Integrated circuit power system with machine learning capability

Abstract

A power system is provided that uses machine learning algorithms to address various issues involving power delivery on integrated circuit systems. The power system can process data on a platform near a target integrated circuit or off-platform in the cloud, so that machine learning algorithms can extract information from the data, process and analyze the data, and perform appropriate actions based on the analysis results. Applying machine learning to integrated circuit power delivery can involve the application of algorithms such as anomaly detection, load prediction, regression, and classification. Operating in this manner, the power system can be provided with improved voltage / frequency scaling capabilities, safety, and power efficiency.

Description

Background Technology

[0001] Integrated circuits such as application-specific integrated circuits (ASICs) and programmable logic devices (PLDs) are frequently used to implement high-performance applications with high throughput or high bandwidth requirements. Such high-performance applications require these integrated circuits to be powered by precisely controlled power supply voltages within tight bandgap and well-controlled sequences.

[0002] Existing system power policies require data associated with the power rail, such as voltage, current, and temperature available to the power system and users. While such power-related data is available, any collected data is not utilized for online analysis due to the significant challenges in processing large volumes of data in practice, where the dimensionality of the data limits the computational power to provide actionable steps for the system.

[0003] Examples of conventional power systems involve data acquisition; however, the data is sampled and stored for offline batch analysis. Data storage does not occur within the system, and data is not continuously captured online, limiting the ability to continuously monitor the power rails for the system. Another example includes online system analysis for buck converters, but the relevant data is not stored in the system's memory, limiting applicability to recursive online methods only. Furthermore, none of the existing methods describe predicting load disturbances at the power rails by monitoring past load behavior.

[0004] The embodiments described herein occur within this context. Attached Figure Description

[0005] Figure 1 This is a diagram of an illustrative power analysis system according to an embodiment.

[0006] Figure 2 It is for operation according to the embodiment. Figure 1 The flowchart illustrates the steps of the power analysis system shown.

[0007] Figure 3 This is a diagram of an illustrative power conversion system according to an embodiment.

[0008] Figure 4A This is a diagram illustrating how data can be reduced via compression according to an embodiment.

[0009] Figure 4B This is a diagram illustrating how data can be reduced via symbolic representation according to an embodiment.

[0010] Figure 5A and Figure 5B This is a diagram illustrating how dynamic signals can be represented in state space according to an embodiment.

[0011] Figures 6A-6C This is a diagram illustrating a machine learning method for detecting anomalies according to an embodiment.

[0012] Figures 7A-7C This is a diagram illustrating a machine learning method for classifying different data types according to an embodiment.

[0013] Figure 8A and Figure 8B This is a diagram of an illustrative predictive control system according to an embodiment.

[0014] Figure 9A This is a diagram illustrating a feedforward neural network for a predictive control system according to an embodiment.

[0015] Figure 9B This is a timing diagram illustrating how extreme load deviations can be predicted according to an embodiment.

[0016] Figure 10 This illustrates how load prediction, according to an embodiment, can help reduce power supply voltage deviation. Detailed Implementation

[0017] This embodiment relates to integrated circuits, and more particularly to power systems that use machine learning algorithms to solve various problems related to power delivery to integrated circuits. The power system can process data locally on a platform including the integrated circuit, power conversion system, and power data processor, or it can process data off-platform in the cloud. Applying machine learning to power delivery to integrated circuits can involve the application of algorithms such as anomaly detection, load prediction, and classification. Power-related data can be extracted from the system, processed into appropriate forms, analyzed, and actions taken to address potential problems.

[0018] Those skilled in the art will recognize that this exemplary embodiment can be practiced without some or all of these specific details. In other instances, well-known operations have not been described in detail so as not to unnecessarily obscure this embodiment.

[0019] Figure 1 Systems such as power analysis system 100 are shown. Figure 1As shown, the power analysis system 100 may include a local platform 102 coupled to an offline platform subsystem such as cloud 112. Platform 102 may include an integrated circuit such as device 104, a power conversion system such as power conversion system 106, and a processor such as power data processor 108. Integrated circuit device 104 may be an application-specific integrated circuit (ASIC), a central processing unit (CPU), a graphics processing unit (GPU), a microcontroller, a microprocessor, a programmable integrated circuit such as a programmable logic device (PLD) or a field-programmable gate array (FPGA), or other suitable types of integrated circuits.

[0020] Device 104 may have one or more power supply terminals (sometimes referred to as power rails). Each power supply terminal may receive a power supply voltage signal from the corresponding power conversion system 106. Figure 1 In the example, device 104 can receive a power supply voltage Vout from power conversion system 106, but can also receive other power supply voltages, such as Vout' and Vout'', from additional power conversion systems.

[0021] Power conversion system 106 may be a DC-DC converter (for example) that receives an input power supply voltage Vin from power supply source 110. In this configuration, power conversion system 106 can be used to convert the input voltage Vin and generate a corresponding output voltage Vout that is then delivered to device 104. Power conversion system 106 may include a data collection block 120 configured to extract and collect data from the power supply. Data collected on platform 102 (sometimes referred to as platform data) may include data associated with voltage, current, power, temperature, mode, and / or any activity at device 104 (for example). Data collection block 120 may include data reduction circuitry to limit the amount and rate of data that needs to be analyzed by system 100.

[0022] The power conversion system 106 can be coupled to the power data processor 108 via a communication bus such as bus 107. Data collected at the power conversion system 106 can be transmitted to the power data processor 108 via bus 107 for local processing. The power data processor 108 may include a local analysis block 122 and a local power policy block 124. The local analysis block 122 performs local analysis on the received data and guides the local power policy block 124 to take appropriate actions to control local power delivery based on the analysis results.

[0023] The power data processor 108 can be coupled to a cloud computing subsystem 112, which is configured to provide cloud-based analytics on collected power conversion data. The cloud 112 may include a remote analytics block 126 and a remote power policy block 128. The remote analytics block 126 can be configured to perform more data-intensive analytics and apply various machine learning algorithms to large volumes of data received from platform 102 or from multiple platforms 102 aggregated from multiple integrated circuit devices. Operating in this manner, the remote analytics block 126 can then control local power delivery by instructing the local power policy block 124 to take appropriate actions at the corresponding platform(s). In other words, the local, edge, or cloud-based analytics block can determine appropriate actions to be taken based on the results of the analytics and / or provide updated power models based on the data.

[0024] Figure 2 This is a flowchart illustrating the steps used to operate the power analysis system 100. Generally, local and remote analysis blocks are configured to perform data processing, data analysis, and data learning, while local and remote policy blocks implement appropriate actions.

[0025] At step 200, the power conversion system 106 may collect data. At step 202, the power conversion system 106 may process the collected data (e.g., by using compression, symbolic representation, etc. to reduce the data).

[0026] At step 204, the reduced data may optionally be fed to the data processor 108 via bus 107 and analyzed by the local analysis block. If needed, the data may also be fed to the cloud and analyzed by a remote analysis block.

[0027] At step 206, the local / remote analytics block can apply machine learning algorithms to aid in data analysis. For example, the analytics block can be configured to perform anomaly detection, data type classification, load prediction, etc.

[0028] Machine learning processes can be supervised or unsupervised. In supervised learning scenarios (e.g., for classification algorithms), the learning steps can be computationally intensive and typically run during a separate learning phase, but are not required during operations after learning is complete. Therefore, supervised learning is more suitable for offline platforms residing (e.g., in the cloud). Data analysis and action steps will still be processed locally because they typically have faster response times.

[0029] In unsupervised learning scenarios (e.g., for anomaly detection), the learning steps will run during normal operation (e.g., learning will run continuously in real time). Therefore, unsupervised learning is more suitable for being split up, allowing time-sensitive and non-computationally intensive tasks to be performed on the platform, while non-time-sensitive and computationally intensive tasks will be performed off-platform in the cloud.

[0030] At step 208, the local / remote power policy block can guide the power conversion system to take appropriate actions based on the results of machine learning analysis. Some exemplary actions that may be taken may include alerting a monitor system on a remote platform, changing the mode or parameters of the DC-DC converter, changing the mode or parameters of other components in the power system, providing the device 104 with an operational status so that the device 104 knows what is happening, changing the Vout level or any voltage level in the power system, changing the operating clock frequency of the device 104, reconfiguring the local platform differently, predicting upcoming loading steps at the device 104, and so on.

[0031] Figure 2 The steps described are merely illustrative and are not intended to limit this embodiment. At least some of the existing steps may be modified or omitted; some steps may be performed in parallel; additional steps may be added or inserted; and the order of some steps may be reversed or changed. For example, these steps may be performed for each power supply voltage Vout, Vout', ... and Vout'' delivered to the target integrated circuit device 104.

[0032] Figure 3 This is a circuit diagram of a suitable implementation of the power conversion system 106. For example... Figure 3 As shown, the power conversion system 106 may include a power stage 300 and associated control circuitry 302 for controlling the power stage 300. The power stage 300 may be configured to convert an input voltage Vin to a corresponding output voltage Vout. The control circuitry 302 may include controller circuitry 314, power manager circuitry 316, local memory circuitry 318, and data reduction circuitry 320. In a suitable arrangement, all circuitry of the control circuitry 302 may be formed as a system-on-a-chip (SoC) on a single integrated circuit die. If desired, circuitry 302 and the power stage 300 may be implemented together on a single chip.

[0033] The reduction circuit 320 can be configured to reduce the amount of data collected via downsampling, binning, or decimation to allow the bus interface 107 to operate at an appropriate rate. The reduced data can be stored locally in memory 318 before being transmitted to the power data processor via path 107.

[0034] Controller 314 can be configured to receive data from data converter 310 (e.g., analog-to-digital data converter), the data being proportional to the control error between Vout and a predetermined setpoint (e.g., a target power supply voltage level or a nominal power supply voltage level). Controller 315 outputs a signal to power stage 300 to control the conversion of input signal Vin to output voltage Vout. For example, the control signal could be a pulse width modulation (PWM) signal controlling a corresponding high-power switching device. In such a PWM system, the ADC sampling rate can be at least as fast as the switching frequency of the PWM signal. A sampling rate of 200 ksps (kilosamples per second) to 2 Msps (megasamples per second) or higher can be used at ADC 310.

[0035] Power manager 316 can receive data from data converter 312 (e.g., ADC circuitry). Converter 312 can be coupled to multiplexer 311 to enable monitoring of various system signals for fault detection. For example, multiplexer 311 can receive Vout, Vin, inductor current I1 in power stage 300 (which can be used as a proxy for the output current of power stage 300); temperature Tj (e.g., the temperature of the system or a specific component such as an inductor or power switch); and so on.

[0036] The power manager 316 can also manage the power on and off of Vout, handle actions for faults and warnings, and perform signal conditioning such as filtering monitored signals. Because the ADC 312 associated with the power manager 316 must sample several signals, it can provide data at a relatively lower rate than the controller 314. For example, a sampling rate of 1-100 ksps might be used at the ADC 312. The bus interface 107 can convey data sampled from monitored signals, the status of fault warnings, and can receive configuration commands and data from the power data processor 108 or other host microcontrollers.

[0037] Memory circuit 318 can be configured to store data from data converter 310 and / or data converter 312. Memory circuit 318 can process sampled data, preferably in conjunction with data reduction circuit 320 and power manager 316 to reduce data by means of deriving signals that are not directly available or by other means of data reduction.

[0038] Figure 3The exemplary arrangement of the power conversion system 106 shown is merely illustrative. System 106 may include more or fewer circuit components to perform data collection functions if desired. Circuits 314, 316, 318, and 320 may be interconnected in other ways and may include any number of data converters to sample any input data of interest.

[0039] At step 202 (see...) Figure 2 Data collection can be reduced in various ways. As examples, data collection can be reduced through extraction, compression, clustering, symbolic representation, or transformation, each of which demonstrates its own benefits and trade-offs. Figure 4A This is a diagram illustrating how data reduction is achieved through compression. The ADC 310 can be configured to oversample the output voltage Vout, making ripple information available (see...). Figure 3 The output voltage Vout can then be derived by memory circuit 318 and stored in reduced form. (e.g.) Figure 4A As illustrated in the diagram, waveform 400 shows an ADC sample from the control ADC 310, while waveform 402 shows a reduced data rate, where each data point 404 includes the average output voltage level (e.g., 1.0V) and the peak ripple voltage (e.g., 10mV) within the same PWM cycle. In this example, compression can provide a reduction of 5-10 times in data size.

[0040] Figure 4B This diagram illustrates how data can be reduced through symbolic representation. The ADC 310 can be controlled to sample the output voltage Vout at a high sampling rate to capture ripple behavior. The voltage level is then binned into various levels to preserve signal statistics and ensure that subsequent analysis of the data remains unchanged despite the reduction in symbolic data. The binning of levels can be performed a priori offline, online (e.g., within the system-on-chip, platform-wide), or a posteriori on a connected device.

[0041] Data reduction can be achieved by using the symbols assigned to the data. This can be done by reducing the bit width required for transmission / storage at each original sample point, or as... Figure 4B As shown, the reduction is provided by assigning a symbol to 420 only when certain voltage threshold levels 410 and 412 are detected to have been crossed. Figure 4B In the example, symbol A is assigned when the voltage falls below threshold 410. Symbol B is assigned when the voltage rises above threshold 410. Symbol C is assigned when the voltage rises above threshold 412. Each symbol can also be paired with a corresponding timestamp to help track the temporal behavior of the sampled signal.

[0042] Dynamic signals can be reduced in a similar way. Consider in Figure 5A The output voltage response shown is illustrated as a function of the sampling time, centered at zero volts for illustrative purposes. Figure 5B The rate of change of the output voltage as a function of the output voltage level is shown. As an example, Figure 5B The state space can be divided into three separate regions, A, B, and C. Region A represents the lowest rate of change with a deviation of approximately + / - 0.05 V from the nominal level. Region B represents the intermediate rate of change with a deviation of + / - 0.05 to 0.1 V from the nominal level. Region C represents the highest rate of change with a deviation of + / - 0.1 V or more from the nominal voltage level. Figure 5A This illustrates how the dynamic signal 500 can be reduced to a series of state-space data symbols (e.g., CCCCCCBBAAAAABBBBAA). In the example in Figure 5, point 510 is the reset state, point 512 is the first sampled symbol, and point 514 is the last sampled symbol.

[0043] Furthermore, additional reduction can be achieved by conveying relevant symbols only when the dynamic signal enters a specific region (e.g., data symbols can be timestamped only when traversing between two different state space regions). If needed, other signals such as output voltage and inductor current at the power stage can be combined on the phase plane and reduced using symbols in a similar manner (e.g., many signals and parameters can be combined to form a high-dimensional sample, which can then be reduced using any suitable technique).

[0044] Processing and computing elements at various points on the platform or in the cloud are configured to process and analyze data for the purpose of taking action based on the data. Such actions may involve classification, prediction, model recognition, etc., to provide user feedback. Such feedback may involve data visualization methods to improve the ability to take action as a result.

[0045] Applying machine learning to SOC power delivery can involve the application of machine learning algorithms such as anomaly detection, classification, prediction, and regression. These analytical algorithms can involve models modified based on the data. Such models can appropriately employ Bayesian methods, neural networks, deep belief networks, kernel functions, regression models, K-nearest neighbor methods, etc. In general, any type of statistical learning process that preserves statistical data can be implemented.

[0046] These machine learning processes provide technical improvements to the overall computing system. For example, anomaly detection enables adaptive voltage scaling for programmable integrated circuits; anomaly detection and regression improve DVFS by automatically learning a dynamic voltage and frequency scaling (DVFS) profile; anomaly detection allows for the identification of power-related attack vectors in security applications; load forecasting enables tighter power supply voltage tolerances, leading to higher frequency gradations; and classification improves power efficiency by identifying the correct power patterns based on analytics-driven power data. These different types of machine learning algorithms are not mutually exclusive and can all be implemented as part of a single system.

[0047] To illustrate an example of anomaly detection, consider, for example... Figure 1 Systems of the type 100 shown are examples of systems in which the integrated circuit device 104 has several high-current voltage rails (e.g., Vout, Vout', Vout'', etc.), each requiring precise voltage control. As an example, the power conversion system 106 can be configured to control each voltage rail to within a few percent of its nominal voltage level (e.g., within 1% or less, within 5% or less, within 10% or less, etc.) and to do so effectively under varying load conditions.

[0048] Each voltage rail (sometimes called a power supply terminal or power supply line) can exhibit its own individual statistical profile. Figure 6A The diagram illustrates a method for anomaly detection that can be used to address the problem of adaptive voltage positioning, a particularly challenging problem for FPGA-based systems where the active circuitry configuration on the FPGA is a priori unknown. Figure 6A As shown in the ramp waveform 600, the voltage (e.g., power supply voltage or other control voltage) can start from a default high voltage level Vhi and can decrease until the machine learning algorithm detects an anomaly. When an anomaly is detected at point 604 (i.e., when the power supply voltage hits the threshold voltage level Vanomaly), the power profile can be slightly increased by a predetermined margin Δ to eliminate the anomaly. The margin Δ can be 5 mV, 10 mV, 50 mV, 100 mV, or other suitable voltage amounts. By operating in this way, a safe reduced voltage level (Vanomaly + Δ) is determined for a given processing load, which saves energy while optimizing performance.

[0049] As an example, Figure 6BThe diagram illustrates the bivariate probability density functions of two power supply voltages, Vin1 and Vin2, for several workloads of two main types. Clustering or modeling identification techniques can be used to distinguish between different workloads. Cluster 620 represents the power distribution under normal operating conditions, while cluster 622 represents the power distribution under abnormal operating conditions. Therefore, machine learning algorithms can be adapted to detect when data switches from cluster 620 to cluster 622, thus signaling when vanomaly has been reached.

[0050] If needed, the anomaly detection algorithm can also be configured to detect anomalies in systems with a single power supply. For example... Figure 6C As shown, the first power distribution curve 610 represents the power density profile centered on voltage V1 under normal operating conditions, while the second power distribution curve 612 represents the power density profile centered on voltage V2 under abnormal operating conditions. In this example, the mean and standard deviation have been shifted between the two profiles 610 and 612. This is used to detect anomalies for 1-2 power rails. Figures 6A-6C The examples provided are merely illustrative. In general, these concepts can be extended to applications with three or more adjustable parameters to detect any suitable type of anomaly.

[0051] To illustrate an example of a classification application, consider the analysis of power supply data for the purpose of supervised classification in a neural network. Figure 7A The diagram illustrates the basic structure used to represent two categories or classes. For example... Figure 7A As shown, the neural network representation 700 may include 16 input nodes 702, which are connected to two hidden nodes 704 (e.g., a first hidden node 708-1 and a second hidden node 708-2). The two hidden nodes may be connected to two output nodes 706. The output nodes may have a bias, which is omitted for clarity.

[0052] The 16 inputs d0-d15 represent 2D input data for two power supply voltages (or currents, etc.) and can be converted to, for example, Figure 7B Bivariate histograms, such as histograms. Figure 7B A plot of the first power supply voltage Vin1 versus the second power supply voltage Vin2 was created. The frequencies at each voltage crossover were plotted for the two distinct classes (e.g., Class 1 and Class 2). Figure 7B As shown, class 1 can display a first two-dimensional (2D) profile distribution, while class 2 can display a second 2D power profile distribution.

[0053] Once trained, the neural network learns the appropriate weights to identify categories with an accuracy of up to 98% or higher. Figure 7C The weights learned at hidden layer 704 are shown. Figure 7C As shown, the first weight matrix 750-1 corresponds to class 1 weights for hidden node 708-1, while the second weight matrix 750-2 corresponds to class 2 weights for hidden node 708-2. Figure 7C In the diagram, white cells in the matrix can represent the absolute dominance of one class and the absence of another, while black cells can represent the complete absence of one class or the absolute dominance of the other. Gray cells can represent a point somewhere between two extremes, where both classes may be present or absent. For example, Figure 7B and 7C Part X in the diagram illustrates the dominant position of Class 1 relative to Class 2 within a specific voltage range. Conversely, Figure 7B and 7C Part Y in the diagram illustrates the dominant position of Class 2 over Class 1 within a specific voltage range. Part Z illustrates the slight advantage of Class 2 over Class 1 within the same voltage range.

[0054] As an example, class 1 could represent patterns for video data, while class 2 could represent patterns for audio data. As another example, class 1 could represent patterns for normal operation, while class 2 could represent patterns for abnormal operation. In other words, classification techniques can also achieve anomaly detection. These examples are merely illustrative. Generally, detection for each class can be matched with a specific weight matrix pattern for different types of data. If desired, radial basis function neural networks (RBFNNs) can also be used to classify different categories in a similar manner.

[0055] To illustrate an example of predictive applications, consider Figure 8A A feedforward DC-DC converter control system 800, where K represents the controller (e.g., Figure 3 (controller 314), G indicates power stage (e.g., Figure 3 In the power stage 808 (300), Zo represents the output impedance of the power stage 808, and Kd is the feedforward control transfer function. The system output can be fed back to circuit 802 via feedback path 812, subtracting the system output from the reference voltage level Vref to control controller stage 804. Circuit 806 can combine the output of controller stage 804 with feedforward controller stage 816 to control power stage 808. Disturbances such as load steps in the output current iout can affect the output Vout via the output impedance Zo. If the feedforward transfer function Kd is designed such that Kd*G equals Zo, perfect control can be achieved by accurately eliminating disturbances to iout.

[0056] According to an embodiment, the feedforward controller 816 can derive its input from a predictor circuit such as a load predictor 814. The predictor 814 can be configured to learn the characteristics of the load from features, such that any load step at iout can be anticipated and applied to the feedforward controller 816 before an iout disturbance event. If desired, the predictor stage 814 can optionally be incorporated into the feedforward controller 816. The feedforward controller 816 can be part of a feedforward control loop. In this way, the operating system 800 can achieve near-perfect control. The predictor 814 can be suitably implemented as a neural network (e.g., an artificial neural network or a radial basis function neural network) or suitably implemented through other machine learning means, such as using Naive Bayes, K-nearest neighbors, regression, logistic regression, and / or kernel methods (to name just a few), either alone or in combination, where bagging bootstrapping and tree methods can be employed.

[0057] Verified predictive input features that can be received at predictor 814 may include load current iout, load voltage (e.g., Vout or some other control voltage), and statistics of the load current / voltage (such as its mean, median, mode, standard deviation, etc.). In multi-supply systems, similar signals of related loads can serve as useful predictive input features. Operational loads related to the characteristics of the load or load operation (e.g., memory bus activity of an FPGA or program counter of a CPU) can also be useful predictive input features. Thus, predictive input features can be expected to be highly dimensional. In some cases, only a small number of predictive input features may be needed to accurately predict load characteristics (e.g., only one feature, only two features, only three features, only four features, etc.).

[0058] Figure 8A This is a schematic diagram of the structure of the control system 800. Figure 8B The diagram illustrates the actual hardware setup. For example... Figure 8B As shown, predictor 814 can influence the mode of power conversion system 106. Predictor 814 can be configured to estimate the appropriate mode of power conversion system 106 for a predicted event. Such a mode can be an aspect of the power stage (e.g., see...). Figure 3 The power stage 300 in the multiphase converter includes modes such as continuous current mode (CCM), discontinuous current mode (DCM), a mode for selecting the number of active phases in the multiphase converter, and a mode for distributing the current through each phase according to the efficiency target / goal.

[0059] Figure 9A It is used to represent association Figure 8A and 8B A diagram illustrating a feedforward neural network 900 of the type of predictive control system described. (See diagram for example.) Figure 9AAs shown, the neural network 900 may include one hundred input nodes 902, a ten-node intermediate layer 904, and two output nodes 906. The one hundred input nodes d0-d99 represent data fed from a sliding window of data points, such as current measurements on a voltage rail. Similar to... Figure 7A The neural network 700, with its two input nodes 906, can also be used to classify data into one of two classes, c1 or c2. In an example where the prediction system is configured to predict load spikes / peaks, the first class is associated with currents that cross or exceed a predetermined threshold within the next three time steps (as an example), while the second class is associated with predicted loading behavior that will not cross the predetermined threshold.

[0060] Figure 9B Simulation results for a neural network prediction system 800 are shown. Waveform 910 represents (e.g., the output current iout) zero-centered current measurement data. Therefore, Figure 9B The plot illustrates the deviation from the nominal current level. The predictor can be configured to predict when waveform 910 exceeds a predetermined threshold D_threshold. Dashed line 912 indicates when the current actually crosses D_threshold within the next three time steps. Line 914 represents the output of the neural network, which predicts when the current will cross D_threshold within the next three time step windows. Figure 9B As shown, the prediction matches the actual behavior very closely. The prediction system can achieve over 99% accuracy on the training set. Near a step of 740 on the x-axis, prediction 914 does not match actual behavior 912, meaning the threshold was crossed but did not actually occur. In this case, line 914 will immediately fall back below after recognizing the error.

[0061] It can take action in response to the predicted event crossing a threshold. For example, predictor 814 can initiate actions such as combining... Figure 8B The described mode switching (e.g., configuring the power conversion system to continuous current mode, discontinuous current mode, mode for selecting the number of active phases in a multiphase converter, mode for allocating current through each phase according to an efficiency object / target, etc.) or efficiently supplying additional current within three time steps prior to the expected event to mitigate voltage deviations at the power supply rail.

[0062] As mentioned above, at least... Figure 8A As described, predictions can be used to inform the feedforward controller 816 to improve control. Figure 10The output voltage deviation in response to a load current step at the CPU or FPGA is shown. Before time t*, load step prediction and feedforward control are active, and the Vout deviation is fairly well controlled at times t1-t3. At time t*, load step prediction and feedforward control are turned off. At times t4-t6, the Vout deviation is quite significant. Therefore, the use of a predictor 814 with a feedforward controller 816 can provide reduced voltage disturbances at any number of power supply rails (as an example).

[0063] The example in Figure 8-10, which shows how machine learning can be applied to predict voltage and current behavior, is merely illustrative. If needed, machine learning algorithms can be used to monitor any verified parameters in the overall system and to anticipate and mitigate any type of unwanted dynamic events.

[0064] Although the methods of operation are described in a specific order, it should be understood that other operations may be performed between the described operations, the described operations may be adjusted so that they occur at slightly different times, or the described operations may be distributed in a system that allows processing operations to occur at various intervals related to the processing, as long as the processing of the overlay operations is performed in the desired manner.

[0065] Example : The following examples relate to other embodiments.

[0066] Example 1 is a system comprising: an integrated circuit; a power stage configured to receive an input voltage and generate a corresponding output voltage to power the integrated circuit; control circuitry configured to extract and process data collected from the power stage; and a power data processor configured to analyze the data and, based on the analysis of the data, guide the control circuitry to take appropriate actions.

[0067] Example 2 is the system of Example 1, wherein the power data processor is optionally configured to use machine learning algorithms to analyze the data.

[0068] Example 3 is the system of Example 2, wherein the power data processor is optionally further configured to use machine learning algorithms to detect anomalies within the system.

[0069] Example 4 is the system of Example 2, wherein the power data processor is optionally further configured to use machine learning algorithms to classify the data into different classes.

[0070] Example 5 is a system of Example 2, wherein the power data processor is optionally further configured to use machine learning algorithms to predict loading events at the integrated circuit.

[0071] Example 6 is a system of any of Examples 1-5, wherein the control circuitry optionally includes reduction circuitry configured to reduce data via compression or symbolic representation.

[0072] Example 7 is a system of Example 6, wherein the control circuitry optionally further includes: a controller configured to control a power stage; a power manager coupled to the controller and a power data processor; and memory circuitry configured to store reduced data.

[0073] Example 8 is a system of any of Examples 1-5, and optionally further includes: a predictive circuit configured to anticipate loading events and control a power stage, wherein the predictor is part of a feedforward control system.

[0074] Example 9 is a system of any of Examples 1-5, wherein the power data processor includes a local analysis block, and the system optionally further includes a remote analysis block configured to remotely analyze data and guide the control circuitry to take appropriate actions based on the analysis of the data.

[0075] Example 10 is a method for operating a power analysis system, the method comprising: using a power conversion system to power an integrated circuit and collecting data while powering the integrated circuit; using the power conversion system to process the collected data; using a power data processor to analyze the processed data by applying machine learning to the processed data; and using the power data processor to guide the power conversion system to take appropriate actions based on the analysis of the data.

[0076] Example 11 is a method of Example 10, wherein the collected data is processed using a power conversion system, optionally including data reduction.

[0077] Example 12 is a method of Example 10, wherein applying machine learning to the processed data may optionally include performing anomaly detection.

[0078] Example 13 is the method of Example 10, wherein applying machine learning to processed data may optionally include classifying the data into different classes.

[0079] Example 14 is the method of Example 10, wherein applying machine learning to the processed data may optionally include predicting when a parameter crosses a predetermined threshold.

[0080] Example 15 is a method of any of Examples 10-14, wherein applying machine learning to processed data may optionally include: performing unsupervised machine learning at a power data processor; and performing supervised machine learning in the cloud, wherein supervised machine learning is more computationally intensive than unsupervised machine learning.

[0081] Example 16 is an apparatus comprising: an integrated circuit; a power conversion system configured to supply voltage to the integrated circuit for output power and further configured to collect data; and an analysis block configured to use machine learning to analyze the collected data and take action on the collected data.

[0082] Example 17 is an apparatus of Example 16, wherein the analysis block optionally uses machine learning to enable dynamic voltage scaling at the integrated circuit by performing anomaly detection on the collected data.

[0083] Example 18 is an apparatus of Example 16, wherein the analysis block optionally uses machine learning to identify attack vectors at the integrated circuit by performing anomaly detection on the collected data.

[0084] Example 19 is an apparatus of Example 16, wherein the analysis block optionally uses machine learning to enable tighter voltage tolerance for the power supply voltage by performing load prediction based on the collected data.

[0085] Example 20 is an apparatus of Example 16, wherein the analysis block optionally uses machine learning to identify appropriate power modes for the power conversion system by classifying the collected data into different classes.

[0086] Example 21 is a power analysis system comprising: means for powering an integrated circuit and collecting data while powering the integrated circuit; means for processing the collected data; means for analyzing the processed data by applying machine learning to the processed data; and means for taking appropriate action based on the analysis of the data.

[0087] Example 22 is a power analysis system of Example 21, wherein the means for processing the collected data may optionally include means for reducing the data.

[0088] Example 23 is a power analysis system of Example 21, wherein the means for applying machine learning to the processed data may optionally include means for performing anomaly detection.

[0089] Example 24 is a power analysis system of Example 21, wherein the means for applying machine learning to the processed data optionally includes means for classifying the data into different classes.

[0090] Example 25 is a power analysis system of Example 21, wherein the means for applying machine learning to the processed data optionally includes means for predicting when a parameter crosses a predetermined threshold.

[0091] For example, all the optional features of the apparatus described herein can be implemented with respect to the methods or processes described herein. The foregoing is merely illustrative of the principles of this disclosure, and various modifications can be made by those skilled in the art. The foregoing embodiments can be implemented individually or in any combination.

Claims

1. A system for power analysis, comprising: integrated circuit; A power stage, which is configured to receive an input voltage and generate a corresponding output voltage to power the integrated circuit; Control circuitry configured to extract and process data collected from the power stage while powering the integrated circuit; The control circuitry includes a data conversion circuit configured to oversample the output voltage to collect the data, making ripple information available. as well as A power data processor is configured to use machine learning algorithms to analyze ripple information to predict load steps and guide the control circuitry to take appropriate actions based on the predicted load steps.

2. The system of claim 1, wherein the power data processor is further configured to use a machine learning algorithm to detect anomalies within the system.

3. The system of claim 1, wherein the power data processor is further configured to use a machine learning algorithm to classify the data into different classes.

4. The system of claim 1, wherein the power data processor is further configured to use a machine learning algorithm to predict loading events at the integrated circuit.

5. The system of claim 1, wherein the control circuit further comprises: The controller is configured to control the power stage; as well as The power manager is coupled to the controller and the power data processor.

6. The system according to any one of claims 1-5, further comprising: The predictive circuit is configured to anticipate loading events and control the power stage, wherein the predictor is part of the feedforward control system.

7. The system according to any one of claims 1-5, wherein the power data processor includes a local analysis block, and the system further includes: A remote analysis block is configured to remotely analyze data and guide the control circuit to take appropriate actions based on the analysis of the data.

8. A method for operating a power analysis system, the method comprising: Using a power conversion system to power an integrated circuit and collecting data while powering the integrated circuit, wherein collecting data while powering the integrated circuit includes oversampling the output voltage of the power conversion system to collect the data, making ripple information available; The collected data is processed using a power conversion system; By applying machine learning to ripple information, ripple information can be analyzed to predict load steps; as well as The power data processor guides the power conversion system to take appropriate actions based on predicted load steps.

9. The method of claim 8, wherein applying machine learning to the ripple information includes performing anomaly detection.

10. The method of claim 8, wherein applying machine learning to the ripple information includes classifying the data into different classes.

11. The method of claim 8, wherein applying machine learning to the ripple information includes predicting when a parameter crosses a predetermined threshold.

12. The method according to any one of claims 8-11, wherein applying machine learning to the ripple information comprises: Unsupervised machine learning is performed at the power data processor; as well as Supervised machine learning is performed in the cloud, where it is more computationally intensive than unsupervised machine learning.

13. An apparatus for power analysis, comprising: integrated circuit; A power conversion system configured to supply voltage to an integrated circuit for output power and further configured to collect data while powering the integrated circuit; A control circuit configured to extract and process the collected data, wherein the control circuit includes a data conversion circuit configured to oversample the output power supply voltage to collect the data, making ripple information available. as well as The analysis block is configured to use machine learning to analyze ripple information to predict load steps and take appropriate actions based on the predicted load steps.

14. The apparatus of claim 13, wherein the analysis block uses machine learning to enable dynamic voltage scaling at the integrated circuit by performing anomaly detection on the collected data.

15. The apparatus of claim 13, wherein the analysis block uses machine learning to identify attack vectors at the integrated circuit by performing anomaly detection on the collected data.

16. The apparatus of claim 13, wherein the analysis block uses machine learning to enable tighter voltage tolerances for the power supply voltage by performing load prediction based on collected data.

17. The apparatus of claim 13, wherein the analysis block uses machine learning to identify suitable power modes for the power conversion system by classifying the collected data into different classes.

18. A power analysis system, comprising: A means for powering an integrated circuit and collecting data while powering the integrated circuit, wherein the means for powering the integrated circuit and collecting data while powering the integrated circuit includes means for oversampling the output voltage of a power analysis system to collect the data so that ripple information is available. A device for processing collected data; A device for analyzing processed data by applying machine learning to ripple information to predict load steps. as well as A device for taking appropriate action based on a predicted load step.

19. The power analysis system of claim 18, wherein the means for applying machine learning to ripple information includes means for performing anomaly detection.

20. The power analysis system of claim 18, wherein the means for applying machine learning to ripple information includes means for classifying the data into different classes.

21. The power analysis system of claim 18, wherein the means for applying machine learning to ripple information includes means for predicting when a parameter crosses a predetermined threshold.

Images