020 can be cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, activation storage 1020 can be completely or partially internal or external to one or more processors or other logic circuits. In at least one embodiment, activation storage 1020 can be selected to be internal or external to a processor, for example, or include DRAM, SRAM, flash memory, or other storage types, depending on the storage available on or off chip, the latency requirements for performing training and / or inference functions, the batch size of data used in inferencing and / or training neural networks, or some combination of these factors. In at least one embodiment, Fig.10 The inference and / or training logic 1015 shown in FIG. 10 may be used in conjunction with an application specific integrated circuit (“ASIC”), such as the ASIC from Google. Processing unit from Graphcore TM Inference Processing Unit (IPU) from Intel Corp (e.g., "Lake Crest") processor. In at least one embodiment, Fig.10 The illustrated inference and / or training logic 1015 may be used in conjunction with central processing unit (“CPU”) hardware, graphics processing unit (“GPU”) hardware, or other hardware such as a field programmable gate array (“FPGA”).

[0115] Fig.11 Inference and / or training logic 1015 in at least one various embodiment is shown. In at least one embodiment, the reasoning and / or training logic 1015 may include, but is not limited to, hardware logic in which computing 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, Fig.11 The inference and / or training logic 1015 shown in FIG. 1 may be used in conjunction with an application specific integrated circuit (ASIC), such as the ASIC from Google. Processing unit from Graphcore TM Inference Processing Unit (IPU) from Intel Corp (e.g., "Lake Crest") processor. In at least one embodiment, Fig.11The reasoning and / or training logic 1015 shown in 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 reasoning and / or training logic 1015 includes but is not limited to data storage 1001 and data storage 1005, which can be used to store weight values ​​and / or other information, including bias values, gradient information, momentum values, and / or other parameter or hyperparameter information. Fig.11 In at least one embodiment shown in , each of the data store 1001 and the data store 1005 is associated with a dedicated computing resource (e.g., computing hardware 1002 and computing hardware 1006). In at least one embodiment, each of the computing hardware 1002 and the computing hardware 1006 includes one or more ALUs that only perform mathematical functions (e.g., linear algebraic functions) on the information stored in the data store 1001 and the data store 1005, respectively, and the results of the execution of the functions are stored in the activation storage 1020.

[0116] In at least one embodiment, each of the data stores 1001 and 1005 and the corresponding computing hardware 1002 and 1006 corresponds to a different layer of the neural network, such that activations obtained from one "storage / compute pair 1001 / 1002" of the data store 1001 and computing hardware 1002 are provided as inputs to the next "storage / compute pair 1005 / 1006" of the data store 1005 and computing hardware 1006, so as to reflect the conceptual organization of the neural network. In at least one embodiment, each storage / compute pair 1001 / 1002 and 1005 / 1006 can correspond to more than one neural network layer. In at least one embodiment, additional storage / compute pairs (not shown) can be included in the inference and / or training logic 1015 after or in parallel with the storage / compute pairs 1001 / 1002 and 1005 / 1006.

[0117] Neural network training and deployment

[0118] Fig.12The training and deployment of a deep neural network in at least one embodiment is shown. In at least one embodiment, an untrained neural network 1206 is trained using a training data set 1202. In at least one embodiment, the training framework 1104 is a PyTorch framework, while in other embodiments, the training framework 1104 is Tensorflow, Boost, Caffe, Microsoft Cognitive Toolkit / CNTK, MXNet, Chainer, Keras, Deeplearning4j, or other training frameworks. In at least one embodiment, the training framework 1104 trains an untrained neural network 1106 and enables it to be trained using the processing resources described herein to generate a trained neural network 1108. In at least one embodiment, 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.

[0119] In at least one embodiment, untrained neural network 1106 is trained using supervised learning, where training data set 1102 includes inputs paired with expected outputs for the inputs, or where training data set 1102 includes inputs with known outputs and outputs for which the neural network is manually graded. In at least one embodiment, untrained neural network 1106 is trained in a supervised manner to process inputs from training data set 1102 and compare the resulting outputs to a set of expected or desired outputs. In at least one embodiment, errors are then propagated back through untrained neural network 1106. In at least one embodiment, training framework 1104 adjusts the weights that control untrained neural network 1106. In at least one embodiment, training framework 1104 includes tools for monitoring the degree to which untrained neural network 1106 converges to a model (e.g., trained neural network 1108), a model suitable for generating correct answers (e.g., results 1114) based on known input data (e.g., new data 1112). In at least one embodiment, the training framework 1104 repeatedly trains the untrained neural network 1106 while adjusting the weights to improve the output of the untrained neural network 1106 using a loss function and an adjustment algorithm (e.g., stochastic gradient descent). In at least one embodiment, the training framework 1104 trains the untrained neural network 1106 until the untrained neural network 1106 reaches a desired accuracy. In at least one embodiment, the trained neural network 1108 can then be deployed to implement any number of machine learning operations.

[0120] In at least one embodiment, untrained neural network 1106 is trained using unsupervised learning, where untrained neural network 1106 attempts to train itself using unlabeled data. In at least one embodiment, unsupervised learning training data set 1102 will include input data without any associated output data or "ground truth" data. In at least one embodiment, untrained neural network 1106 can learn groupings within training data set 1102 and can determine how individual inputs are related to untrained data set 1102. In at least one embodiment, unsupervised training can be used to generate a self-organizing map, which is a type of trained neural network 1108 that is capable of performing operations useful for reducing the dimensionality of new data 1112. In at least one embodiment, unsupervised training can also be used to perform anomaly detection, which allows identification of data points in new data set 1112 that deviate from the normal pattern of new data set 1112.

[0121] In at least one embodiment, semi-supervised learning can be used, which is a technique in which a mixture of labeled and unlabeled data is included in the training data set 1102. In at least one embodiment, the training framework 1104 can be used to perform incremental learning, such as through a transfer learning technique. In at least one embodiment, incremental learning enables the trained neural network 1108 to adapt to new data 1112 without forgetting the knowledge that was infused into the network during initial training.

[0122] Data Center

[0123] Fig.13 An example data center 1300 is shown in which at least one embodiment may be used. In at least one embodiment, data center 1300 includes a data center infrastructure layer 1310 , a framework layer 1320 , a software layer 1330 , and an application layer 1340 .

[0124] In at least one embodiment, Fig.13As shown, the data center infrastructure layer 1310 may include a resource coordinator 1312, grouped computing resources 1314, and node computing resources ("node CRs") 1316(1)-1316(N), where "N" represents any complete positive integer. In at least one embodiment, the node CRs 1316(1)-1316(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 devices (e.g., dynamic read-only memories), storage devices (e.g., 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 the node CRs 1316(1)-1316(N) may be a server having one or more of the above computing resources.

[0125] In at least one embodiment, the grouped computing resources 1314 may include separate groups of node CRs housed in one or more racks (not shown), or many racks (also not shown) housed in data centers at various geographic locations. The separate groups of node CRs within the grouped computing resources 1314 may include computing, networks, memory, or storage resources that can be configured or allocated to support groupings of one or more workloads. In at least one embodiment, several node CRs including a CPU or processor may be grouped in one or more racks to provide computing resources to support one or more workloads. In at least one embodiment, one or more racks may also include any number of power modules, cooling modules, and network switches in any combination.

[0126] In at least one embodiment, resource coordinator 1322 may configure or otherwise control one or more nodes CR 1316(1)-1316(N) and / or grouped computing resources 1314. In at least one embodiment, resource coordinator 1322 may include a software design infrastructure ("SDI") management entity for data center 1300. In at least one embodiment, resource coordinator 1322 may include hardware, software, or some combination thereof.

[0127] In at least one embodiment, Fig.13As shown, the framework layer 1320 includes a job scheduler 1332, a configuration manager 1334, a resource manager 1336, and a distributed file system 1338. In at least one embodiment, the framework layer 1320 may include a framework that supports the software 1332 of the software layer 1330 and / or one or more applications 1342 of the application layer 1340. In at least one embodiment, the software 1332 or the application 1342 may include a web-based service software or application, such as a service or application provided by Amazon Web Services, Google Cloud, and Microsoft Azure. In at least one embodiment, the framework layer 1320 may be, but is not limited to, a free and open source software network application framework, such as Apache SparkTM (hereinafter referred to as "Spark") that can utilize the distributed file system 1338 for large-scale data processing (e.g., "big data"). In at least one embodiment, the job scheduler 1332 may include a Spark driver to facilitate scheduling of workloads supported by various layers of the data center 1300. In at least one embodiment, the configuration manager 1334 may be capable of configuring different layers, such as the software layer 1330 and the framework layer 1320 including Spark and a distributed file system 1338 for supporting large-scale data processing. In at least one embodiment, the resource manager 1336 may be capable of managing cluster or group computing resources mapped to or allocated to support the distributed file system 1338 and the job scheduler 1332. In at least one embodiment, the cluster or group computing resources may include grouped computing resources 1314 on the data center infrastructure layer 1310. In at least one embodiment, the resource manager 1336 may coordinate with the resource coordinator 1312 to manage these mapped or allocated computing resources.

[0128] In at least one embodiment, software 1332 included in software layer 1330 may include software used by at least a portion of node CRs 1316(1)-1316(N), grouped computing resources 1314, and / or distributed file system 1338 of framework layer 1320. 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.

[0129] In at least one embodiment, one or more applications 1342 included in application layer 1340 may include one or more types of applications used by at least a portion of node CRs 1316(1)-1316(N), grouped computing resources 1314, and / or distributed file system 1338 of framework layer 1320. The one or more types of applications may include, but are not limited to, any number of genomics applications, cognitive computing, 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.

[0130] In at least one embodiment, any of configuration manager 1334, resource manager 1336, and resource coordinator 1312 can implement any number and type of self-modification actions based on any number and type of data acquired in any technically feasible manner. In at least one embodiment, the self-modification actions can relieve a data center operator of data center 1300 from making potentially bad configuration decisions and can avoid underutilized and / or poorly performing portions of the data center.

[0131] In at least one embodiment, data center 1300 may include tools, services, software, or other resources to train one or more machine learning models or 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 may be trained by calculating weight parameters according to a neural network architecture using the software and computing resources described above with respect to data center 1300. In at least one embodiment, by using weight parameters calculated by one or more training techniques described herein, information may 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 1300.

[0132] In at least one embodiment, the data center can use a CPU, an application-specific integrated circuit (ASIC), a GPU, an FPGA, or other hardware to use the above resources to perform training and / or reasoning. In addition, one or more of the above software and / or hardware resources can be configured as a service to allow users to train or perform information reasoning, such as image recognition, speech recognition, or other artificial intelligence services.

[0133] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Provides details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 1015 may be in the system Fig.13 In some embodiments, the present invention provides a method for performing inference or prediction 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.

[0134] In accordance with at least one embodiment, data center infrastructure 1310 can receive input text and cause the input to be directed to corresponding components of application layer 1340 and software layer 1330 for purposes of training and / or inference as described herein.

[0135] Computer Systems

[0136] Fig.14 1400, which may be a system of interconnected devices and components, a system on a chip (SOC), or some combination thereof, formed with a processor 1400, which may include an execution unit to execute instructions. In at least one embodiment, in accordance with the present disclosure, such as the embodiments described herein, the computer system 1400 may include, but is not limited to, components, such as a processor 1402, whose execution unit includes logic to execute algorithms for processing data. In at least one embodiment, the computer system 1400 may include a processor, such as a processor available from Intel Corporation of Santa Clara, California. Processor family, XeonTM, XScaleTM and / or StrongARMTM, Core TM or Nervana TM microprocessor, although other systems (including PCs with other microprocessors, engineering workstations, set-top boxes, etc.) may also be used. In at least one embodiment, computer system 1400 may execute a version of the WINDOWS operating system available from Microsoft Corporation of Redmond, Wash., although other operating systems (e.g., UNIX and Linux), embedded software, and / or graphical user interfaces may also be used.

[0137] Embodiments may be used in other devices, such as handheld devices and embedded applications. Some examples of handheld devices include cellular phones, Internet Protocol (Internet Protocol) 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, 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.

[0138] In at least one embodiment, the computer system 1400 may include, but is not limited to, a processor 1402, which may include, but is not limited to, one or more execution units 1408 to perform machine learning model training and / or reasoning according to the techniques described herein. In at least one embodiment, the system is a single-processor desktop or server system, but in another embodiment, the system may be a multi-processor system. In at least one embodiment, the processor 1402 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 that implements an instruction set combination, or any other processor device, such as a digital signal processor. In at least one embodiment, the processor 1402 may be coupled to a processor bus 1410, which may transmit data signals between the processor 1402 and other components in the computer system 1400.

[0139] In at least one embodiment, processor 1402 may include, but is not limited to, level 1 ("L1") internal cache memory ("cache") 1404. In at least one embodiment, processor 1402 may have a single internal cache or multiple levels of internal cache. In at least one embodiment, cache memory may reside external to processor 1402. Other embodiments may also include a combination of internal and external caches, depending on the particular implementation and needs. In at least one embodiment, register file 1406 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.

[0140] In at least one embodiment, an execution unit 1408, including but not limited to logic to perform integer and floating point operations, is also located in the processor 1402. The processor 1402 may also include a microcode ("ucode") read-only memory ("ROM") for storing microcode for certain macroinstructions. In at least one embodiment, the execution unit 1408 may include logic for processing a packed instruction set 1409. In at least one embodiment, by including the packed instruction set 1409 in the instruction set of the general purpose processor 1402, and the associated circuitry to execute the instructions, operations used by many multimedia applications may be performed using packed data in the general purpose processor 1402. In one or more embodiments, many multimedia applications may be executed faster and more efficiently by using the full width of the processor's data bus to perform operations on packed data, which may not require the transfer of smaller units of data on the processor's data bus to perform one or more operations one data element at a time.

[0141] In at least one embodiment, execution unit 1408 may also be used in microcontrollers, embedded processors, graphics devices, DSPs, and other types of logic circuits. In at least one embodiment, computer system 1400 may include, but is not limited to, memory 1420. In at least one embodiment, memory 1420 may be implemented as a dynamic random access memory ("DRAM") device, a static random access memory ("SRAM") device, a flash memory device, or other storage device. Memory 1420 may store instructions 1419 and / or data 1421 represented by data signals that may be executed by processor 1402.

[0142] In at least one embodiment, the system logic chip may be coupled to the processor bus 1410 and the memory 1420. In at least one embodiment, the system logic chip may include, but is not limited to, a memory controller hub ("MCH") 1416, and the processor 1402 may communicate with the MCH 1416 via the processor bus 1410. In at least one embodiment, the MCH 1416 may provide a high bandwidth memory path 1418 to the memory 1420 for instruction and data storage and for storage of graphics commands, data, and textures. In at least one embodiment, the MCH 1416 may initiate data signals between the processor 1402, the memory 1420, and other components in the computer system 1400, and bridge data signals between the processor bus 1410, the memory 1420, and the system I / O 1422. 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, the MCH 1416 may be coupled to the memory 1420 via a high bandwidth memory path 1418 , and the graphics / video card 1412 may be coupled to the MCH 1416 via an Accelerated Graphics Port (“AGP”) interconnect 1414 .

[0143] In at least one embodiment, computer system 1400 may use system I / O 1422 as a proprietary hub interface bus to couple MCH 1416 to I / O controller hub ("ICH") 1430. In at least one embodiment, ICH 1430 may provide direct connection to certain I / O devices through 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 used to connect peripheral devices to memory 1420, chipsets, and processor 1402. Examples may include, but are not limited to, an audio controller 1429, a firmware hub ("Flash BIOS") 1428, a wireless transceiver 1426, a data store 1424, a traditional I / O controller 1423 including user input and keyboard interface, a serial expansion port 1427 (e.g., a universal serial bus (USB)), and a network controller 1434. Data store 1424 may include a hard drive, a floppy drive, a CD-ROM device, a flash memory device, or other mass storage device.

[0144] In at least one embodiment, Fig.14 The system is shown as comprising interconnected hardware devices or "chips", while in other embodiments, Fig.14 An exemplary system on a chip ("SoC") may be shown. In at least one embodiment, Fig.14The devices shown in can be interconnected with a proprietary interconnect, a standardized interconnect (eg, PCIe), or some combination thereof. In at least one embodiment, one or more components of system 1400 are interconnected using a compute express link (CXL) interconnect.

[0145] Reasoning and / or training logic 1015 is used to perform reasoning and / or training operations related to one or more embodiments. Fig.10 and / or Fig.11 Provide details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 1015 may be Fig.14 for use in a system 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.

[0146] In some embodiments, for example, a video data stream may be received via expansion port 1427 or wireless transceiver 1426 and then directed to processor 1402 and / or video graphics card 1412 for processing. Depending on whether these components are part of a device (such as an autonomous vehicle) or a separate device, the output may then go to a control system via I / O or be transmitted to the vehicle via a wireless transceiver.

[0147] Fig.15 1 is a block diagram illustrating an electronic device 1500 for utilizing a processor 1510 in at least one embodiment. In at least one embodiment, the electronic device 1500 may be, for example but not limited to, a notebook computer, a tower server, a rack server, a blade server, a laptop computer, a desktop computer, a tablet computer, a mobile device, a phone, an embedded computer, or any other suitable electronic device.

[0148] In at least one embodiment, system 1500 may include, but is not limited to, a processor 1510 communicatively coupled to any suitable number or kind of components, peripherals, modules, or devices. In at least one embodiment, processor 1510 is coupled using a bus or interface, such as an I2C bus, a system management bus ("SMBus"), a low pin count (LPC) bus, a serial peripheral interface ("SPI"), a high-definition audio ("HDA") bus, a serial advanced technology attachment ("SATA") bus, a universal serial bus ("USB") (versions 1, 2, 3), or a universal asynchronous receiver / transmitter ("UART") bus. In at least one embodiment, Fig.15 A system is shown that includes interconnected hardware devices or "chips", while in other embodiments, Fig.15 An exemplary system on a chip ("SoC") may be shown.

[0149] In at least one embodiment, Fig.15 The devices shown in can be interconnected with a proprietary interconnect, a standardized interconnect (e.g., PCIe), or some combination thereof. In at least one embodiment, Fig.15 One or more components of the system are interconnected using Compute Express Link (CXL) interconnect lines.

[0150] In at least one embodiment, Fig.15 It may include a display 1524, a touch screen 1525, a touch pad 1530, a near field communication unit (“NFC”) 1545, a sensor hub 1540, a thermal sensor 1546, a fast chipset (“EC”) 1535, a trusted platform module (“TPM”) 1538, a BIOS / firmware / flash (“BIOS, FW Flash”) 1522, a DSP 1560, a drive (“SSD” or “HDD”) 1520 (e.g., a solid state disk (“SSD”) or a hard disk drive (“HDD”)), a wireless local area network unit (“WLAN”) 1550, a Bluetooth unit 1552, a wireless wide area network unit (“WWAN”) 1556, a global positioning system (GPS) 1555, a camera (“USB 3.0 camera”) 1554 (e.g., a USB 3.0 camera), or a low power double data rate (“LPDDR”) memory unit (“LPDDR3”) 1515 implemented with, for example, the LPDDR3 standard. These components may each be implemented in any suitable manner.

[0151] In at least one embodiment, other components may be communicatively coupled to the processor 1510 through the components discussed above. In at least one embodiment, an accelerometer 1541, an ambient light sensor (“ALS”) 1542, a compass 1543, and a gyroscope 1544 may be communicatively coupled to the sensor hub 1540. In at least one embodiment, a thermal sensor 1539, a fan 1537, a keyboard 1546, and a touchpad 1530 may be communicatively coupled to the EC 1535. In at least one embodiment, a speaker 1563, an earphone 1564, and a microphone (“mic”) 1565 may be communicatively coupled to an audio unit (“audio codec and class D amplifier”) 1564, which in turn may be communicatively coupled to the DSP 1560. In at least one embodiment, the audio unit 1564 may include, for example, but not limited to, an audio encoder / decoder (“codec”) and a class D amplifier. In at least one embodiment, a SIM card (“SIM”) 1557 may be communicatively coupled to the WWAN unit 1556. In at least one embodiment, components such as the WLAN unit 1550 and the Bluetooth unit 1552 and the WWAN unit 1556 may be implemented as a next generation form factor (NGFF).

[0152] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Provides details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 1015 may be in the system Fig.15 for use in reasoning or predicting 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.

[0153] Fig.16 A computer system 1600 is shown in at least one embodiment. In at least one embodiment, the computer system 1600 is configured to implement the various processes and methods described throughout this disclosure.

[0154] In at least one embodiment, computer system 1600 includes, but is not limited to, at least one central processing unit ("CPU") 1602 connected to a communication bus 1610 implemented using any suitable protocol, such as PCI ("Peripheral Component 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, computer system 1600 includes, but is not limited to, main memory 1604 and control logic (e.g., implemented as hardware, software, or a combination thereof), and data may be stored in main memory 1604 in the form of random access memory ("RAM"). In at least one embodiment, a network interface subsystem ("network interface") 1622 provides an interface to other computing devices and networks for receiving data from computer system 1600 and transmitting data to other systems.

[0155] In at least one embodiment, computer system 1600 includes, but is not limited to, input device 1608, parallel processing system 1612, and display device 1606, which may be implemented using a cathode ray tube ("CRT"), a liquid crystal display ("LCD"), a light emitting diode ("LED"), a plasma display, or other suitable display technology. In at least one embodiment, user input is received from input device 1608 (such as a keyboard, mouse, touch pad, microphone, etc.). In at least one embodiment, each of the foregoing modules may be located on a single semiconductor platform to form a processing system.

[0156] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Provides details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 1015 may be in the system Fig.16 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 architectures, or neural network use cases described herein.

[0157] Fig.17 A computer system 1700 in at least one embodiment is shown. In at least one embodiment, the computer system 1700 includes, but is not limited to, a computer 1710 and a USB disk 1720. In at least one embodiment, the computer 1710 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 1710 includes, but is not limited to, a server, a cloud instance, a laptop computer, and a desktop computer.

[0158] In at least one embodiment, the USB disk 1720 includes, but is not limited to, a processing unit 1730, a USB interface 1740, and a USB interface logic 1750. In at least one embodiment, the processing unit 1730 may be any instruction execution system, device, or device capable of executing instructions. In at least one embodiment, the processing unit 1730 may include, but is not limited to, any number and type of processing cores (not shown). In at least one embodiment, the processing core 1730 includes an application specific integrated circuit ("ASIC") that is optimized to perform any number and type of operations associated with machine learning. For example, in at least one embodiment, the processing core 1730 is a tensor processing unit ("TPC") that is optimized to perform machine learning reasoning operations. In at least one embodiment, the processing core 1730 is a visual processing unit ("VPU") that is optimized to perform machine vision and machine learning reasoning operations.

[0159] In at least one embodiment, USB interface 1740 can be any type of USB connector or USB socket. For example, in at least one embodiment, USB interface 1740 is a USB 3.0 Type-C socket for data and power. In at least one embodiment, USB interface 1740 is a USB 3.0 Type-A connector. In at least one embodiment, USB interface logic 1750 can include any number and type of logic that enables processing unit 1730 to connect to a device (e.g., computer 1710) via USB connector 1740.

[0160] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Provides details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 1015 may be in the system Fig.17 In some embodiments, the present invention provides a method for performing inference or prediction 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.

[0161] Fig.18 1 is a block diagram illustrating an exemplary system on a chip integrated circuit 1800 that can be manufactured using one or more IP cores in at least one embodiment. In at least one embodiment, the integrated circuit 1800 includes one or more application processors 1805 (e.g., CPU), at least one graphics processor 1810, and may additionally include an image processor 1815 and / or a video processor 1820, any of which may be a modular IP core. In at least one embodiment, the integrated circuit 1800 includes peripheral or bus logic, which includes a USB controller 1825, a UART controller 1830, a SPI / SDIO controller 1835, and an I.sup.2S / I.sup.2C controller 1840. In at least one embodiment, the integrated circuit 1800 may include a display device 1845 coupled to one or more of a high-definition multimedia interface (HDMI) controller 1850 and a mobile industry processor interface (MIPI) display interface 1855. In at least one embodiment, storage may be provided by a flash subsystem 1860, including flash memory and a flash controller. In at least one embodiment, a memory interface may be provided via a memory controller 1865 for accessing SDRAM or SRAM memory devices. In at least one embodiment, some integrated circuits also include an embedded security engine 1870 .

[0162] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding inference and / or training logic 1015. In at least one embodiment, inference and / or training logic 1015 may be used in integrated circuit 1800 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.

[0163] For example, inference and / or training logic 1015 can accept an input video stream and generate inferences for objects represented in the video stream, as described herein.In at least some embodiments, image processor 1815 can be used to process video frames as they are received.

[0164] Figures 19A-19BAn exemplary integrated circuit and associated graphics processor according to various embodiments described herein are shown, which can be manufactured using one or more IP cores. In addition to the illustrations, other logic and circuits may be included in at least one embodiment, including additional graphics processors / cores, peripheral interface controllers, or general-purpose processor cores.

[0165] Figures 19A-19B is a block diagram illustrating an exemplary graphics processor for use within a SoC according to embodiments described herein. Fig.19A An exemplary graphics processor 1910 is shown that in at least one embodiment is a system on a chip integrated circuit, which may be manufactured using one or more IP cores. Fig.19B Another exemplary graphics processor 1940 of a system on a chip integrated circuit in at least one embodiment is shown, which can be manufactured using one or more IP cores. In at least one embodiment, Fig.19A The graphics processor 1910 is a low power graphics processor core. In at least one embodiment, Fig.19B The graphics processor 1940 is a higher performance graphics processor core. In at least one embodiment, each graphics processor 1910, 1940 may be Fig.18 A variant of the graphics processor 1810.

[0166] In at least one embodiment, the graphics processor 1910 includes a vertex processor 1905 and one or more fragment processors 1915A-1915N (e.g., 1915A, 1915B, 1915C, 1915D to 1915N-1 and 1915N). In at least one embodiment, the graphics processor 1910 can execute different shader programs via separate logic, so that the vertex processor 1905 is optimized to perform operations for the vertex shader program, while one or more fragment processors 1915A-1915N perform fragment (e.g., pixel) shading operations for fragments or pixels or shader programs. In at least one embodiment, the vertex processor 1905 performs the vertex processing stage of the 3D graphics pipeline and generates primitives and vertex data. In at least one embodiment, the one or more fragment processors 1915A-1915N use the primitives and vertex data generated by the vertex processor 1905 to generate a frame buffer displayed on a display device. In at least one embodiment, one or more fragment processors 1915A-1915N are optimized to execute fragment shader programs as provided in the OpenGL API, which can be used to perform similar operations as pixel shader programs provided in the Direct 3D API.

[0167] In at least one embodiment, graphics processor 1910 additionally includes one or more memory management units (MMUs) 1920A-1920B, one or more caches 1925A-1925B, and one or more circuit interconnects 1930A-1930B. In at least one embodiment, one or more MMUs 1920A-1920B provide a mapping of virtual to physical addresses for graphics processor 1910, including for vertex processors 1905 and / or fragment processors 1915A-1915N, which may reference vertex or image / texture data stored in memory in addition to vertex or image / texture data stored in one or more caches 1925A-1925B. In at least one embodiment, one or more MMUs 1920A-1920B provide a mapping of virtual to physical addresses for graphics processor 1910, including for vertex processors 1905 and / or fragment processors 1915A-1915N, which may reference vertex or image / texture data stored in memory in addition to vertex or image / texture data stored in one or more caches 1925A-1925B.

[0168] The 1920A-1920B can synchronize with other MMUs in the system, including Fig.18 One or more MMUs associated with one or more application processors 1805, image processor 1815, and / or video processor 1820 enable each processor 1805-1820 to participate in a shared or unified virtual memory system. In at least one embodiment, one or more circuit interconnects 1930A-1930B enable graphics processor 1910 to connect to other IP cores within the SoC via an internal bus of the SoC or via a direct connection.

[0169] In at least one embodiment, graphics processor 1940 includes Fig.19A One or more MMUs 1920A-1920B, caches 1925A-1925B, and circuit interconnects 1930A-1930B of a graphics processor 1910. In at least one embodiment, a graphics processor 1940 includes one or more shader cores 1955A-1955N (e.g., 1955A, 1955B, 1955C, 1955D, 1955E, 1955F to 1955N-1 and 1955N) that provide a unified shader core architecture in which 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 plurality of shader cores may vary. In at least one embodiment, the graphics processor 1940 includes an inter-core task manager 1945 that acts as a thread dispatcher to dispatch execution threads to one or more shader cores 1955A-1955N and a tiling unit 1958 to accelerate tile operations for tile-based rendering, in which rendering operations of a scene are subdivided in image space, for example, to exploit local spatial coherence within a scene or to optimize use of internal caches.

[0170] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Provides details about the inference and / or training logic 1015. In at least one embodiment, the inference and / or training logic 1015 may be implemented in an integrated circuit. Fig.19A and / or Fig.19B Inference and / or training logic 1015 may be used to perform inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions or architectures, or neural network use cases described herein. For example, inference and / or training logic 1015 may accept an input video stream and generate inferences for objects represented in the video stream, as discussed herein.

[0171] Figures 20A-20B Additional exemplary graphics processor logic according to the embodiments described herein is shown. In at least one embodiment, Fig. 20A shows that it can be included in Fig.18 The graphics core 2000 within the graphics processor 1810 may, in at least one embodiment, be Fig.19B Unified shader cores 1955A-1955N. Fig. 20B A highly parallel general purpose graphics processing unit 2030 suitable for deployment on a multi-chip module in at least one embodiment is shown.

[0172] In one embodiment, graphics core 2000 includes a shared instruction cache 2002, texture unit 2018, and cache / shared memory 2020, which may be shared with execution resources within graphics core 2000. In at least one embodiment, graphics core 2000 may include multiple slices 2001A-2001N or partitions of each core, and a graphics processor may include multiple instances of graphics core 2000. Slices 2001A-2001N may include support logic including local instruction caches 2004A-2004N, thread schedulers 2006A-2006N, thread dispatchers 2008A-2008N, and a set of registers 2010A-2010N. In at least one embodiment, the slices 2001A-2001N may include a set of additional function units (AFU 2012A-2012N), floating point units (FPU 2014A-2014N), integer arithmetic logic units (ALU 2016A-2016N), address calculation units (ACU

[0173] 2013A-2013N), Double Precision Floating Point Unit (DPFPU 2015A-2015N), and Matrix Processing Unit (MPU2017A-2017N).

[0174] In one embodiment, the FPU 2014A-2014N can perform single-precision (32-bit) and half-precision (16-bit) floating-point operations, while the DPFPU 2015A-2015N can perform double-precision (64-bit) floating-point operations. In one embodiment, the ALU 2016A-2016N can perform variable-precision integer operations with 8-bit, 16-bit, and 32-bit precision, and can be configured for mixed-precision operations. In one embodiment, the MPU 2017A-2017N can also be configured for mixed-precision matrix operations, including half-precision floating-point operations and 8-bit integer operations. In one embodiment, the MPU 2017A-2017N can perform various matrix operations to accelerate machine learning application frameworks, including general matrix-to-matrix multiplication (GEMM) that enables support for acceleration. In one embodiment, the AFU 2012A-2012N can perform additional logical operations that are not supported by floating-point or integer units, including trigonometric operations (e.g., Sine, Cosine, etc.).

[0175] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding inference and / or training logic 1015. In at least one embodiment, inference and / or training logic 1015 may be used in graphics core 2000 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.

[0176] Fig. 20BA general purpose processing unit (GPGPU) 2030 in at least one embodiment is shown, which can be configured to enable highly parallel computing operations to be performed by a graphics processing unit array. In at least one embodiment, GPGPU 2030 can be directly linked to other instances of GPGPU 2030 to create a multi-GPU cluster to increase the training speed for deep neural networks. In at least one embodiment, GPGPU 2030 includes a host interface 2032 to enable connection with a host processor. In at least one embodiment, host interface 2032 is a PCI Express interface. In at least one embodiment, host interface 2032 can be a manufacturer-specific communication interface or communication structure. In at least one embodiment, GPGPU 2030 receives commands from a host processor and uses a global scheduler 2034 to dispatch execution threads associated with those commands to a group of computing clusters 2036A-2036H. In at least one embodiment, computing clusters 2036A-2036H share cache memory 2038. In at least one embodiment, cache memory 2038 may be used as a higher level cache for cache memory within computing clusters 2036A-2036H.

[0177] In at least one embodiment, GPGPU 2030 includes memory 2044A-2044B coupled to compute cluster 2036A-2036H via a set of memory controllers 2042A-2042B. In at least one embodiment, memory 2044A-2044B 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.

[0178] In at least one embodiment, computing clusters 2036A-2036H each include a set of graphics cores, such as Fig. 20A The graphics core 2000 may include multiple types of integer and floating point logic units that may be used to perform computational operations within a range of precision 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 2036A-2036H may be configured to perform 16-bit or 32-bit floating point operations, while a different subset of the floating point units may be configured to perform 64-bit floating point operations.

[0179] In at least one embodiment, multiple instances of GPGPU 2030 can be configured to operate as a computing cluster. In at least one embodiment, the communications used by computing clusters 2036A-2036H for synchronization and data exchange vary between embodiments. In at least one embodiment, multiple instances of GPGPU 2030 communicate through host interface 2032. In at least one embodiment, GPGPU 2030 includes an I / O hub 2039 that couples GPGPU 2030 with GPU link 2040, enabling direct connection to other instances of GPGPU 2030. In at least one embodiment, GPU link 2040 is coupled to a dedicated GPU to GPU bridge that enables communication and synchronization between multiple instances of GPGPU 2030. In at least one embodiment, GPU link 2040 is coupled to a high-speed interconnect to send and receive data to other GPGPUs or parallel processors. In at least one embodiment, multiple instances of GPGPU 2030 are located in separate data processing systems and communicate via network devices accessible via host interface 2032. In at least one embodiment, GPU link 2040 may be configured to enable connection to a host processor, in addition to or in place of host interface 2032 .

[0180] In at least one embodiment, the GPGPU 2030 may be configured to train a neural network. In at least one embodiment, the GPGPU 2030 may be used within an inference platform. In at least one embodiment in which the GPGPU 2030 is used for inference, the GPGPU may include fewer computing clusters 2036A-2036H relative to when the GPGPU is used to train a neural network. In at least one embodiment, the memory technology associated with the memory 2044A-2044B may differ between the inference and training configurations, with higher bandwidth memory technology being dedicated to the training configuration. In at least one embodiment, the inference configuration of the GPGPU 2030 may support inference specific instructions. For example, in at least one embodiment, the inference configuration may provide support for one or more 8-bit integer dot product instructions that may be used during inference operations of a deployed neural network.

[0181] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding inference and / or training logic 1015. In at least one embodiment, inference and / or training logic 1015 may be used in GPGPU 2030 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 as described herein.

[0182] Fig.21 2100. In at least one embodiment, the computing system 2100 includes a processing subsystem 2101 having one or more processors 2102 and a system memory 2104 communicating via an interconnect path that may include a memory hub 2105. In at least one embodiment, the memory hub 2105 may be a separate component within a chipset component or may be integrated within one or more processors 2102. In at least one embodiment, the memory hub 2105 is coupled to an I / O subsystem 2111 via a communication link 2106. In an embodiment, the I / O subsystem 2111 includes an I / O hub 2107 that can enable the computing system 2100 to receive input from one or more input devices 2108. In at least one embodiment, the I / O hub 2107 can enable a display controller, which is included in one or more processors 2102, to provide output to one or more display devices 2110A. In at least one embodiment, the one or more display devices 2110A coupled to the I / O hub 2107 may include local, internal, or embedded display devices.

[0183] In at least one embodiment, the processing subsystem 2101 includes one or more parallel processors 2112 coupled to the memory hub 2105 via a bus or other communication link 2113. In at least one embodiment, the communication link 2113 can be one of many standard-based communication link technologies or protocols, such as but not limited to PCI Express, or can be a communication interface or communication structure for a vendor. In at least one embodiment, the one or more parallel processors 2112 form a parallel or vector processing system in a computational concentration, which can include a large number of processing cores and / or processing clusters, such as a multi-integrated core (MIC) processor. In at least one embodiment, the one or more parallel processors 2112 form a graphics processing subsystem that can output pixels to one of the one or more display devices 2110A coupled via the I / O hub 2107. In at least one embodiment, the one or more parallel processors 2112 can also include a display controller and a display interface (not shown) to enable direct connection to one or more display devices 2110B.

[0184] In at least one embodiment, a system storage unit 2114 can be connected to the I / O hub 2107 to provide a storage mechanism for the computing system 2100. In at least one embodiment, an I / O switch 2116 can be used to provide an interface mechanism to enable connections between the I / O hub 2107 and other components, such as a network adapter 2118 and / or a wireless network adapter 2119 that can be integrated into the platform, as well as various other devices that can be added through one or more additional devices 2120. In at least one embodiment, the network adapter 2118 can be an Ethernet adapter or another wired network adapter. In at least one embodiment, the wireless network adapter 2119 can include one or more of Wi-Fi, Bluetooth, near field communication (NFC), or other network devices including one or more radios.

[0185] In at least one embodiment, computing system 2100 may include other components not explicitly shown, including USB or other port connections, optical storage drives, video capture devices, etc., which may also be connected to I / O hub 2107. Fig.21 The communication paths interconnecting the various components in the system can be implemented using any suitable protocol, such as a PCI (Peripheral Component Interconnect) based protocol (e.g., PCI-Express), or other bus or point-to-point communication interfaces and / or protocols (e.g., NV-Link high-speed interconnect or interconnect protocol).

[0186] In at least one embodiment, one or more parallel processors 2112 include circuits optimized for graphics and video processing (including, for example, video output circuits) and constitute a graphics processing unit (GPU). In at least one embodiment, one or more parallel processors 2112 include circuits optimized for general processing. In at least one embodiment, the components of the computing system 2100 can be integrated with one or more other system elements on a single integrated circuit. For example, in at least one embodiment, one or more parallel processors 2112, memory hub 2105, one or more processors 2102, and I / O hub 2107 can be integrated into a system on chip (SoC) integrated circuit. In at least one embodiment, the components of the computing system 2100 can 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 computing system 2100 can be integrated into a multi-chip module (MCM), which can be interconnected with other multi-chip modules into a modular computing system.

[0187] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Provide details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 1015 may be Fig.21 for use in system 2100 for inferring 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.

[0188] processor

[0189] Fig. 22 2200 according to at least one embodiment. In at least one embodiment, the various components of the parallel processor 2200 may be implemented using one or more integrated circuit devices, such as a programmable processor, an application specific integrated circuit (ASIC), or a field programmable gate array (FPGA). In at least one embodiment, the parallel processor 2200 shown is a processor according to an exemplary embodiment. Fig.21 A variation of one or more parallel processors 2112 is shown.

[0190] In at least one embodiment, parallel processor 2200 includes parallel processing unit 2202. In at least one embodiment, parallel processing unit 2202 includes I / O unit 2204, which enables communication with other devices, including other instances of parallel processing unit 2202. In at least one embodiment, I / O unit 2204 can be directly connected to other devices. In at least one embodiment, I / O unit 2204 is connected to other devices by using a hub or switch interface (e.g., memory hub 2105). In at least one embodiment, the connection between memory hub 2105 and I / O unit 2204 forms communication link 2113. In at least one embodiment, I / O unit 2204 is connected to host interface 2206 and memory crossbar switch 2216, wherein host interface 2206 receives commands for performing processing operations and memory crossbar switch 2216 receives commands for performing memory operations.

[0191] In at least one embodiment, when the host interface 2206 receives the command buffer via the I / O unit 2204, the host interface 2206 can direct work operations to execute those commands to the front end 2208. In at least one embodiment, the front end 2208 is coupled with a scheduler 2210, which is configured to distribute commands or other work items to the processing cluster array 2212. In at least one embodiment, the scheduler 2210 ensures that the processing cluster array 2212 is properly configured and in a valid state before assigning tasks to the processing cluster array 2212 in the processing cluster array 2212. In at least one embodiment, the scheduler 2210 is implemented by firmware logic executed on a microcontroller. In at least one embodiment, the microcontroller-implemented scheduler 2210 can be configured to perform complex scheduling and work distribution operations at coarse and fine granularity, thereby achieving fast preemption and context switching of threads executed on the processing array 2212. In at least one embodiment, the host software can prove the workload for scheduling on the processing array 2212 through one of a plurality of graphics processing doorbells. In at least one embodiment, the workload may then be automatically distributed across the processing array 2212 by scheduler 2210 logic within a microcontroller that includes scheduler 2210 .

[0192] In at least one embodiment, the processing cluster array 2212 may include up to "N" processing clusters (e.g., cluster 2214A, cluster 2214B, through cluster 2214N). In at least one embodiment, each cluster 2214A-2214N of the processing cluster array 2212 may execute a large number of concurrent threads. In at least one embodiment, the scheduler 2210 may allocate work to the clusters 2214A-2214N of the processing cluster array 2212 using various scheduling and / or work allocation algorithms, which may vary depending on the workload generated by each program or type of calculation. In at least one embodiment, scheduling may be handled dynamically by the scheduler 2210, or may be partially assisted by compiler logic during the compilation of program logic configured to be executed by the processing cluster array 2212. In at least one embodiment, different clusters 2214A-2214N of the processing cluster array 2212 may be allocated to process different types of programs or to perform different types of calculations.

[0193] In at least one embodiment, processing cluster array 2212 can be configured to perform various types of parallel processing operations. In at least one embodiment, processing cluster array 2212 is configured to perform general-purpose parallel computing operations. For example, in at least one embodiment, processing cluster array 2212 can include logic to perform processing tasks including filtering video and / or audio data, performing modeling operations including physics operations, and performing data transformations.

[0194] In at least one embodiment, the processing cluster array 2212 is configured to perform parallel graphics processing operations. In at least one embodiment, the processing cluster array 2212 may include additional logic to support the execution of such graphics processing operations, including but not limited to texture sampling logic to perform texture operations, as well as tessellation logic and other vertex processing logic. In at least one embodiment, the processing cluster array 2212 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 2202 may transfer data from the system memory via the I / O unit 2204 for processing. In at least one embodiment, during processing, the transferred data may be stored to an on-chip memory (e.g., parallel processor memory 2222) during processing and then written back to the system memory.

[0195] In at least one embodiment, when the parallel processing unit 2202 is used to perform graphics processing, the scheduler 2210 can be configured to divide the processing workload into tasks of approximately equal size to better distribute graphics processing operations to multiple clusters 2214A-2214N of the processing cluster array 2212. In at least one embodiment, portions of the processing cluster array 2212 can be configured to perform different types of processing. For example, in at least one embodiment, a first portion can be configured to perform vertex shading and topology generation, a second portion can be configured to perform tessellation and geometry shading, and a third portion can 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 2214A-2214N can be stored in a buffer to allow the intermediate data to be transferred between the clusters 2214A-2214N for further processing.

[0196] In at least one embodiment, the processing cluster array 2212 can receive processing tasks to be executed via the scheduler 2210, which receives commands defining the processing tasks from the front end 2208. In at least one embodiment, the processing tasks can include an index of data to be processed, which can include surface (patch) data, primitive data, vertex data, and / or pixel data, as well as state parameters and commands defining how to process the data (e.g., what program to execute). In at least one embodiment, the scheduler 2210 can be configured to obtain an index corresponding to a task, or can receive an index from the front end 2208. In at least one embodiment, the front end 2208 can be configured to ensure that the processing cluster array 2212 is configured in a valid state before starting a workload specified by an incoming command buffer (e.g., a batch-buffer, a push buffer, etc.).

[0197] In at least one embodiment, each of the one or more instances of parallel processing unit 2202 can be coupled to parallel processor memory 2222. In at least one embodiment, parallel processor memory 2222 can be accessed via memory crossbar switch 2216, which can receive memory requests from processing cluster array 2212 and I / O unit 2204. In at least one embodiment, memory crossbar switch 2216 can access parallel processor memory 2222 via memory interface 2218. In at least one embodiment, memory interface 2218 can include multiple partition units (e.g., partition unit 2220A, partition unit 2220B to partition unit 2220N), which can each be coupled to a portion of parallel processor memory 2222 (e.g., memory units). In at least one embodiment, the plurality of partition units 2220A-2220N are configured to be equal to the number of memory cells, such that the first partition unit 2220A has a corresponding first memory cell 2224A, the second partition unit 2220B has a corresponding memory cell 2224B, and the Nth partition unit 2220N has a corresponding Nth memory cell 2224N. In at least one embodiment, the number of partition units 2220A-2220N may not be equal to the number of memory devices.

[0198] In at least one embodiment, the memory units 2224A-2224N 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, the memory units 2224A-2224N 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 can be stored across the memory units 2224A-2224N, allowing the partition units 2220A-2220N to write to portions of each rendering target in parallel to efficiently use the available bandwidth of the parallel processor memory 2222. In at least one embodiment, local instances of the parallel processor memory 2222 can be eliminated to facilitate a unified memory design that utilizes system memory in combination with local cache memory.

[0199] In at least one embodiment, any of the clusters 2214A-2214N of the processing cluster array 2212 can process data to be written to any memory unit 2224A-2224N within the parallel processor memory 2222. In at least one embodiment, the memory crossbar switch 2216 can be configured to transmit the output of each cluster 2214A-2214N to any partition unit 2220A-2220N or another cluster 2214A-2214N, and the cluster 2214A-2214N can perform other processing operations on the output. In at least one embodiment, each cluster 2214A-2214N can communicate with the memory interface 2218 through the memory crossbar switch 2216 to read from or write to various external storage devices. In at least one embodiment, memory crossbar switch 2216 has connections to memory interface 2218 to communicate with I / O unit 2204, and connections to local instances of parallel processor memory 2222, thereby enabling processing units within different processing clusters 2214A-2214N to communicate with system memory or other memory that is not local to parallel processing unit 2202. In at least one embodiment, memory crossbar switch 2216 can use virtual channels to separate traffic flows between clusters 2214A-2214N and partition units 2220A-2220N.

[0200] In at least one embodiment, multiple instances of parallel processing unit 2202 may be provided on a single plug-in card, or multiple plug-in cards may be interconnected. In at least one embodiment, different instances of parallel processing unit 2202 may be configured to interoperate, even if 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 parallel processing unit 2202 may include higher precision floating point units relative to other instances. In at least one embodiment, a system incorporating one or more instances of parallel processing unit 2202 or parallel processor 2200 may be implemented in a variety of configurations and form factors, including but not limited to desktop, laptop or handheld personal computers, servers, workstations, game consoles, and / or embedded systems.

[0201] Fig.23 is a block diagram of a partition unit 2320 in at least one embodiment. In at least one embodiment, the partition unit 2320 is Fig. 222320N). In at least one embodiment, partition unit 2320 includes L2 cache 2321, frame buffer interface 2325, and ROP 2326 (raster operation unit). L2 cache 2321 is a read / write cache that is configured to perform load and store operations received from memory crossbar switch 2316 and ROP 2326. In at least one embodiment, L2 cache 2321 outputs read misses and urgent write-back requests to frame buffer interface 2325 for processing. In at least one embodiment, updates can also be sent to the frame buffer via frame buffer interface 2325 for processing. In at least one embodiment, frame buffer interface 2325 communicates with memory units (such as ROPs) in parallel processor memory. Fig. 22 interacts with one of the memory units 2224A-2224N (e.g., within parallel processor memory 2222).

[0202] In at least one embodiment, ROP 2326 is a processing unit that performs raster operations such as stenciling, z-testing, blending, etc. In at least one embodiment, ROP 2326 then outputs processed graphics data stored in graphics memory. In at least one embodiment, ROP 2326 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 can be lossless compression logic that utilizes one or more of a variety of compression algorithms. The type of compression performed by ROP 2326 can 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.

[0203] In at least one embodiment, ROP 2326 is included within each processing cluster (e.g., Fig. 22 In at least one embodiment, read and write requests for pixel data are transmitted through memory crossbar switch 2316 rather than pixel fragment data transmission. In at least one embodiment, the processed graphics data can be displayed on a display device (such as Fig.21 2100), routed by processor 2102 for further processing, or by Fig. 22 One of the processing entities within parallel processor 2200 is routed for further processing.

[0204] Fig.24 is a block diagram of a processing cluster 2414 within a parallel processing unit in at least one embodiment. In at least one embodiment, a processing cluster is Fig. 22In at least one embodiment, processing cluster 2414 can be configured to execute many threads in parallel, where the term "thread" refers to an instance of a specific program executed on a specific set of input data. In at least one embodiment, single instruction multiple data (SIMD) instruction issuance technology is used to support the parallel execution of a large number of threads without providing multiple independent instruction units. In at least one embodiment, single instruction multiple thread (SIMT) technology is used to support the parallel execution of a large number of synchronized threads, which uses an instruction unit that is configured to issue instructions to a set of processing engines within each processing cluster.

[0205] In at least one embodiment, the operation of the processing cluster 2214 can be controlled by a pipeline manager 2432 that assigns processing tasks to SIMT parallel processors. In at least one embodiment, the pipeline manager 2432 Fig. 22 The scheduler 2210 receives instructions and manages the execution of these instructions through the graphics multiprocessor 2434 and / or the texture unit 2436. In at least one embodiment, the graphics multiprocessor 2434 is an exemplary instance of a SIMT parallel processor. However, in at least one embodiment, various types of SIMT parallel processors of different architectures may be included in the processing cluster 2414. In at least one embodiment, one or more instances of the graphics multiprocessor 2434 may be included in the processing cluster 2414. In at least one embodiment, the graphics multiprocessor 2434 may process data, and the data crossbar switch 2440 may be used to distribute the processed data to one of multiple possible destinations (including other shader units). In at least one embodiment, the pipeline manager 2432 may facilitate the distribution of processed data by specifying the destination of the processed data to be distributed via the data crossbar switch 2240.

[0206] In at least one embodiment, each graphics multiprocessor 2434 within a processing cluster 2414 may include the same set of function execution logic (e.g., arithmetic logic unit, load store unit, etc.). In at least one embodiment, the function execution logic may be configured in a pipelined manner, where new instructions may be issued before previous instructions are completed. In at least one embodiment, the function execution logic supports a variety of operations, including integer and floating point arithmetic, comparison operations, Boolean operations, shifts, and calculations of various algebraic functions. In at least one embodiment, the same functional unit hardware may be utilized to perform different operations, and any combination of functional units may be present.

[0207] In at least one embodiment, the instructions transmitted to the processing cluster 2414 constitute threads. In at least one embodiment, a group of threads executed across a group of parallel processing engines is a thread group. In at least one embodiment, the thread group executes the program on different input data. In at least one embodiment, each thread within a thread group may be assigned to a different processing engine within the graphics multiprocessor 2434. In at least one embodiment, a thread group may include fewer threads than the number of processing engines within the graphics multiprocessor 2434. 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 the cycle that is processing the thread group. In at least one embodiment, a thread group may also include more threads than the number of processing engines within the graphics multiprocessor 2434. In at least one embodiment, when a thread group includes more threads than the number of processing engines within the graphics multiprocessor 2434, processing may be performed in consecutive clock cycles. In at least one embodiment, multiple thread groups may be executed simultaneously on the graphics multiprocessor 2434.

[0208] In at least one embodiment, the graphics multiprocessor 2434 includes internal cache memory to perform load and store operations. In at least one embodiment, the graphics multiprocessor 2434 can abandon the internal cache and use the cache memory within the processing cluster 2414 (e.g., L1 cache 2248). In at least one embodiment, each graphics multiprocessor 2434 can also access the partition unit (e.g., Fig. 22 L2 caches within the partition units 2220A-2220N) of the graphics multiprocessor 2434 are shared among all processing clusters 2414 and can be used to transfer data between threads. In at least one embodiment, the graphics multiprocessor 2434 can also access off-chip global memory, which can include one or more of the local parallel processor memory and / or system memory. In at least one embodiment, any memory external to the parallel processing unit 2402 can be used as global memory. In at least one embodiment, the processing cluster 2414 includes multiple instances of the graphics multiprocessor 2434, which can share instructions and data that can be stored in the L1 cache 2448.

[0209] In at least one embodiment, each processing cluster 2414 may include a memory management unit ("MMU") 2445 configured to map virtual addresses to physical addresses. In at least one embodiment, one or more instances of MMU 2445 may reside in Fig. 222218. In at least one embodiment, the MMU 2245 includes a set of page table entries (PTEs) that are used to map virtual addresses to physical addresses of tiles (talk more about tiles) and in at least one embodiment to cache line indices. In at least one embodiment, the MMU 2445 may include an address translation lookaside buffer (TLB) or cache that may reside within the graphics multiprocessor 2434 or L1 cache or processing cluster 2414. In at least one embodiment, the physical addresses are processed to assign surface data access locality for efficient request interleaving between partition units. In at least one embodiment, the cache line index may be used to determine whether a request for a cache line is a hit or miss.

[0210] In at least one embodiment, the processing clusters 2414 can be configured such that each graphics multiprocessor 2434 is coupled to a texture unit 2436 to perform texture mapping operations, which may involve, for example, determining texture sample locations, reading texture data, and filtering texture data. In at least one embodiment, texture data is read from an internal texture L1 cache (not shown) or from an L1 cache within the graphics multiprocessor 2434, and texture data is retrieved from an L2 cache, local parallel processor memory, or system memory as needed. In at least one embodiment, each graphics multiprocessor 2434 outputs processed tasks to a data crossbar 2440 to provide the processed tasks to another processing cluster 2414 for further processing or to store the processed tasks in an L2 cache, local parallel processor memory, or system memory via a memory crossbar 2416. In at least one embodiment, a preROP 2442 (pre-raster operations unit) is configured to receive data from the graphics multiprocessor 2434, direct the data to a ROP unit, which may communicate with a partition unit (e.g., Fig. 22 In at least one embodiment, the PreROP 2442 unit can perform optimizations for color blending, organize pixel color data, and perform address translation.

[0211] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding inference and / or training logic 1015. In at least one embodiment, inference and / or training logic 1015 can be used in graphics processing cluster 2214 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 architectures, or neural network use cases described herein.

[0212] Fig.25 A graphics multiprocessor 2534 in at least one embodiment is shown. In at least one embodiment, the graphics multiprocessor 2534 is coupled to a pipeline manager 2532 of a processing cluster 2514. In at least one embodiment, the graphics multiprocessor 2534 has an execution pipeline that includes, but is not limited to, an instruction cache 2552, an instruction unit 2554, an address mapping unit 2556, a register file 2558, one or more general purpose graphics processing unit (GPGPU) cores 2562, and one or more load / store units 2566. The GPGPU cores 2562 and the load / store units 2566 are coupled to cache memory 2572 and shared memory 2570 via a memory and cache interconnect 2568.

[0213] In at least one embodiment, the instruction cache 2552 receives a stream of instructions to be executed from the pipeline manager 2532. In at least one embodiment, the instructions are cached in the instruction cache 2552 and dispatched for execution by the instruction unit 2554. In one embodiment, the instruction unit 2554 can dispatch instructions as thread groups (e.g., warps), assigning each thread of the thread group to a different execution unit within the GPGPU core 2562. In at least one embodiment, the instructions can access any local, shared, or global address space by specifying an address within the unified address space. In at least one embodiment, the address mapping unit 2556 can be used to convert addresses in the unified address space into different memory addresses that can be accessed by the load / store unit 2566.

[0214] In at least one embodiment, register file 2558 provides a set of registers for the functional units of graphics multiprocessor 2534. In at least one embodiment, register file 2558 provides temporary storage for operands for data paths of functional units (e.g., GPGPU core 2562, load / store unit 2566) connected to graphics multiprocessor 2534. In at least one embodiment, register file 2558 is divided between each functional unit such that a dedicated portion of register file 2558 is allocated to each functional unit. In at least one embodiment, register file 2558 is divided between different warps being executed by graphics multiprocessor 2534.

[0215] In at least one embodiment, the GPGPU cores 2562 may each include a floating point unit (FPU) and / or an integer arithmetic logic unit (ALU) for executing instructions of the graphics multiprocessor 2534. The GPGPU cores 2562 may be similar in architecture or the architecture may be different. In at least one embodiment, the first part of the GPGPU core 2562 includes a single-precision FPU and an integer ALU, while the second part 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 arithmetic or enable variable-precision floating-point arithmetic. In at least one embodiment, the graphics multiprocessor 2534 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 may also include fixed or special-function logic.

[0216] In at least one embodiment, the GPGPU core 2562 includes SIMD logic capable of executing a single instruction to multiple sets of data. In one embodiment, the GPGPU core 2562 can physically execute SIMD4, SIMD8 and SIMD9 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 by a single SIMD instruction. For example, in at least one embodiment, eight SIMT threads that perform the same or similar operations can be executed in parallel by a single SIMD8 logic unit.

[0217] In at least one embodiment, the memory and cache interconnect 2568 is an interconnect network that connects each functional unit of the graphics multiprocessor 2534 to the register file 2558 and the shared memory 2570. In at least one embodiment, the memory and cache interconnect 2568 is a crossbar interconnect that allows the load / store unit 2566 to implement load and store operations between the shared memory 2570 and the register file 2558. In at least one embodiment, the register file 2558 can operate at the same frequency as the GPGPU core 2562, so that the latency of data transfer between the GPGPU core 2562 and the register file 2558 is very low. In at least one embodiment, the shared memory 2570 can be used to enable communication between threads executing on the functional units within the graphics multiprocessor 2534. In at least one embodiment, the cache memory 2572 can be used as, for example, a data cache to cache texture data communicated between the functional units and the texture unit 2536. In at least one embodiment, the shared memory 2570 can also be used as a program-managed cache. In at least one embodiment, in addition to automatically cached data stored in cache memory 2572, threads executing on GPGPU core 2562 may programmatically store data in shared memory.

[0218] 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 can be communicatively coupled to the host processor / core via a bus or other interconnect (e.g., a high-speed interconnect such as PCIe or NVLink). In at least one embodiment, the GPU can be integrated with the core on the same package or chip and communicatively coupled to the core via an internal processor bus / interconnect (such as can be inside the package or chip). In at least one embodiment, regardless of the manner in which the GPU is connected, the processor core can assign work to the GPU in the form of a sequence of commands / instructions contained in a work descriptor. In at least one embodiment, the GPU then uses dedicated circuits / logic to efficiently process these commands / instructions.

[0219] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Details are provided regarding the inference and / or training logic 1015. In at least one embodiment, the inference and / or training logic 1015 may be used in the graphics multiprocessor 2234 to perform inference or prediction operations based at least in part on weight parameters computed using the neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0220] Fig.26 2600, which may include logic circuits for executing instructions. In at least one embodiment, the processor 2600 may execute instructions, including x86 instructions, ARM instructions, special instructions for application-specific integrated circuits (ASICs), and the like. In at least one embodiment, the processor 2610 may include registers for storing packed data, such as 64-bit wide MMXTM registers in microprocessors enabled with MMX technology by Intel Corporation of Santa Clara, California. In at least one embodiment, the MMX registers available in integer and floating point form may operate with packed data elements that accompany single instruction multiple data ("SIMD") and streaming SIMD extension ("SSE") instructions. In at least one embodiment, 128-bit wide XMM registers associated with SSE2, SSE3, SSE4, AVX or higher (generally referred to as "SSEx") technology may hold such packed data operands. In at least one embodiment, the processor 2610 may execute instructions to accelerate machine learning or deep learning algorithms, training, or reasoning.

[0221] In at least one embodiment, the processor 2600 includes an in-order front end ("front end") 2601 to extract instructions to be executed and prepare instructions for later use in the processor pipeline. In at least one embodiment, the front end 2601 may include several units. In at least one embodiment, an instruction prefetcher 2626 fetches instructions from a memory and provides the instructions to an instruction decoder 2628, which in turn decodes or interprets the instructions. For example, in at least one embodiment, the instruction decoder 2628 decodes the received instructions into one or more operations of so-called "microinstructions" or "microoperations" (also referred to as "micro-operations" or "microinstructions") that are executable by the machine. In at least one embodiment, the instruction decoder 2628 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, the trace cache 2630 can assemble the decoded microinstructions into a program-ordered sequence or trace in the microinstruction queue 2634 for execution. In at least one embodiment, when trace cache 2630 encounters a complex instruction, microcode ROM 2632 provides the microinstructions necessary to complete the operation.

[0222] In at least one embodiment, some instructions may be converted into a single micro-operation, while other instructions may require several micro-operations to complete the entire operation. In at least one embodiment, if more than four micro-operations are required to complete an instruction, the instruction decoder 2628 may access the microcode ROM 2632 to execute the instruction. In at least one embodiment, the instruction may be decoded into a small number of micro-operations for processing at the instruction decoder 2628. In at least one embodiment, if multiple micro-operations are required to complete the operation, the instruction may be stored in the microcode ROM 2632. In at least one embodiment, the trace cache 2630 references the entry point programmable logic array ("PLA") to determine the correct micro-instruction pointer for reading the microcode sequence from the microcode ROM 2632 to complete one or more instructions according to at least one embodiment. In at least one embodiment, after the microcode ROM 2632 completes the micro-operation sequencing of the instruction, the front end 2601 of the machine may resume fetching micro-operations from the trace cache 2630.

[0223] In at least one embodiment, an out-of-order execution engine ("out-of-order engine") 2603 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 stream to optimize performance when the instructions go down the pipeline and are scheduled for execution. The out-of-order execution engine 2603 includes, but is not limited to, a distributor / register renamer 2640, a memory microinstruction queue 2642, an integer / floating point microinstruction queue 2644, a memory scheduler 2646, a fast scheduler 2602, a slow / general floating point scheduler ("slow / general FP scheduler") 2604, and a simple floating point scheduler ("simple FP scheduler") 2606. In at least one embodiment, the fast scheduler 2602, the slow / general floating point scheduler 2604, and the simple floating point scheduler 2606 are also collectively referred to as "microinstruction schedulers 2602, 2604, 2606". The distributor / register renamer 2640 allocates the machine buffers and resources required for each microinstruction to execute in sequence. In at least one embodiment, the allocator / register renamer 2640 renames the logical registers into entries in the register file. In at least one embodiment, the allocator / register renamer 2640 also allocates entries for each microinstruction in one of two microinstruction queues, a memory microinstruction queue 2642 for memory operations and an integer / floating point microinstruction queue 2644 for non-memory operations, in front of the memory scheduler 2646 and the microinstruction schedulers 2602, 2604, 2606. In at least one embodiment, the microinstruction schedulers 2602, 2604, 2606 determine when the microinstructions are ready to execute based on the readiness of their slave 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 2602 of at least one embodiment can schedule on every half of the main clock cycle, while the slow / general floating point scheduler 2604 and the simple floating point scheduler 2606 can schedule once per main processor clock cycle. In at least one embodiment, microinstruction schedulers 2602, 2604, 2606 arbitrate on dispatch ports to schedule microinstructions for execution.

[0224] In at least one embodiment, execution block b11 includes, but is not limited to, integer register file / branch network 2608, floating point register file / branch network ("FP register file / branch network") 2610, address generation units ("AGU") 2612 and 2614, fast arithmetic logic units ("fast ALU") 2616 and 2618, slow arithmetic logic unit ("slow ALU") 2620, floating point ALU ("FP") 2622, and floating point move unit ("FP move") 2624. In at least one embodiment, integer register file / branch network 2608 and floating point register file / bypass network 2610 are also referred to herein as "register files 2608, 2610". In at least one embodiment, AGUs 2612 and 2614, fast ALUs 2616 and 2618, slow ALU 2620, floating point ALU 2622, and floating point move unit 2624 are also referred to herein as "execution units 2612, 2614, 2616, 2618, 2620, 2622, and 2624". In at least one embodiment, execution block b11 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).

[0225] In at least one embodiment, register files 2608, 2610 can be arranged between microinstruction schedulers 2602, 2604, 2606 and execution units 2612, 2614, 2616, 2618, 2620, 2622 and 2624. In at least one embodiment, integer register file / branch network 2608 performs integer operations. In at least one embodiment, floating point register file / branch network 2610 performs floating point operations. In at least one embodiment, each of register files 2608, 2610 can include but is not limited to a branch network that can bypass or forward the just completed result that has not yet been written to the register file to a new slave object. In at least one embodiment, register files 2608, 2610 can communicate data with each other. In at least one embodiment, integer register file / branch network 2608 can 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, floating point register file / branch network 2610 may include, but is not limited to, 128-bit wide entries, since floating point instructions typically have operands that are 64 to 128 bits wide.

[0226] In at least one embodiment, the execution units 2612, 2614, 2616, 2618, 2620, 2622, 2624 can execute instructions. In at least one embodiment, the register files 2608, 2610 store integer and floating point data operand values ​​that the microinstructions need to execute. In at least one embodiment, the processor 2600 may include, but is not limited to, any number of execution units 2612, 2614, 2616, 2618, 2620, 2622, 2624 and combinations thereof. In at least one embodiment, the floating point ALU 2622 and the floating point move unit 2624 can perform floating point, MMX, SIMD, AVX and SSE or other operations, including specialized machine learning instructions. In at least one embodiment, the floating point ALU 2622 may include, but is not limited to, a 64-bit by 64-bit floating point divider to perform division, square root and remainder micro-operations. In at least one embodiment, floating point hardware may be used to process instructions involving floating point values. In at least one embodiment, ALU operations can be passed to fast ALUs 2616, 2618. In at least one embodiment, fast ALUs 2616, 2618 can perform fast operations with an effective delay of half a clock cycle. In at least one embodiment, most complex integer operations go to slow ALUs 2620 because slow ALUs 2620 can include, but are not limited to, integer execution hardware for long-latency type operations, such as multipliers, shifts, flag logic, and branch processing. In at least one embodiment, memory load / store operations can be performed by ALUs 2612, 2614. In at least one embodiment, fast ALUs 2616, fast ALUs 2618, and slow ALUs 2620 can perform integer operations on 64-bit data operands. In at least one embodiment, fast ALUs 2616, fast ALUs 2618, and slow ALUs 2620 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 2622 and floating point move unit 2624 can be implemented to support a range of operands having bits of various widths. In at least one embodiment, the floating point ALU 2622 and floating point move unit 2624 can operate on 128-bit wide packed data operands in conjunction with SIMD and multimedia instructions.

[0227] In at least one embodiment, microinstruction schedulers 2602, 2604, 2606 schedule dependent operations before the parent load completes execution. In at least one embodiment, since microinstructions can be speculatively scheduled and executed in processor 2600, processor 2600 can also include logic for handling memory misses. In at least one embodiment, if a data load misses in the data cache, there may be a dependent operation running in the pipeline that temporarily does not have the correct data for the scheduler. 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 independent operations may be allowed to complete. 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 for text string comparison operations.

[0228] In at least one embodiment, the term "register" may refer to an onboard processor storage location that can be used as part of an instruction to identify an operand. In at least one embodiment, registers may be those that can be used from outside the processor (from a programmer's perspective). In at least one embodiment, registers may not be limited to a particular type of circuit. On the contrary, in at least one embodiment, registers 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 by circuits within the processor, such as dedicated physical registers, physical registers dynamically allocated using register renaming, a combination of dedicated and dynamically allocated physical registers, etc. In at least one embodiment, integer registers store 32-bit integer data. The register file of at least one embodiment also includes eight multimedia SIMD registers for packaging data.

[0229] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding the inference and / or training logic 1015. In at least one embodiment, part or all of the inference and / or training logic 1015 may be incorporated into the execution block 2611 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 the execution block 2611. In addition, weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALUs of the execution block 2611 to perform one or more machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0230] Fig. 27A deep learning application processor 2700 in at least one embodiment is shown. In at least one embodiment, the deep learning application processor 2700 uses instructions that, if executed by the deep learning application processor 2700, cause the deep learning application processor 2700 to perform some or all of the processes and techniques described throughout this disclosure. In at least one embodiment, the deep learning application processor 2700 is an application specific integrated circuit (ASIC). In at least one embodiment, the application processor 2700 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 2700 includes, but is not limited to, processing clusters 2710(1)-2710(12), inter-chip links (“ICLs”) 2720(1)-2720(12), inter-chip controllers (“ICCs”) 2730(1)-2730(2), second generation high bandwidth memory (“HBM2”) 2740(1)-2740(4), memory controllers (“Mem Ctrlr”) 2742(1)-2742(4), high bandwidth memory physical layers (“HBM PHY”) 2744(1)-2744(4), a management controller central processing unit (“management controller CPU”) 2750, serial peripheral interface, inter-integrated circuit, and general purpose input / output blocks (“SPI, I2C, GPIO”) 2760, peripheral component interconnect express controller and direct memory access blocks (“PCIe controller and DMA”) 2770, and sixteen-lane peripheral component interconnect express ports (“PCI Express”). x 16”)2780.

[0231] In at least one embodiment, the processing cluster 2710 may perform deep learning operations, including inference or prediction operations based on weight parameters calculated based on one or more training techniques, including those described herein. In at least one embodiment, each processing cluster 2710 may include, but is not limited to, any number and type of processors. In at least one embodiment, the deep learning application processor 2700 may include any number and type of processing clusters 2700. In at least one embodiment, the inter-chip link 2720 is bidirectional. In at least one embodiment, the inter-chip link 2720 and the inter-chip controller 2730 enable multiple deep learning application processors 2700 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, the deep learning application processor 2700 may include any number (including zero) and type of ICL 2720 and ICC 2730.

[0232] In at least one embodiment, HBM2 2740 provides a total of 32GB of memory. HBM2 2740(i) is associated with both memory controller 2742(i) and HBM PHY 2744(i). In at least one embodiment, any number of HBM2 2740 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 2742 and HBM PHY 2744. In at least one embodiment, SPI, I2C, GPIO 3360, PCIe controller 2760 and DMA 2770 and / or PCIe2780 can be replaced with any number and type of blocks to implement any number and type of communication standards in any technically feasible manner.

[0233] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding the inference and / or training logic 1015. 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 2700. In at least one embodiment, the deep learning application processor 2700 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 2700. In at least one embodiment, the processor 2700 can be used to perform one or more of the neural network use cases described herein.

[0234] Fig.28 2800. In at least one embodiment, the neuromorphic processor 2800 may receive one or more inputs from a source external to the neuromorphic processor 2800. In at least one embodiment, these inputs may be transmitted to one or more neurons 2802 within the neuromorphic processor 2800. In at least one embodiment, the neurons 2802 and their components may be implemented using circuits or logic including one or more arithmetic logic units (ALUs). In at least one embodiment, the neuromorphic processor 2800 may include, but is not limited to, thousands of instances of neurons 2802, but any suitable number of neurons 2802 may be used. In at least one embodiment, each instance of a neuron 2802 may include a neuron input 2804 and a neuron output 2806. In at least one embodiment, a neuron 2802 may generate an output that may be transmitted to the inputs of other instances of the neuron 2802. In at least one embodiment, the neuron input 2804 and the neuron output 2806 may be interconnected via a synapse 2808.

[0235] In at least one embodiment, the neurons 2802 and synapses 2808 may be interconnected such that the neuromorphic processor 2800 operates to process or analyze information received by the neuromorphic processor 2800. In at least one embodiment, the neuron 2802 may send an output pulse (or "trigger" or "spike") when the input received through the neuron input 2804 exceeds a threshold. In at least one embodiment, the neuron 2802 may sum or integrate the signal received at the neuron input 2804. For example, in at least one embodiment, the neuron 2802 may be implemented as a leaky integrate-trigger neuron, where if the sum (referred to as the "membrane potential") exceeds a threshold, the neuron 2802 may generate an output (or "trigger") using a transfer function such as a sigmoid or threshold function. In at least one embodiment, the leaky integrate-trigger neuron may sum the signal received at the neuron input 2804 into a membrane potential, and may apply an attenuation factor (or leakage) to reduce the membrane potential. In at least one embodiment, a leaky integrate-trigger neuron may trigger if multiple input signals are received at the neuron input 2804 fast enough to exceed a threshold (such as before the membrane potential decays too low to trigger). In at least one embodiment, the neuron 2802 may be implemented using a circuit or logic that receives an input, integrates the input to a 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. In addition, in at least one embodiment, the neuron 2802 may include, but is not limited to, a comparator circuit or logic that generates an output spike at the neuron output 2806 when the result of applying the transfer function to the neuron input 2804 exceeds a threshold. In at least one embodiment, once the neuron 2802 triggers, it may 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, the neuron 2802 may resume normal operation after a suitable period of time (or repair period).

[0236] In at least one embodiment, neurons 2802 may be interconnected via synapses 2808. In at least one embodiment, synapses 2808 may be operable to transmit a signal from an output of a first neuron 2802 to an input of a second neuron 2802. In at least one embodiment, a neuron 2802 may transmit information over more than one instance of synapse 2808. In at least one embodiment, one or more instances of a neuron output 2806 may be connected to an instance of a neuron input 2804 in the same neuron 2802 via an instance of synapse 2808. In at least one embodiment, an instance of a neuron 2802 that produces an output to be transmitted over an instance of synapse 2808 may be referred to as a "pre-synaptic neuron" relative to that instance of synapse 2808. In at least one embodiment, an instance of a neuron 2802 that receives an input transmitted through an instance of synapse 2808 may be referred to as a "post-synaptic neuron" relative to an instance of synapse 2808. In at least one embodiment, with respect to various instances of synapses 2808, because an instance of neuron 2802 can receive input from one or more instances of synapses 2808 and can also transmit output through one or more instances of synapses 2808, a single instance of neuron 2802 can be both a "pre-synaptic neuron" and a "post-synaptic neuron."

[0237] In at least one embodiment, neurons 2802 may be organized into one or more layers. Each instance of a neuron 2802 may have a neuron output 2806 that may fan out to one or more neuron inputs 2804 through one or more synapses 2808. In at least one embodiment, a neuron output 2806 of a neuron 2802 in a first layer 2810 may be connected to a neuron input 2804 of a neuron 2802 in a second layer 2812. In at least one embodiment, a layer 2810 may be referred to as a "feed-forward layer". In at least one embodiment, each instance of a neuron 2802 in an instance of a first layer 2810 may fan out to each instance of a neuron 2802 in a second layer 2812. In at least one embodiment, a first layer 2810 may be referred to as a "fully connected feed-forward layer". In at least one embodiment, each instance of a neuron 2802 in each instance of a second layer 2812 may fan out to less than all instances of a neuron 2802 in a third layer 2814. In at least one embodiment, a second layer 2812 may be referred to as a "sparsely connected feed-forward layer". In at least one embodiment, neurons 2802 in the second layer 2812 may fan out to neurons 2802 in multiple other layers, including neurons 2802 in the (same) second layer 2812. In at least one embodiment, the second layer 2812 may be referred to as a "recurrent layer." The neuromorphic processor 2800 may include, but is not limited to, any suitable combination of recurrent layers and feed-forward layers, including but not limited to sparsely connected feed-forward layers and fully connected feed-forward layers.

[0238] In at least one embodiment, the neuromorphic processor 2800 may include, but is not limited to, a reconfigurable interconnect architecture or a dedicated hardwired interconnect to connect the synapses 2808 to the neurons 2802. In at least one embodiment, the neuromorphic processor 2800 may include, but is not limited to, circuitry or logic that allows synapses to be assigned to different neurons 2802 as needed, depending on the neural network topology and neuron fan-in / fan-out. For example, in at least one embodiment, the synapses 2808 may be connected to the neurons 2802 using an interconnect architecture such as a network on a chip or through dedicated connections. In at least one embodiment, the synaptic interconnects and their components may be implemented using circuitry or logic.

[0239] Fig.292900, which may be a discrete graphics processing unit or may be a graphics processor integrated with multiple processing cores. In at least one embodiment, the graphics processor 2900 communicates with registers on the graphics processor 2900 and commands placed in memory via a memory-mapped I / O interface. In at least one embodiment, the graphics processor 2900 includes a memory interface 2914 for accessing memory. In at least one embodiment, the memory interface 2914 is an interface to local memory, one or more internal caches, one or more shared external caches, and / or to system memory.

[0240] In at least one embodiment, the graphics processor 2900 also includes a display controller 2902 for driving display output data to a display device 2920. In at least one embodiment, the display controller 2902 includes hardware for one or more overlay planes of the display device 2920 and a combination of multiple layers of video or user interface elements. In at least one embodiment, the display device 2920 can be an internal or external display device. In at least one embodiment, the display device 2920 is a head-mounted display device, such as a virtual reality (VR) display device or an augmented reality (AR) display device. In at least one embodiment, the graphics processor 2900 includes a video codec engine 2906 to encode, decode, or transcode media to, from, or between one or more media coding formats, including but not limited to Moving Picture Experts Group (MPEG) formats (e.g., MPEG-2), Advanced Video Coding (AVC) formats (e.g., H.264 / MPEG-4 AVC, and Society of Motion Picture and Television Engineers (SMPTE) 421M / VC-1), and Joint Photographic Experts Group (JPEG) formats (e.g., JPEG) and MotionJPEG (MJPEG) formats.

[0241] In at least one embodiment, graphics processor 2900 includes a block image transfer (BLIT) engine 2904 to perform two-dimensional (2D) rasterizer operations, including, for example, bit boundary block transfers. However, in at least one embodiment, 2D graphics operations are performed using one or more components of a graphics processing engine (GPE) 2910. In at least one embodiment, GPE 2910 is a compute engine for performing graphics operations, including three-dimensional (3D) graphics operations and media operations.

[0242] In at least one embodiment, GPE 2910 includes a 3D pipeline 2912 for performing 3D operations, such as rendering three-dimensional images and scenes using processing functions that operate on 3D primitive shapes (e.g., rectangles, triangles, etc.). 3D pipeline 2912 includes programmable and fixed function elements that perform various tasks and / or spawn execution threads to 3D / media subsystem 2915. While 3D pipeline 2912 can be used to perform media operations, in at least one embodiment, GPE 2910 also includes a media pipeline 2916 for performing media operations, such as video post-processing and image enhancement.

[0243] In at least one embodiment, the media pipeline 2916 includes fixed function or programmable logic units for performing one or more specialized media operations, such as video decode acceleration, video deinterlacing, and video encode acceleration, in place of or on behalf of the video codec engine 2906. In at least one embodiment, the media pipeline 2916 also includes a thread generation unit for generating threads for execution on the 3D / media subsystem 2915. In at least one embodiment, the generated threads perform computations for the media operations on one or more graphics execution units included in the 3D / media subsystem 2915.

[0244] In at least one embodiment, the 3D / media subsystem 2915 includes logic for executing threads generated by the 3D pipeline 2912 and the media pipeline 2916. In at least one embodiment, the 3D pipeline 2912 and the media pipeline 2916 send thread execution requests to the 3D / media subsystem 2915, which includes thread dispatch logic for arbitrating the various requests and dispatching them to available thread execution resources. In at least one embodiment, the execution resources include an array of graphics execution units for processing 3D and media threads. In at least one embodiment, the 3D / media subsystem 2915 includes one or more internal caches for thread instructions and data. In at least one embodiment, the subsystem 2915 also includes shared memory, which includes registers and addressable memory to share data between threads and store output data.

[0245] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Detail is provided regarding the inference and / or training logic 1015. In at least one embodiment, part or all of the inference and / or training logic 1015 may be incorporated into the processor 2900. For example, in at least one embodiment, the training and / or inference techniques described herein may use one or more ALUs included in the 3D pipeline 2912. In addition, in at least one embodiment, the inference and / or training operations described herein may use Fig.10 or Fig.11 In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of the graphics processor 2900 to execute one or more machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0246] Fig.30 is a block diagram of the hardware logic of a graphics processor core 3000 in at least one embodiment described herein. In at least one embodiment, the graphics processor core 3000 is included within a graphics core array. In at least one embodiment, the graphics processor core 3000 (sometimes referred to as a core slice) may be one or more graphics cores within a modular graphics processor. In at least one embodiment, the graphics processor core 3000 is an example of a graphics core slice, and the graphics processors described herein may include multiple graphics core slices based on a target power and performance envelope. In at least one embodiment, each graphics core 3000 may include a fixed function block 3030 coupled to multiple sub-cores 3001A-3001F, also referred to as sub-slices, which include modular blocks of general and fixed function logic.

[0247] In at least one embodiment, fixed function block 3030 includes a geometry / fixed function pipeline 3036, which may be shared by all sub-cores in graphics processor 3000, for example, in lower performance and / or lower power graphics processor implementations. In at least one embodiment, geometry / fixed function pipeline 3036 includes a 3D fixed function pipeline, a video front end unit, a thread generator and thread dispatcher, and a unified return buffer manager that manages a unified return buffer.

[0248] In at least one fixed embodiment, functional block 3030 also includes a graphics SoC interface 3037, a graphics microcontroller 3038, and a media pipeline 3039. Graphics SoC interface 3037 provides an interface between graphics core 3000 and other processor cores in the integrated circuit system on chip. In at least one embodiment, graphics microcontroller 3038 is a programmable subprocessor that can be configured to manage various functions of graphics processor 3000, including thread dispatching, scheduling, and preemption. In at least one embodiment, media pipeline 3039 includes logic that helps decode, encode, pre-process and / or post-process multimedia data including image and video data. In at least one embodiment, media pipeline 3039 implements media operations via requests to calculation or sampling logic within sub-cores 3001-3001F.

[0249] In at least one embodiment, the SoC interface 3037 enables the graphics core 3000 to communicate with a general application processor core (e.g., a CPU) and / or other components within the SoC, including memory hierarchy elements such as a shared last level cache, system RAM, and / or embedded on-chip or packaged DRAM. In at least one embodiment, the SoC interface 3037 may also enable communication with fixed function devices within the SoC (e.g., a camera imaging pipeline), and enable the use and / or implementation of global memory atomics that may be shared between the graphics core 3000 and the CPU within the SoC. In at least one embodiment, the SoC interface 3037 may also implement power management controls for the graphics core 3000 and enable interfaces between the clock domain of the graphics core 3000 and other clock domains within the SoC. In at least one embodiment, the SoC interface 3037 enables command buffers to be received from a command stream converter and a global thread dispatcher, which are configured to provide commands and instructions to each of one or more graphics cores within the graphics processor. In at least one embodiment, commands and instructions may be dispatched to media pipeline 3039 when media operations are to be performed, or may be assigned to geometry and fixed function pipelines (e.g., geometry and fixed function pipeline 3036, geometry and fixed function pipeline 3014) when graphics processing operations are to be performed.

[0250] In at least one embodiment, the graphics microcontroller 3038 can be configured to perform various scheduling and management tasks for the graphics core 3000. In at least one embodiment, the graphics microcontroller 3038 can perform graphics and / or computational workload scheduling on various graphics parallel engines within the execution unit (EU) arrays 3002A-3002F, 3004A-3004F in the sub-cores 3001A-3001F. In at least one embodiment, host software executing on a CPU core of a SoC including the graphics core 3000 can submit a workload to one of a plurality of graphics processor doorbells, which invokes scheduling operations on the appropriate graphics engine. In at least one embodiment, the scheduling operations include determining which workload to run next, submitting the workload to the command stream converter, preempting existing workloads running on the engine, monitoring the progress of the workload, and notifying the host software when the workload is completed. In at least one embodiment, graphics microcontroller 3038 may also facilitate low power or idle states for graphics core 3000, thereby providing graphics core 3000 with the ability to save and restore registers across low power state transitions within graphics core 3000 independent of an operating system and / or graphics driver software on the system.

[0251] In at least one embodiment, graphics core 3000 may have up to N modular sub-cores more or less than the sub-cores 3001A-3001F shown. For each set of N sub-cores, in at least one embodiment, graphics core 3000 may also include shared function logic 3010, shared and / or cache memory 3012, geometry / fixed function pipelines 3014, and additional fixed function logic 3016 to accelerate various graphics and compute processing operations. In at least one embodiment, shared function logic 3010 may include logic units (e.g., samplers, math and / or inter-thread communication logic) that may be shared by each of the N sub-cores within graphics core 3000. Shared and / or cache memory 3012 may be the last level cache for the N sub-cores 3001A-3001F within graphics core 3000, and may also be used as shared memory accessible by multiple sub-cores. In at least one embodiment, geometry / fixed function pipeline 3014 may be included in place of geometry / fixed function pipeline 3036 within fixed function block 3030 and may include the same or similar logic units.

[0252] In at least one embodiment, the graphics core 3000 includes additional fixed function logic 3016, which may include various fixed function acceleration logic for use by the graphics core 3000. In at least one embodiment, the additional fixed function logic 3016 includes an additional geometry pipeline for use in position-only shading. In position-only shading, there are at least two geometry pipelines, and in the full geometry pipeline and the culling pipeline within the geometry / fixed function pipeline 3016, 3036, it is an additional geometry pipeline that may be included in the additional fixed function logic 3016. In at least one embodiment, the culling pipeline is a trimmed version of the full geometry pipeline. In at least one embodiment, the full pipeline and the culling pipeline can execute different instances of an application, each with a separate environment. In at least one embodiment, position-only shading can hide long culling runs for discarded triangles, so that shading can be completed earlier in some cases. For example, in at least one embodiment, the culling pipeline logic in the additional fixed function logic 3016 can execute the position shader in parallel with the main application and generate critical results faster than the full pipeline because the culling pipeline obtains and masks the position attributes of the vertex without performing rasterization and rendering the pixels to the frame buffer. In at least one embodiment, the culling pipeline can use the generated critical results to calculate visibility information for all triangles, regardless of whether the triangles are culled. In at least one embodiment, the full pipeline (which can be called a replay pipeline in this case) can consume visibility information to skip culled triangles to mask only visible triangles that are ultimately passed to the rasterization stage.

[0253] In at least one embodiment, the additional fixed-function logic 3016 may also include machine learning acceleration logic, such as fixed-function matrix multiplication logic, for implementing optimizations including for machine learning training or inference.

[0254] In at least one embodiment, a set of execution resources is included within each graphics sub-core 3001A-3001F, which can be used to perform graphics, media, and compute operations in response to requests from a graphics pipeline, a media pipeline, or a shader program. In at least one embodiment, the graphics sub-core 3001A-3001F includes multiple EU arrays 3002A-3002F, 3004A-3004F, thread dispatch and inter-thread communication (TD / IC) logic 3003A-3003F, 3D (e.g., texture) samplers 3005A-3005F, media samplers 3006A-3006F, shader processors 3007A-3007F, and shared local memory (SLM) 3008A-3008F. Each of the EU arrays 3002A-3002F, 3004A-3004F includes a plurality of execution units, which are general purpose graphics processing units capable of servicing graphics, media, or compute operations, performing floating point and integer / fixed point logic operations, including graphics, media, or compute shader programs. In at least one embodiment, the TD / IC logic 3003A-3003F performs local thread dispatch and thread control operations for the execution units within the sub-core and facilitates communication between threads executed on the execution units of the sub-core. In at least one embodiment, the 3D samplers 3005A-3005F can read data associated with textures or other 3D graphics into memory. In at least one embodiment, the 3D samplers can read texture data differently based on the configured sampling state and texture format associated with a given texture. In at least one embodiment, the media samplers 3006A-3006F can perform similar read operations based on the type and format associated with the media data. In at least one embodiment, each graphics sub-core 3001A-3001F may alternatively include a unified 3D and media sampler. In at least one embodiment, threads executing on execution units within each sub-core 3001A-3001F may utilize shared local memory 3008A-3008F within each sub-core to enable threads executing within a thread group to execute using a pool of on-chip memory.

[0255] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Detail is provided regarding the inference and / or training logic 1015. In at least one embodiment, some or all of the inference and / or training logic 1015 may be incorporated into the graphics processor 3010. For example, in at least one embodiment, the training and / or inference techniques described herein may be used in the 3D pipeline 3010, the graphics microcontroller 3038, the geometry and fixed function pipelines 3014 and 3036, or the graphics processor 3010. Fig.30 In addition, in at least one embodiment, the inference and / or training operations described herein may use the addition Fig.10 or Fig.11 In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of the graphics processor 3000 to execute one or more of the machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0256] Figures 31A-31B Thread execution logic 3100 is shown for an array of processing elements that comprise a graphics processor core in at least one embodiment. Fig.31A At least one embodiment is shown in which thread execution logic 3100 is used. Fig.31B Exemplary internal details of an execution unit in at least one embodiment are shown.

[0257] like Fig.31AAs shown in, in at least one embodiment, the thread execution logic 3100 includes a shader processor 3102, a thread dispatcher 3104, an instruction cache 3106, a scalable execution unit array including a plurality of execution units 3108A-3108N, a sampler 3110, a data cache 3112, and a data port 3114. In at least one embodiment, the scalable execution unit array can be dynamically scaled by enabling or disabling one or more execution units (e.g., any one of execution units 3108A, 3108B, 3108C, 3108D to 3108N-1 and 3108N), for example, based on the computational requirements of the workload. In at least one embodiment, the scalable execution units are interconnected by an interconnect structure that links to each execution unit. In at least one embodiment, the thread execution logic 3100 includes one or more connections to a memory (such as a system memory or a cache memory) through one or more of the instruction cache 3106, the data port 3114, the sampler 3110, and the execution units 3108A-3108N. In at least one embodiment, each execution unit (e.g., 3108A) is an independently programmable general-purpose computing unit capable of executing multiple simultaneous hardware threads while processing multiple data elements in parallel for each thread. In at least one embodiment, the array of execution units 3108A-3108N is scalable to include any number of individual execution units.

[0258] In at least one embodiment, execution units 3108A-3108N are mainly used to execute shader programs. In at least one embodiment, shader processor 3102 can process various shader programs and dispatch execution threads associated with shader programs via thread dispatcher 3104. In at least one embodiment, thread dispatcher 3104 includes logic for arbitrating thread initialization celebrations from graphics and media pipelines and instantiating requested threads on one or more execution units in execution units 3108A-3108N. For example, in at least one embodiment, a geometry pipeline can dispatch vertices, tessellation or geometry shaders to thread execution logic for processing. In at least one embodiment, thread dispatcher 3104 can also process runtime thread generation requests from executing shader programs.

[0259] In at least one embodiment, the execution units 3108A-3108N support an instruction set that includes native support for many standard 3D graphics shader instructions, so that shader programs in graphics libraries (such as Direct 3D and OpenGL) can be executed with minimal translation. In at least one embodiment, the execution unit supports vertex and geometry processing (e.g., vertex programs, geometry programs, vertex shaders), pixel processing (e.g., pixel shaders, fragment shaders) and general processing (e.g., calculation and media shaders). In at least one embodiment, each execution unit 3108A-3108N includes one or more arithmetic logic units (ALUs) that can perform multiple-issue single instruction multiple data (SIMD), and multi-threaded operations implement an efficient execution environment despite higher latency memory access. In at least one embodiment, each hardware thread within each execution unit has a dedicated high-bandwidth register file and associated independent thread state. In at least one embodiment, execution is multiple issues per clock to the pipeline, and the pipeline is capable of integer, single-precision and double-precision floating-point operations, SIMD branch functions, logical operations, prior operations, and other operations. In at least one embodiment, while waiting for data from memory or one or more shared functions, dependency logic within execution units 3108A-3108N causes the waiting thread to sleep until the requested data is returned. In at least one embodiment, while a particular waiting thread is sleeping, hardware resources can be dedicated to processing other threads. For example, in at least one embodiment, during a delay associated with a vertex shader operation, an execution unit can perform operations on a pixel shader, a fragment shader, or another type of shader program (including a different vertex shader).

[0260] In at least one embodiment, each of the execution units 3108A-3108N operates on an array of data elements. In at least one embodiment, the number of data elements is the "execution size" or number of lanes of an instruction. In at least one embodiment, an execution lane is a logical unit used for the execution of data element access, masking, and flow control within an instruction. In at least one embodiment, the multiple lanes may be independent of the multiple physical arithmetic logic units (ALUs) or floating point units (FPUs) for a particular graphics processor. In at least one embodiment, the execution units 3108A-3108N support integer and floating point data types.

[0261] In at least one embodiment, the execution unit instruction set includes SIMD instructions. In at least one embodiment, various data elements can be stored in registers as packed data types, and the execution unit will process various elements based on the data size of those elements. For example, in at least one embodiment, when operating on a 256-bit wide vector, 256 bits of the vector are stored in registers, and the execution unit operates on the vector as four separate 64-bit packed data elements (quadword (QW) size data elements), eight separate 32-bit packed data elements (doubleword (DW) size data elements), sixteen separate 16-bit packed data elements (word (W) size data elements) or thirty-two separate 8-bit data elements (byte (B) size data elements). However, in at least one embodiment, different vector widths and register sizes are possible.

[0262] In at least one embodiment, one or more execution units may be combined into fused execution units 3109A-3109N with thread control logic (3107A-3107N) for fused EUs. In at least one embodiment, multiple EUs may be merged into one EU group. In at least one embodiment, the number of EUs in the fused EU group may be configured to execute separate SIMD hardware threads. The number of EUs in the fused EU group may vary according to various embodiments. In at least one embodiment, each EU may execute various SIMD widths, including but not limited to SIMD8, SIMD16, and SIMD32. In at least one embodiment, each fused graphics execution unit 3109A-3109N includes at least two execution units. For example, in at least one embodiment, the fused execution unit 3109A includes a first EU 3108A, a second EU 3108B, and a thread control logic 3107A shared with the first EU 3108A and the second EU 3108B. In at least one embodiment, thread control logic 3107A controls threads executing on fused graphics execution unit 3109A, allowing each EU within fused execution units 3109A-3109N to execute using an instruction pointer register.

[0263] In at least one embodiment, one or more internal instruction caches (e.g., 3106) are included in the thread execution logic 3100 to cache thread instructions for the execution unit. In at least one embodiment, one or more data caches (e.g., 3112) are included to cache thread data during thread execution. In at least one embodiment, a sampler 3110 is included to provide texture sampling for 3D operations and media sampling for media operations. In at least one embodiment, the sampler 3110 includes a specialized texture or media sampling function to process the texture or media data during the sampling process before providing the sampled data to the execution unit.

[0264] During execution, in at least one embodiment, the graphics and media pipeline sends thread initiation requests to the thread execution logic 3100 through the thread generation and dispatch logic. In at least one embodiment, once a set of geometric objects have been processed and rasterized into pixel data, the pixel processor logic (e.g., pixel shader logic, fragment shader logic, etc.) within the shader processor 3102 is called to further calculate output information and cause the results to be written to the output surface (e.g., color buffer, depth buffer, template buffer, etc.). In at least one embodiment, the pixel shader or fragment shader calculates the values ​​of various vertex attributes to be interpolated on the rasterized object. In at least one embodiment, the pixel processor logic within the shader processor 3102 then executes the pixel or fragment shader program provided by the application program interface (API). In at least one embodiment, in order to execute the shader program, the shader processor 3102 dispatches the thread to the execution unit (e.g., 3108A) via the thread dispatcher 3104. In at least one embodiment, the shader processor 3102 uses the texture sampling logic in the sampler 3110 to access the texture data in the texture map stored in the memory. In at least one embodiment, arithmetic operations on texture data and input geometry data calculate pixel color data for each geometry fragment, or discard one or more pixels for further processing.

[0265] In at least one embodiment, data port 3114 provides a memory access mechanism for thread execution logic 3100 to output processed data to memory for further processing on the graphics processor output pipeline. In at least one embodiment, data port 3114 includes or is coupled to one or more cache memories (e.g., data cache 3112) to cache data for memory access via the data port.

[0266] like Fig.31BAs shown, in at least one embodiment, the graphics execution unit 3108 may include an instruction fetch unit 3137, a general register file array (GRF) 3124, an architectural register file array (ARF) 3126, a thread arbiter 3122, an issue unit 3130, a branch unit 3132, a set of SIMD floating point units (FPUs) 3134, and in at least one embodiment, a set of dedicated integer SIMD ALUs 3135. In at least one embodiment, the GRF 3124 and ARF 3126 include a set of general register files and architectural register files associated with each simultaneous hardware thread that can be active in the graphics execution unit 3108. In at least one embodiment, the per-thread architectural state is maintained in the ARF 3126, while data used during thread execution is stored in the GRF 3124. In at least one embodiment, the execution state of each thread, including the instruction pointer of each thread, can be saved in thread-specific registers in the ARF 3126.

[0267] In at least one embodiment, graphics execution unit 3108 has an architecture that is a combination of simultaneous multithreading (SMT) and fine-grained interleaved multithreading (IMT). In at least one embodiment, the architecture has a modular configuration that can be fine-tuned at design time based on the target number of simultaneous threads and the number of registers per execution unit, where execution unit resources are logically allocated for executing multiple simultaneous threads.

[0268] In at least one embodiment, the graphics execution unit 3108 may issue multiple instructions together, each of which may be a different instruction. In at least one embodiment, the thread arbiter 3122 of the graphics execution unit thread 3108 may dispatch the instruction to one of the issue unit 3130, the branch unit 3142, or the SIMD FPU 3134 for execution. In at least one embodiment, each execution thread may access 128 general registers in the GRF 3124, each of which may store 32 bytes and may be accessed as a SIMD 8 element vector of 32-bit data elements. In at least one embodiment, each execution unit thread may access 4KB in the GRF 3124, although the embodiment is not limited thereto, and more or less register resources may be provided in other embodiments. In at least one embodiment, although the number of threads per execution unit may also vary depending on the embodiment, up to seven threads may be executed simultaneously. In at least one embodiment in which seven threads may access 4KB, the GRF 3124 may store a total of 28KB. In at least one embodiment, flexible addressing modes may allow registers to be addressed together to efficiently build wider registers or rectangular block data structures representing strides.

[0269] In at least one embodiment, memory operations, sampler operations, and other longer latency system communications are scheduled via "send" instructions executed by message passing send unit 3130. In at least one embodiment, dispatching branch instructions to dedicated branch unit 3132 facilitates SIMD divergence and eventual convergence.

[0270] In at least one embodiment, the graphics execution unit 3108 includes one or more SIMD floating point units (FPUs) 3134 to perform floating point operations. In at least one embodiment, one or more FPUs 3134 also support integer calculations. In at least one embodiment, one or more FPUs 3134 can SIMD perform up to M 32-bit floating point (or integer) operations, or SIMD perform up to 2M 16-bit integers or 16-bit floating point operations. In at least one embodiment, at least one of the one or more FPUs provides extended mathematical capabilities to support high-throughput a priori mathematical functions and double-precision 64-bit floating points. In at least one embodiment, there is also a set of 8-bit integer SIMD ALUs 3135, and can be specifically optimized to perform operations related to machine learning calculations.

[0271] In at least one embodiment, an array of multiple instances of graphics execution unit 3108 may be instantiated in graphics sub-core groupings (e.g., sub-slices). In at least one embodiment, execution unit 3108 may execute instructions across multiple execution lanes. In at least one embodiment, each thread executing on graphics execution unit 3108 executes on a different lane.

[0272] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Detailed information about the reasoning and / or training logic 1015 is provided. In at least one embodiment, part or all of the reasoning and / or training logic 1015 may be incorporated into the execution logic 3100. In addition, in at least one embodiment, other than Fig.10 or Fig.11 In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of execution logic 3100 to execute one or more machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0273] Fig.32A parallel processing unit ("PPU") 3200 in at least one embodiment is shown. In at least one embodiment, the PPU 3200 is configured with machine-readable code that, if executed by the PPU 3200, causes the PPU 3200 to perform some or all of the processes and techniques described throughout this disclosure. In at least one embodiment, the PPU 3200 is a multithreaded processor implemented on one or more integrated circuit devices, and utilizes multithreading as a latency hiding technique designed to process computer-readable instructions (also referred to as machine-readable instructions or simply instructions) executed in parallel on multiple threads. In at least one embodiment, a thread refers to an execution thread and is an instance of a group of instructions configured to be executed by the PPU 3200. In at least one embodiment, the PPU 3200 is a graphics processing unit ("GPU") configured to implement a graphics rendering pipeline for processing three-dimensional ("3D") graphics data to generate two-dimensional ("2D") image data for display on a display device, such as a liquid crystal display ("LCD") device. In at least one embodiment, the PPU 3200 is used to perform computations such as linear algebra operations and machine learning operations. Fig.32 The example parallel processor is shown for illustrative purposes only, and should be construed as a non-limiting example of a processor architecture contemplated within the scope of the present disclosure, and any suitable processor may be employed in addition and / or in place thereof.

[0274] In at least one embodiment, one or more PPUs 3200 are configured to accelerate high performance computing ("HPC"), data center, and machine learning applications. In at least one embodiment, PPU 3200 is configured to accelerate deep learning systems and applications, including the following non-limiting examples: autonomous vehicle platforms, deep learning, high-precision speech, image, text recognition systems, intelligent video analysis, molecular simulation, drug discovery, disease diagnosis, weather forecasting, big data analysis, astronomy, molecular dynamics simulation, financial modeling, robotics, factory automation, real-time language translation, online search optimization, and personalized user recommendations, etc.

[0275] In at least one embodiment, the PPU 3200 includes, but is not limited to, an input / output ("I / O") unit 3206, a front end unit 3210, a scheduler unit 3212, a work distribution unit 3214, a hub 3216, a crossbar switch ("Xbar") 3220, one or more general processing clusters ("GPCs") 3218, and one or more partition units ("memory partition units") 3222. In at least one embodiment, the PPU 3200 is connected to a host processor or other PPUs 3200 via one or more high-speed GPU interconnects ("GPU interconnects") 3208. In at least one embodiment, the PPU 3200 is connected to a host processor or other peripheral devices via an interconnect 3202. In an embodiment, the PPU 3200 is connected to a local memory including one or more memory devices ("memory") 3204. In at least one embodiment, the memory devices 3204 include, but are not limited to, one or more dynamic random access memory ("DRAM") devices. In at least one embodiment, one or more DRAM devices are configured and / or configurable as a high bandwidth memory ("HBM") subsystem with multiple DRAM dies stacked within each device.

[0276] In at least one embodiment, the high-speed GPU interconnect 3208 may refer to a wire-based multi-lane communication link that the system uses to scale and includes one or more PPUs 3200 in conjunction with one or more central processing units ("CPUs"), supporting cache coherence between the PPU 3200 and the CPU and CPU mastering. In at least one embodiment, the high-speed GPU interconnect 3208 transmits data and / or commands to other units of the PPU 3200, such as one or more copy engines, video encoders, video decoders, power management units, and / or other units in the PPU 3200 through the hub 3216. Fig.32 Other components that may not be explicitly shown.

[0277] In at least one embodiment, I / O unit 3206 is configured to receive data from a host processor ( Fig.32In at least one embodiment, the I / O unit 3206 communicates with the host processor directly through the system bus 3202 or through one or more intermediate devices (e.g., a memory bridge). In at least one embodiment, the I / O unit 3206 can communicate with one or more other processors (e.g., one or more PPUs 3200) via the system bus 3202. In at least one embodiment, the I / O unit 3206 implements a Peripheral Component Interconnect Express ("PCIe") interface for communicating over a PCIe bus. In at least one embodiment, the I / O unit 3206 implements an interface for communicating with external devices.

[0278] In at least one embodiment, the I / O unit 3206 decodes packets received via the system bus 3202. In at least one embodiment, at least some of the packets represent commands configured to cause the PPU 3200 to perform various operations. In at least one embodiment, the I / O unit 3206 sends the decoded commands to various other units of the PPU 3200 as specified by the commands. In at least one embodiment, the commands are sent to the front end unit 3210 and / or to the hub 3216 or other units of the PPU 3200, such as one or more replication engines, video encoders, video decoders, power management units, etc. ( Fig.32 In at least one embodiment, I / O unit 3206 is configured to route communications between various logical units of PPU 3200.

[0279] In at least one embodiment, a program executed by a host processor encodes a command stream in a buffer that provides a workload to the PPU 3200 for processing. In at least one embodiment, the workload includes instructions and data to be processed by those instructions. In at least one embodiment, the buffer is an area in memory that is accessible (e.g., read / write) by both the host processor and the PPU 3200—the host interface unit can be configured to access the buffer in system memory connected to the system bus 3202 via memory requests transmitted via the I / O unit 3206 through the system bus 3202. In at least one embodiment, the host processor writes the command stream to the buffer and then sends a pointer indicating the start of the command stream to the PPU 3200, so that the front end unit 3210 receives pointers to one or more command streams and manages one or more command streams, reads commands from the command streams and forwards the commands to the various units of the PPU 3200.

[0280] In at least one embodiment, the front end unit 3210 is coupled to a scheduler unit 3212 that configures the various GPCs 3218 to process tasks determined by one or more command streams. In at least one embodiment, the scheduler unit 3212 is configured to track state information related to the various tasks managed by the scheduler unit 3212, where the state information may indicate which GPC 3218 the task is assigned to, whether the task is active or inactive, a priority associated with the task, etc. In at least one embodiment, the scheduler unit 3212 manages multiple tasks that are executed on one or more GPCs 3218.

[0281] In at least one embodiment, the scheduler unit 3212 is coupled to a work distribution unit 3214, which is configured to dispatch tasks for execution on the GPCs 3218. In at least one embodiment, the work distribution unit 3214 tracks a plurality of scheduled tasks received from the scheduler unit 3212 and the work distribution unit 3214 manages a pending task pool and an active task pool for each GPC 3218. In at least one embodiment, the pending task pool includes a plurality of time slots (e.g., 32 time slots) containing tasks assigned to be processed by a particular GPC 3218; the active task pool may include a plurality of time slots (e.g., 4 time slots) for tasks actively processed by the GPC 3218, such that as one of the tasks in the GPC 3218 completes execution, the task is evicted from the active task pool of the GPC 3218 and one of the other tasks is selected from the pending task pool and scheduled for execution on the GPC 3218. In at least one embodiment, if the active task is idle on GPC 3218, such as while waiting for data dependencies to be resolved, the active task is evicted from GPC 3218 and returned to the pending task pool, while another task in the pending task pool is selected and scheduled for execution on GPC 3218.

[0282] In at least one embodiment, work distribution unit 3214 communicates with one or more GPCs 3218 via XBar 3220. In at least one embodiment, XBar 3220 is an interconnect network that couples many units of PPU 3200 to other units of PPU 3200 and can be configured to couple work distribution unit 3214 to a specific GPC 3218. In at least one embodiment, other units of one or more PPU 3200 can also be connected to XBar 3220 through hub 3216.

[0283] In at least one embodiment, tasks are managed by a scheduler unit 3212 and assigned to one of the GPCs 3218 by a work distribution unit 3214. The GPC 3218 is configured to process tasks and produce results. In at least one embodiment, the results can be consumed by other tasks in the GPC 3218, routed to a different GPC 3218 via an XBar 3220, or stored in memory 3204. In at least one embodiment, the results can be written to the memory 3204 via a partition unit 3222, which implements a memory interface for writing data to or reading data from the memory 3204. In at least one embodiment, the results can be transmitted to another PPU 3204 or a CPU via a high-speed GPU interconnect 3208. In at least one embodiment, the PPU 3200 includes, but is not limited to, U partition units 3222, which is equal to the number of separate and distinct memory devices 3204 coupled to the PPU 3200. In at least one embodiment, the following is combined Fig.34 The partition unit 3222 is described in more detail.

[0284] In at least one embodiment, the host processor executes a driver core that implements an application programming interface (API) that enables one or more applications executing on the host processor to schedule operations for execution on the PPU 3200. In one embodiment, multiple computing applications are executed simultaneously by the PPU 3200, and the PPU 3200 provides isolation, quality of service (“QoS”), and independent address spaces for the multiple computing applications. In at least one embodiment, the application generates instructions (e.g., in the form of API calls) that cause the driver core to generate one or more tasks for execution by the PPU 3200, and the driver core outputs the tasks to one or more streams processed by the PPU 3200. In at least one embodiment, each task includes one or more related groups of threads, which may be referred to as warps. In at least one embodiment, a warp includes multiple related threads (e.g., 32 threads) that can be executed in parallel. In at least one embodiment, a cooperative thread may refer to multiple threads that include instructions for executing tasks and exchanging data through shared memory. In at least one embodiment, in combination Fig.34 Threads and cooperating threads are described in greater detail according to at least one embodiment.

[0285] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11Details are provided regarding the inference and / or training logic 1015. In at least one embodiment, the deep learning application processor is used to train a machine learning model (such as a neural network) to predict or infer information provided to the PPU 3200. In at least one embodiment, the deep learning application processor 3200 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 the PPU 3200. In at least one embodiment, the PPU 3200 can be used to perform one or more of the neural network use cases described herein.

[0286] Fig.33 A general processing cluster ("GPC") 3300 is shown in at least one embodiment. In at least one embodiment, GPC 3300 is Fig.32 3218. In at least one embodiment, each GPC 3300 includes, but is not limited to, multiple hardware units for processing tasks, and each GPC 3300 includes, but is not limited to, a pipeline manager 3302, a pre-raster operations unit ("PROP") 3304, a raster engine 3308, a work distribution crossbar ("WDX") 3316, a memory management unit ("MMU") 3318, one or more data processing clusters ("DPCs") 3306, and any suitable combination of components.

[0287] In at least one embodiment, the operation of the GPC 3300 is controlled by a pipeline manager 3302. In at least one embodiment, the pipeline manager 3302 manages the configuration of one or more DPCs 3306 to process tasks assigned to the GPC 3300. In at least one embodiment, the pipeline manager 3302 configures at least one of the one or more DPCs 3306 to implement at least a portion of a graphics rendering pipeline. In at least one embodiment, the DPC 3306 is configured to execute vertex shader programs on a programmable streaming multiprocessor (“SM”) 3314. In at least one embodiment, the pipeline manager 3302 is configured to route packets received from the work distribution unit to appropriate logic units within the GPC 3300, and in at least one embodiment, some packets may be routed to fixed function hardware units in the PROP 3304 and / or the raster engine 3308, while other packets may be routed to the DPC 3306 for processing by the primitive engine 3312 or the SM 3314. In at least one embodiment, pipeline manager 3302 configures at least one of DPCs 3306 to implement a neural network model and / or a computational pipeline.

[0288] In at least one embodiment, PROP unit 3304 is configured to route data generated by raster engine 3308 and DPC 3306 to a raster operations ("ROP") unit in partition unit 3222 in at least one embodiment, in conjunction with Fig.32 Described in more detail. In at least one embodiment, the PROP unit 3304 is configured to perform optimizations for color blending, organize pixel data, perform address translation, and the like. In at least one embodiment, the raster engine 3308 includes, but is not limited to, a plurality of fixed-function hardware units configured to perform various raster operations, and in at least one embodiment, the raster engine 3308 includes, but is not limited to, a setup engine, a coarse raster engine, a culling engine, a clipping engine, a fine raster engine, a tile aggregation engine, and any suitable combination thereof. In at least one embodiment, the setup engine receives the transformed vertices and generates plane equations associated with the geometric primitives defined by the vertices; the plane equations are transmitted to the coarse raster engine to generate coverage information for the base primitives (e.g., an x, y coverage mask for the tile); the output of the coarse raster engine is transmitted to the culling engine, where fragments associated with primitives that fail the z test are culled, and to the clipping engine, where the fragments that are outside the frustum are clipped. In at least one embodiment, the clipped and culled fragments are passed to the fine raster engine to generate properties for the pixel fragments based on the plane equations generated by the setup engine. In at least one embodiment, the output of the raster engine 3308 includes fragments to be processed by any appropriate entity (e.g., by a fragment shader implemented within the DPC 3306).

[0289] In at least one embodiment, each DPC 3306 included in GPC 3300 includes, but is not limited to, an M-pipeline controller ("MPC") 3310; a primitive engine 3312; one or more SMs 3314; and any suitable combination thereof. In at least one embodiment, MPC 3310 controls the operation of DPC 3306, routing packets received from pipeline manager 3302 to appropriate units in DPC 3306. In at least one embodiment, packets associated with vertices are routed to primitive engine 3312, which is configured to fetch vertex attributes associated with the vertices from memory; conversely, packets associated with shader programs may be sent to SM 3314.

[0290] In at least one embodiment, SM 3314 includes, but is not limited to, a programmable stream processor configured to process tasks represented by multiple threads. In at least one embodiment, SM 3314 is multithreaded and configured to execute multiple threads (e.g., 32 threads) from a specific thread group simultaneously, and implements a single instruction, multiple data ("SIMD") architecture, wherein each thread in a group of threads (e.g., a warp) is configured to process different data sets based on the same instruction set. In at least one embodiment, all threads in a thread group execute the same instruction. In at least one embodiment, SM 3314 implements a single instruction, multiple thread ("SIMT") architecture, wherein each thread in a group of threads is configured to process different data sets based on the same instruction set, but wherein individual threads in a thread group are allowed to diverge during execution. In at least one embodiment, a program counter, call stack, and execution state are maintained for each warp, thereby achieving concurrency between warps and serial execution within the warp when threads in the warp diverge. In another embodiment, a program counter, call stack, and execution state are maintained for each individual thread, thereby enabling equal concurrency between all threads within and between warps. In at least one embodiment, execution state is maintained for each individual thread, and threads executing the same instruction can be converged and executed in parallel to improve efficiency. At least one embodiment of SM 3314 is described in more detail below.

[0291] In at least one embodiment, MMU 3318 is used between GPC 3300 and memory partition unit (e.g., Fig.32 The MMU 3318 provides an interface between the partition unit 3222 of the memory, and provides virtual address to physical address translation, memory protection, and arbitration of memory requests. In at least one embodiment, the MMU 3318 provides one or more translation lookaside buffers ("TLBs") for performing translation of virtual addresses to physical addresses in memory.

[0292] The reasoning and / or training logic 1015 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig.10 and / or Fig.11 Details are provided regarding inference and / or training logic 1015. In at least one embodiment, the deep learning application processor is used to train a machine learning model (such as a neural network) to predict or infer information provided to the GPC 3300. In at least one embodiment, the GPC 3300 is used to infer or predict information based on a machine learning model (e.g., a neural network) that has been trained by another processor or system or the GPC 3300. In at least one embodiment, the GPC 3300 can be used to perform one or more of the neural network use cases described herein.

[0293] Fig.34 A memory partition unit 3400 of a parallel processing unit ("PPU") according to at least one embodiment is shown. In at least one embodiment, the memory partition unit 3400 includes, but is not limited to, a raster operation ("ROP") unit 3402; a level 2 ("L2") cache 3404; a memory interface 3406; and any suitable combination thereof. The memory interface 3406 is coupled to a memory. The memory interface 3406 may implement a 32, 64, 128, 1024 bit data bus, or a similar implementation for high speed data transfer. In at least one embodiment, the PPU includes U memory interfaces 3406, one memory interface 3406 for each pair of partition units 3400, wherein each pair of partition units 3400 is connected to a corresponding memory device. For example, in at least one embodiment, the PPU may be connected to up to Y memory devices, such as a high bandwidth memory stack or a graphics double data rate version 5 synchronous dynamic random access memory ("GDDR5 SDRAM").

[0294] In at least one embodiment, the memory interface 3406 implements a high bandwidth memory second generation ("HBM2") memory interface, and Y is equal to half of U. In at least one embodiment, the HBM2 memory stack is located on the same physical package as the PPU, providing significant power and area savings compared to GDDR5 SDRAM systems. In at least one embodiment, each HBM2 stack includes, but is not limited to, four memory dies, and Y is equal to 4, each HBM2 stack includes two 128-bit channels per die for a total of 8 channels and a data bus width of 1024 bits. In at least one embodiment, the memory supports single error correction double error detection ("SECDED") error correction code ("ECC") to protect data. ECC provides higher reliability for computing applications that are sensitive to data corruption.

[0295] In at least one embodiment, the PPU implements a multi-level memory hierarchy. In at least one embodiment, the memory partition unit 3400 supports unified memory to provide a single unified virtual address space for the central processing unit ("CPU") and PPU memory, thereby enabling data sharing between virtual memory systems. In at least one embodiment, the frequency of PPU accesses to memory located on other processors is tracked to ensure that memory pages are moved to the physical memory of the PPU that accesses the page more frequently. In at least one embodiment, the high-speed GPU interconnect 3208 supports address translation services that allow the PPU to directly access the CPU's page tables and provide full access to the CPU's memory through the PPU.

[0296] In at least one embodiment, the copy engine transfers data between multiple PPUs or between a PPU and a CPU. In at least one embodiment, the copy engine can generate a page fault for an address that is not mapped in a page table, and the memory partition unit 3400 then services the page fault, maps the address into a page table, and then the copy engine performs the transfer. In at least one embodiment, fixed (or non-pageable) memory is operated for multiple copy engines between multiple processors, thereby substantially reducing the available memory. In at least one embodiment, in the event of a hardware page fault, the address can be passed to the copy engine without regard to whether it is a resident memory page, and the copy process is transparent.

[0297] According to at least one embodiment, from Fig.32 Data from memory 3204 or other system memory is retrieved by the memory partition unit 3400 and stored in an L2 cache 3404, which is located on the chip and shared between various GPCs. In at least one embodiment, each memory partition unit 3400 includes, but is not limited to, at least a portion of an L2 cache associated with a corresponding memory device. In at least one embodiment, lower levels of cache are implemented in various units within a GPC. In at least one embodiment, each SM 3314 can implement a level 1 ("L1") cache, where the L1 cache is private memory dedicated to a particular SM 3314, and data is retrieved from the L2 cache 3404 and stored in each L1 cache for processing in the functional units of the SM 3314. In at least one embodiment, the L2 cache 3404 is coupled to a memory interface 3406 and an XBar 3220.

[0298] In at least one embodiment, ROP unit 3402 performs graphics raster operations related to pixel color, such as color compression, pixel blending, etc. In at least one embodiment, ROP unit 3402 performs depth testing in conjunction with raster engine 3308, receiving the depth of a sample position associated with a pixel fragment from a culling engine of raster engine 3308. In at least one embodiment, the depth is tested against a corresponding depth in a depth buffer for the sample position associated with the fragment. In at least one embodiment, if the fragment passes the depth test for the sample position, ROP unit 3402 updates the depth buffer and sends the result of the depth test to raster engine 3308. It will be appreciated that the number of partition units 3400 can be different than the number of GPCs, and therefore, each ROP unit 3402 can be coupled to each GPC in at least one embodiment. In at least one embodiment, ROP unit 3402 tracks packets received from different GPCs and determines to which to route the results generated by ROP unit 3402 via XBar 3220.

[0299] Fig.35 Streaming multiprocessor ("SM") 3500 is shown in at least one embodiment. In at least one embodiment, SM 3500 is Fig.33 SMs. In at least one embodiment, SM 3500 includes, but is not limited to, instruction cache 3502; one or more scheduler units 3504; register file 3508; one or more processing cores ("cores") 3510; one or more special function units ("SFUs") 3512; one or more load / store units ("LSUs") 3514; interconnect network 3516; shared memory / level 1 ("L1") cache 3518; and any suitable combination thereof. In at least one embodiment, a work distribution unit schedules tasks for execution on a general processing cluster ("GPC") of a parallel processing unit ("PPU"), and each task is assigned to a specific data processing cluster ("DPC") within a GPC, and if the task is associated with a shader program, the task is assigned to one of the SMs 3500. In at least one embodiment, scheduler unit 3504 receives tasks from the work distribution unit and manages instruction scheduling for one or more thread blocks assigned to SM 3500. In at least one embodiment, the scheduler unit 3504 schedules thread blocks to execute as warps of parallel threads, where each thread block is assigned at least one warp. In at least one embodiment, each warp executes a thread. In at least one embodiment, the scheduler unit 3504 manages a plurality of different thread blocks, assigns warps to different thread blocks, and then dispatches instructions from a plurality of different cooperative groups to various functional units (e.g., processing core 3510, SFU 3512, and LSU 3514) in each clock cycle.

[0300] In at least one embodiment, a cooperative group may refer to a programming model for organizing groups of communicating threads that allows developers to express the granularity at which threads are communicating, thereby being able to express richer and more efficient parallel decompositions. In at least one embodiment, a cooperative launch API supports synchronization between thread blocks to execute parallel algorithms. In at least one embodiment, the application of the programming model provides a single, simple construct for synchronizing cooperative threads: a barrier across all threads of a thread block (e.g., a syncthreads() function). However, in at least one embodiment, programmers can define thread groups at a granularity smaller than a thread block and synchronize within the defined group to achieve higher performance, design flexibility, and software reuse in the form of a collection group scope function interface. In at least one embodiment, cooperative groups enable programmers to explicitly define thread groups at sub-block and multi-block granularity and perform collective operations, such as synchronizing threads in cooperative groups. The programming model supports clean composition across software boundaries, so that libraries and utility functions can be safely synchronized in their local environment without having to make assumptions about convergence. In at least one embodiment, cooperative group primitives enable new patterns of cooperative parallelism, including but not limited to producer-consumer parallelism, opportunistic parallelism, and global synchronization over the entire thread block grid.

[0301] In at least one embodiment, the dispatch unit 3506 is configured to send instructions to one or more of the functional units, and the scheduler unit 3504 includes, but is not limited to, two dispatch units 3506 that enable two different instructions from the same thread warp to be dispatched per clock cycle. In at least one embodiment, each scheduler unit 3504 includes a single dispatch unit 3506 or additional dispatch units 3506.

[0302] In at least one embodiment, each SM 3500 includes, in at least one embodiment, but is not limited to, a register file 3508 that provides a set of registers for the functional units of the SM 3500. In at least one embodiment, the register file 3508 is divided between each functional unit, so that a dedicated portion of the register file 3508 is allocated to each functional unit. In at least one embodiment, the register file 3508 is divided between different thread warps executed by the SM 3500, and the register file 3508 provides temporary storage for operands of the data path connected to the functional unit. In at least one embodiment, each SM 3500 includes, but is not limited to, a plurality of L processing cores 3510. In at least one embodiment, the SM 3500 includes, but is not limited to, a large number (e.g., 128 or more) of different processing cores 3510. In at least one embodiment, each processing core 3510 includes, in at least one embodiment, but is not limited to, a full pipeline, single-precision, double-precision, and / or mixed-precision processing unit, including, but not limited to, a floating-point arithmetic logic unit and an integer arithmetic logic unit. In at least one embodiment, the floating-point arithmetic logic unit implements the IEEE 754-2008 standard for floating-point arithmetic. In at least one embodiment, processing core 3510 includes, but is not limited to, 64 single precision (32-bit) floating point cores, 64 integer cores, 32 double precision (64-bit) floating point cores, and 8 tensor cores.

[0303] According to at least one embodiment, the tensor core is configured to perform matrix operations. In at least one embodiment, one or more tensor cores are included in the processing core 3510. In at least one embodiment, the tensor core is configured to perform deep learning matrix arithmetic, such as convolution operations for neural network training and reasoning. In at least one embodiment, each tensor core operates on a 4×4 matrix and performs a matrix multiplication and accumulation operation D=A×B+C, where A, B, C, and D are 4×4 matrices.

[0304] In at least one embodiment, the matrix multiplication inputs A and B are 16-bit floating point matrices, and the accumulation matrices C and D are 16-bit floating point or 32-bit floating point matrices. In at least one embodiment, the tensor core performs a 32-bit floating point accumulation operation on the 16-bit floating point input data. In at least one embodiment, the 16-bit floating point multiplication uses 64 operations and obtains a full-precision product, which is then accumulated with other intermediate products using 32-bit floating point addition to perform a 4x4x4 matrix multiplication. In at least one embodiment, the tensor core is used to perform larger two-dimensional or higher dimensional matrix operations composed of these smaller elements. In at least one embodiment, the API (such as the CUDA 9C++ API) exposes specialized matrix loads, matrix multiplications and accumulations, and matrix storage operations to efficiently use tensor cores from CUDA-C++ programs. In at least one embodiment, at the CUDA level, the warp level interface assumes a 16×16 size matrix across all 32 warp threads.

[0305] In at least one embodiment, each SM 3500 includes, but is not limited to, M SFUs 3512 that perform special functions (e.g., attribute evaluation, reciprocal square root, etc.). In at least one embodiment, the SFUs 3512 include, but are not limited to, tree traversal units configured to traverse a hierarchical tree data structure. In at least one embodiment, the SFUs 3512 include, but are not limited to, texture units configured to perform texture map filtering operations. In at least one embodiment, the texture units are configured to load texture maps (e.g., 2D arrays of texture pixels) from memory and sample the texture maps to produce sampled texture values ​​for use by shader programs executed by the SM 3500. In at least one embodiment, the texture maps are stored in shared memory / L1 cache 3518. In at least one embodiment, according to at least one embodiment, the texture units use mip-maps (e.g., texture maps with different levels of detail) to implement texture operations (such as filtering operations). In at least one embodiment, each SM 3500 includes, but is not limited to, two texture units.

[0306] In at least one embodiment, each SM 3500 includes, but is not limited to, N LSUs 3514 that implement load and store operations between the shared memory / L1 cache 3518 and the register file 3508. In at least one embodiment, each SM 3500 includes, but is not limited to, an interconnect network 3516 that connects each functional unit to the register file 3508, and the LSUs 3514 connect to the register file 3508 and the shared memory / L1 cache 3518. In at least one embodiment, the interconnect network 3516 is a crossbar switch that can be configured to connect any functional unit to any register in the register file 3508, and to connect the LSUs 3514 to memory locations in the register file 3508 and the shared memory / L1 cache 3518.

[0307] In at least one embodiment, shared memory / L1 cache 3518 is an array of on-chip memory that allows data storage and communication between SM 3500 and primitive engines and between threads in SM 3500 in at least one embodiment. In at least one embodiment, shared memory / L1 cache 3518 includes, but is not limited to, 128KB of storage capacity and is located in the path from SM 3500 to the partition unit. In at least one embodiment, shared memory / L1 cache 3518 is used in at least one embodiment for caching reads and writes. In at least one embodiment, one or more of shared memory / L1 cache 3518, L2 cache, and memory is a backing store.

[0308] In at least one embodiment, data cache and shared memory functions are combined into a single memory block, providing improved performance for both types of memory accesses. In at least one embodiment, the capacity is used by programs that do not use shared memory or as a cache, for example, if the shared memory is configured to use half of the capacity, texture and load / store operations can use the remaining capacity. According to at least one embodiment, the integration within the shared memory / L1 cache 3518 enables the shared memory / L1 cache 3518 to be used as a high throughput pipeline for streaming data while providing high bandwidth and low latency access to frequently reused data. In at least one embodiment, when configured for general parallel computing, a simpler configuration can be used compared to graphics processing. In at least one embodiment, the fixed-function graphics processing unit is bypassed, thereby creating a simpler programming model. In at least one embodiment, in a general parallel computing configuration, the work distribution unit directly allocates and distributes blocks of threads to DPCs. In at least one embodiment, threads in a block execute the same program, use unique thread IDs in computations to ensure that each thread generates unique results, use SM 3500 to execute the program and perform computations, use shared memory / L1 cache 3518 to communicate between threads, and use LSU 3514 to read and write global memory through shared memory / L1 cache 3518 and memory partitioning units. In at least one embodiment, when configured for general parallel computing, SM 3500 writes commands to scheduler unit 3504 that can be used to start new work on a DPC.

[0309] In at least one embodiment, the PPU is included in or coupled to a desktop computer, a laptop computer, a tablet computer, a server, a supercomputer, a smartphone (e.g., wireless, handheld device), a personal digital assistant ("PDA"), a digital camera, a vehicle, a head mounted display, a handheld electronic device, etc. In at least one embodiment, the PPU is implemented on a single semiconductor substrate. In at least one embodiment, the PPU is included in a system on a chip ("SoC") along with one or more other devices (e.g., additional PPUs, memory, a reduced instruction set computer ("RISC") CPU, one or more memory management units ("MMU"), a digital-to-analog converter ("DAC"), etc.).

[0310] In at least one embodiment, the PPU may be included on a graphics card that includes one or more storage devices. The graphics card may be configured to connect to a PCIe slot on a desktop computer motherboard. In at least one embodiment, the PPU may be an integrated graphics processing unit ("iGPU") included in a chipset of the motherboard.

[0311] Reasoning and / or training logic 1015 is used to perform reasoning and / or training operations related to one or more embodiments. Fig.10 and / or Fig.11 Provides details about the inference and / or training logic 1015. In at least one embodiment, the deep learning application processor is used to train a machine learning model (such as a neural network) to predict or infer information provided to the SM 3500. In at least one embodiment, the SM 3500 is used to infer or predict information based on a machine learning model (e.g., a neural network) that has been trained by another processor or system or by the SM 3500. In at least one embodiment, the SM 3500 can be used to perform one or more of the neural network use cases described herein.

[0312] In at least one embodiment, a single semiconductor platform may refer to a unique single semiconductor-based integrated circuit or chip. In at least one embodiment, a multi-chip module with increased connectivity may be used that emulates on-chip operations and provides substantial improvements over implementations utilizing a central processing unit ("CPU") and bus. In at least one embodiment, the various modules may also be placed separately or in various combinations of semiconductor platforms, depending on the needs of the user.

[0313] Other variations are within the spirit of the present disclosure. Thus, while the disclosed technology is susceptible to various modifications and alternative constructions, certain illustrated embodiments thereof are shown in the drawings and have been described in detail above. However, it should be understood that there is no intention to limit the disclosure to one or more specific forms disclosed, but on the contrary, it is intended to cover all modifications, alternative constructions, and equivalents that fall within the spirit and scope of the disclosure as defined by the appended claims.

[0314] Unless otherwise noted, unless otherwise noted or clearly contradicted by the context, the use of the terms "a", "an" and "the" and similar references in the context of describing the disclosed embodiments (particularly in the context of the appended claims) should be interpreted as covering both the singular and the plural. The terms "comprising", "having", "including" and "containing" should be interpreted as open-ended terms (meaning "including but not limited to"). The term "connected" when unmodified refers to a physical connection and should be understood as partly or completely contained in, connected to or connected together, even if there are any intervening. Unless otherwise indicated herein, the reference to numerical ranges herein is intended only to be used as a shorthand method of referring to each individual value falling within the range separately, and each individual value is incorporated into the specification as if it were recited separately herein. Unless otherwise indicated or contradicted by the context, the use of the term "group" (e.g., "a group of items") or "subset" should be interpreted as a non-empty set including one or more members. In addition, unless otherwise indicated or contradicted by the context, the term "subset" of a corresponding set does not necessarily mean a proper subset of the corresponding set, but a subset and a corresponding set may be equal.

[0315] Unless otherwise expressly indicated or clearly contradicted by the context, joint language such as phrases of the form "at least one of A, B, and C" or "at least one of A, B, and C" is understood in the context to be generally used to refer to items, and the term etc. can be A or B or C, or any non-empty subset of the set of A and B and C. For example, in a set having three members, the conjunction phrases "at least one of A, B, and C" and "at least one of A, B, and C" refer to any of the following sets: {A}, {B}, {C}, {A, B}, {A, C}, {B, C}, {A, B, C}. Therefore, this joint language is not intended to imply that certain embodiments require the presence of at least one of A, at least one of B, and at least one of C. In addition, unless otherwise stated or contradicted by the context, the term "plurality" indicates a plural state (e.g., "plurality of items" means a plurality of items). The number of items in the plural is at least two, but can be plural when explicitly or by the context indicates. In addition, unless otherwise stated or clear from the context, the phrase "based on" means "based at least in part on" rather than "based only on".

[0316] Unless otherwise noted herein or clearly contradictory to the environment, the operations of the processes described herein may be performed in any suitable sequence. In at least one embodiment, processes such as those described herein (or their variations and / or combinations thereof) are performed under the control of one or more computer systems configured with executable instructions, and are implemented as codes (e.g., executable instructions, one or more computer programs, or one or more applications) that are jointly executed on one or more processors by hardware or a combination thereof. In at least one embodiment, the code is stored on a computer-readable storage medium, for example, in the form of a computer program, which includes a plurality of instructions that can be executed by one or more processors. In at least one embodiment, the computer-readable storage medium is a non-transitory computer-readable storage medium that does not include a transient signal (e.g., a transient electrical or electromagnetic transmission of a propagation), but includes a non-transitory data storage circuit (e.g., a buffer, a cache, and a queue) in a transceiver of a transient signal. In at least one embodiment, code (e.g., executable code or source code) is stored on a set of one or more non-transitory computer-readable storage media (or other memory for storing executable instructions) on which executable instructions are stored, and when executed by one or more processors of a computer system, the computer system causes the computer system to perform the operations described herein. In at least one embodiment, a set of non-transitory computer-readable storage media includes multiple non-transitory computer-readable storage media and one or more single non-transitory storage media of multiple non-transitory computer-readable storage media lacking all codes, and multiple non-transitory computer-readable storage media store all codes together. In at least one embodiment, executable instructions are executed so that different instructions are executed by different processors, for example, a non-transitory computer-readable storage medium stores instructions, and a main central processing unit ("CPU") executes some instructions, while a graphics processing unit ("GPU") executes other instructions. In at least one embodiment, different components of the computer system have separate processors, and different processors execute different subsets of instructions.

[0317] Thus, in at least one embodiment, a computer system is configured to implement one or more services that individually or collectively perform the operations of the processes described herein, and such a computer system is configured with applicable hardware and / or software capable of implementing the operations. In addition, a computer system that implements at least one embodiment of the present disclosure is a single device, and in another embodiment, is a distributed computer system that includes multiple devices that operate in different ways, so that the distributed computer system performs the operations described herein, and so that a single device does not perform all operations.

[0318] The use of any and all examples or exemplary language (e.g., "such as") provided herein is intended only to better illustrate embodiments of the present disclosure and does not limit the scope of the disclosure unless otherwise stated. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of the disclosure.

[0319] All references, including publications, patent applications, and patents, cited herein are hereby incorporated by reference to the same extent as if each reference were individually and specifically indicated to be incorporated by reference.

[0320] In the description and claims, the terms "coupled" and "connected" and their derivatives may be used. It should be understood that these terms may not be intended as synonyms for each other. On the contrary, in specific examples, "connected" or "coupled" may be used to indicate that two or more elements are in direct or indirect physical or electrical contact with each other. "Coupled" may also mean that two or more elements are not in direct contact with each other, but still cooperate or interact with each other.

[0321] Unless otherwise specifically stated, it is understood that throughout the specification, terms such as "processing", "calculating", "computing", "determining", etc. refer to the actions and / or processes of a computer or computing system or similar electronic computing device to process and / or convert data represented as physical quantities (e.g., electronic) in registers and / or memories of the computing system into other data similarly represented as physical quantities in memories, registers or other such information storage, transmission or display devices of the computing system.

[0322] In a similar manner, the term "processor" may refer to any device or part of a device that processes electronic data from registers and / or memory and converts the electronic data into other electronic data that can be stored in registers and / or memory. As a non-limiting example, a "processor" may be a CPU or a GPU. A "computing platform" may include one or more processors. As used herein, a "software" process may include, for example, software and / or hardware entities that perform work over time, such as tasks, threads, and intelligent agents. Moreover, each process may refer to multiple processes to execute instructions continuously or intermittently, sequentially or in parallel. Because a system may embody one or more methods and a method may be considered a system, the terms "system" and "method" are used interchangeably herein.

[0323] In this document, reference may be made to obtaining, acquiring, receiving or inputting analog or digital data into a subsystem, a computer system or a computer-implemented machine. The process of obtaining, acquiring, receiving or inputting analog and digital data may be accomplished in a variety of ways, such as by receiving data as a parameter of a function call or a call to an application programming interface. In some embodiments, the process of obtaining, acquiring, receiving or inputting analog or digital data may be accomplished by transmitting data via a serial or parallel interface. In another embodiment, the process of obtaining, acquiring, receiving or inputting analog or digital data may be accomplished by transmitting data from a providing entity to an acquiring entity via a computer network. Reference may also be made to providing, outputting, transmitting, sending or presenting analog or digital data. In various examples, the process of providing, outputting, transmitting, sending or presenting analog or digital data may be accomplished by transmitting data as an input or output parameter of a function call, an application programming interface or an inter-process communication mechanism.

[0324] Although the above discussion sets forth example implementations of the described techniques, other architectures may be used to implement the described functionality and are intended to be within the scope of the present disclosure. In addition, although specific responsibility allocations are defined above for discussion purposes, various functions and responsibilities may be allocated and divided in different ways depending on the circumstances.

[0325] Furthermore, although the subject matter has been described in language specific to structural features and / or methodological acts, it should be understood that the subject matter claimed in the appended claims is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as exemplary forms of implementing the claims.

Claims

1. A processor, comprising: One or more arithmetic logic units (ALUs) to facilitate detection of the one or more objects in the image based at least in part on one or more pseudo-labels corresponding to the one or more objects.

2. The processor of claim 1 , wherein the one or more ALUs are further configured to: The one or more objects are detected using a neural network, the neural network being trained in part by generating a region of interest (ROI) score for each of a set of object candidates, the one or more pseudo labels being determined using the ROI scores.

3. The processor of claim 2, wherein the one or more ALUs are further configured to: The one or more objects are detected using a neural network that is trained in part by providing the ROI scores to at least one refinement branch that is capable of generating refined ROI scores for the one or more pseudo-labels. 4 . The processor of claim 3 , wherein the at least one refinement branch performs at least one of ROI quantization and selection, or removing pseudo labels for redundant candidates.

5. The processor of claim 1 , wherein the one or more ALUs are further configured to: The one or more objects are detected using a neural network trained in part by selecting the one or more pseudo-labels corresponding to the one or more objects having a refined ROI score exceeding an object threshold.

6. A system comprising: one or more processors configured to detect the one or more objects in the image using one or more neural networks based at least in part on one or more pseudo-labels corresponding to the one or more objects in the image; as well as One or more memories for storing the one or more neural networks.

7. A processor, comprising: One or more arithmetic logic units (ALUs) for facilitating training one or more neural networks for detecting one or more objects in an image based at least in part on one or more pseudo-labels corresponding to the one or more objects in the image.

8. A system comprising: One or more processors for facilitating training of one or more neural networks for detecting one or more objects in an image based at least in part on one or more pseudo-labels corresponding to the one or more objects in the image.

9. A method comprising: One or more neural networks are trained to detect one or more objects in an image based at least in part on one or more pseudo-labels corresponding to the one or more objects.

10. A system comprising: a camera configured to capture an image; one or more processors configured to detect the one or more objects in the image using one or more neural networks based at least in part on one or more pseudo-labels corresponding to the one or more objects; as well as A storage device for storing information about the detected object.

Images