SYSTEMS, DEVICES AND METHODS FOR PERFORMANCE ASSESSMENT

DE102024101476A1Pending Publication Date: 2025-07-24INFINEON TECHNOLOGIES AG
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
DE102024101476
Authority / Receiving Office
DE · DE
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-01-18
Publication Date
2025-07-24

Smart Images

  • Figure 00000022_0000
    Figure 00000022_0000
  • Figure 00000024_0000
    Figure 00000024_0000
  • Figure 00000025_0000
    Figure 00000025_0000
Patent Text Reader

Abstract

A hardware power estimator (HPE) configured to provide a power estimate for an electronic device includes a scaling circuit configured to receive an input and produce a scaled nonlinear output of the input, wherein the scaling circuit applies an approximation of an activation function to the input. The HPE further includes multiply-add circuitry (MADD) configured to produce an output indicative of a mathematical power estimate of the electronic device using the output of the scaling. The input to the scaling circuit includes data values representing state information or activity information of one or more components or intellectual property (IP) blocks of the electronic device.
Need to check novelty before this filing date? Find Prior Art

Description

Technical area

[0001] Various embodiments generally relate to the performance estimation of electronic devices. background

[0002] Managing power consumption in embedded devices is critical for extending battery life and minimizing a system's environmental footprint. As this aspect of system design becomes increasingly important, estimating system-level power consumption becomes increasingly challenging. This complexity applies to a wide range of scenarios, from IoT applications with numerous embedded devices to automotive applications such as engine control modules and powertrain modules, as well as industrial equipment equipped with sensors. For example, automotive microcontrollers incorporate accelerators whose dynamic power consumption varies considerably depending on the configuration and data pipeline.

[0003] There has been a paradigm shift toward power-efficient methods for integrating a system-on-chip (SoC). Power-aware designs and technologies based on accurate pre-silicon power consumption are widely used in modern SoCs. These techniques must also be applied at the silicon / hardware level to manage system power more efficiently. Estimating dynamic power consumption during runtime enables faster and better power management systems (such as DVC, DVFS, etc.) to be used in a given system. Furthermore, estimating power consumption for a complete system—the sum of all individual SoCs / ASICs, with real-time operating systems / software on complex hardware—leads to global optimization of power / energy consumption.

[0004] Currently, estimating the average power consumption of microcontrollers and ASICs for specific use cases requires gate-level RTL simulations or time-based activity simulations. Simulation time constraints and the complexity of converting user code to vectors prevent rapid estimation of average power consumption for iterative performance optimization of complex use cases. Furthermore, implementation limitations make it impractical to use the same techniques to realize hardware-based power estimators. Therefore, modern SoCs use power management IPs that aggregate various IP logic states to achieve the overall system power state (deep sleep, sleep, idle, standby, etc.). These states are then used by the power management controller to optimize overall performance.

[0005] Microcontrollers generally contain multiple processing units, different types of memory cells, and multiple instances of peripherals. Each of the IPs, in turn, supports various configurations. For example, CPU cores can operate at different frequencies, with or without lock-step between cores, and with different activity factors (IPC rates). IO interfaces such as CAN support different baud rates, different power management states, and so on. Theoretically, there are several million combinations of unique configurations for such microcontrollers. Furthermore, a typical application dynamically switches microcontrollers between some of these configurations. For example, accelerators used in ADAS or visual computing systems support multiple configurations with different data and bus load sizes.This further increases the complexity of estimating or modeling the dynamic performance of such systems. In special cases, such as microcontrollers, the power consumption of digital logic depends on its configuration, unlike general-purpose microprocessors, where the size and complexity of the software can affect power consumption. Currently, estimating the average dynamic power consumption of such microcontroller user software requires either on-silicon power measurement or gate-level silicon simulation (in the pre-silicon phase). Brief description of the drawings

[0006] In the drawings, like reference characters generally refer to the same parts throughout the several views. The drawings are not necessarily to scale; rather, emphasis is placed upon illustrating the principles of the invention. In the following description, various embodiments of the invention are described with reference to the following drawings, in which: Fig. 1A is a diagram illustrating a microcontroller unit according to at least one example embodiment of the present disclosure. Fig. Figure 1B is a table of configurations for a microcontroller unit. Fig. 2 is a block diagram of a hardware performance estimator according to one or more aspects of the present disclosure. Fig. 3 shows an example of a scaling circuit according to one or more aspects of the present disclosure. Fig. 4 shows diagrams of activation functions according to one or more aspects of the present disclosure. The Fig. 5-7 show examples of multiply-add (MADD) or multiply-accumulate (MAC) circuits according to aspects of the present disclosure. Fig. 8 shows a hardware power estimation network according to one or more aspects of the present disclosure. The Fig. 9 and Fig. 10 each depict an exemplary flowchart and environment for determining the weights or weighted set of values for HPEs in accordance with aspects of the present disclosure. The Fig. 11 and Fig. 12 show methods according to aspects of the present disclosure. Description

[0007] The following detailed description refers to the accompanying drawings which show, by way of illustration, specific details and embodiments of the invention.

[0008] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any embodiment or design described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments or designs.

[0009] The words "plurality" and "multiple" in the description or claims expressly refer to a set greater than one. The terms "group (of)", "set (of), "collection (of), "series (of), "sequence (of), "grouping (of)," etc., and similar terms in the description or claims, refer to a set equal to or greater than one, that is, one or more. Any term expressed in the plural that is not expressly stated to be "plural" or "multiple" also refers to a set equal to or greater than one. The terms "proper subset", "reduced subset", and "smaller subset" refer to a subset of a set that is not equal to the set, that is, a subset of a set that contains fewer elements than the set.

[0010] The terms “at least one” and “one or more” may be understood to include a numerical quantity greater than or equal to one (e.g., one, two, three, four, [...], etc.).

[0011] Unless otherwise indicated, the use of the ordinal adjectives "first", "second", "third", etc., to describe a common object merely means that different instances of similar objects are meant, and is not intended to imply that the objects so described must be in any particular order, be it temporal, spatial, rank, or in any other way.

[0012] The term "data," as used herein, may be understood to include, for example, information in any suitable analog or digital form, provided as a file, a portion of a file, a set of files, a signal or stream, a portion of a signal or stream, a set of signals or streams, and the like. Furthermore, the term "data" may also be used to mean a reference to information, e.g., in the form of a pointer. However, the term "data" is not limited to the aforementioned examples, but may take various forms and represent any type of information as understood in the art.

[0013] The term "processor" or "controller," as used herein, can be understood as any type of unit that enables the processing of data, signals, etc. The data, signals, etc. can be processed according to one or more specific functions performed by the processor or controller.

[0014] A processor or controller may be or include an analog circuit, a digital circuit, a mixed-signal circuit, a logic circuit, a processor, a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a field programmable gate array (FPGA), an integrated circuit, an application-specific integrated circuit (ASIC), etc., or any combination thereof. Any other implementation of the respective functions, described in more detail below, may also be considered a processor, controller, or logic circuit.It is understood that two (or more) of the processors, controllers, or logic circuits described herein may be implemented as a single unit with equivalent functionality or the like, and conversely, that any individual processor, controller, or logic circuit described herein may be implemented as two (or more) separate units with equivalent functionality or the like.

[0015] A "circuit" here refers to any type of logic-implementing unit, which may include specialized hardware or a processor for executing software. Thus, a circuit may be an analog circuit, a digital circuit, a mixed-signal circuit, a logic circuit, a processor, a microprocessor, a signal processor, a central processing unit ("CPU"), a graphics processing unit ("GPU"), a neuromorphic computing unit (NCU), a digital signal processor ("DSP"), a field-programmable gate array ("FPGA"), an integrated circuit, an application-specific integrated circuit ("ASIC"), etc., or any combination thereof.Any other implementation of the respective functions, which is described in more detail below, may also be understood as a "circuit." It is understood that two (or more) of the circuits described herein may be implemented as a single circuit with substantially the same functionality. Conversely, any individual circuit described herein may be implemented as two (or more) separate circuits with substantially the same functionality. Furthermore, the term "circuit" may refer to two or more circuits that together form a single circuit.

[0016] As used herein, the terms "module," "component," "system," "circuit," "element," "interface," "slice," "circuitry," etc., refer to a set of one or more electronic components, a computer-related unit, hardware, software (e.g., in execution), and / or firmware. A circuit or similar term can be, for example, a processor, a process running on a processor, a controller, an object, an executable program, a storage device, and / or a computer with a processing device. To illustrate, an application running on a server and the server can also be a circuit. One or more circuits can be located within the same circuit, and the circuits can be localized on one computer and / or distributed across two or more computers.Here, a set of elements or a set of other circuits can be described, where the term “set” can be interpreted as “one or more”.

[0017] As used here, a "signal" can be transmitted or passed through a signal chain where the signal is processed to change properties such as phase, amplitude, frequency, etc. The signal can be said to be the same signal even if the properties are adjusted. As long as a signal encodes the same information, it can generally be considered the same signal.

[0018] A signal "indicative" of a value or other information can be a digital or analog signal that encodes or otherwise conveys the value or other information in a manner that can be decoded by a component that receives the signal and / or triggers an appropriate action. The signal can be stored or buffered in a computer-readable storage medium before being received by the receiving component. The receiving component can retrieve the signal from the storage medium. Furthermore, a "value" that is "indicative" of a quantity, state, or parameter can be physically embodied as a digital signal, an analog signal, or as stored bits that encode or otherwise convey the value.

[0019] When an element is described as being "connected" or "coupled" to another element, it may be physically connected or coupled to the other element so that current and / or electromagnetic radiation (e.g., a signal) can flow along a conductive path formed by the elements. Conductive, inductive, or capacitive elements may be present between the element and the other element when the elements are described as being coupled or connected to each other. Furthermore, when one element is coupled or connected to another, it may be capable of inducing a flow of voltage or current, or the propagation of an electromagnetic wave, in the other element without physical contact or intervening components.When a voltage, current, or signal is "applied" to an element, the voltage, current, or signal can be conducted to the element through a physical connection or through capacitive, electromagnetic, or inductive coupling that does not involve a physical connection.

[0020] The term "memory" is defined herein as a non-transitory, computer-readable medium in which data or information can be stored for retrieval. When reference is made to "memory," it can mean volatile or non-volatile memory, including random-access memory (RAM), read-only memory (ROM), flash memory, solid-state memory, magnetic tape, hard disk drive, optical drive, etc., or any combination thereof. In addition, registers, shift registers, processor registers, data buffers, etc. are also considered to be part of the memory. A single component referred to as "memory" or "a memory" may consist of more than one different type of memory and may therefore refer to an entire component comprising one or more types of memory.Each individual memory component can be divided into several collectively equivalent memory components, and vice versa. Furthermore, while memory may be depicted separately from one or more other components (as in the drawings), it may also be integrated with other components, for example, on a common integrated chip or a controller with embedded memory.

[0021] The term “software” refers to any type of executable instructions, including firmware.

[0022] Unless expressly stated otherwise, terms used herein such as “processing,” “calculating,” “computing,” “determining,” “representing,” “displaying,” or the like may refer to actions or processes of a machine (e.g., a computer / processor / etc.) that manipulates or transforms data represented as physical (e.g., electronic, magnetic, or optical) quantities in one or more memories (e.g., volatile memory, non-volatile memory, or a combination thereof), registers, or other components of the machine that receive, store, transmit, or display information.

[0023] Example embodiments of the present disclosure may be implemented by one or more computers (or computing devices) that retrieve and execute computer-executable instructions recorded on a storage medium (e.g., a non-transitory, computer-readable storage medium) to perform the functions of one or more embodiments of the disclosure described herein. The computer(s) may include one or more central processing units (CPUs), a microprocessor unit (MPUs), or other circuitry, and may include a network of separate computers or separate computer processors. The computer-executable instructions may be provided to the computer, for example, via a network or a non-transitory, computer-readable storage medium.The storage medium may include, for example, one or more hard disks, random access memory (RAM), read-only memory (ROM), distributed computing system memory, an optical drive (such as a compact disc (CD), a digital versatile disc (DVD), or a Blu-ray disc (BD)), a flash memory device, a memory card, or the like. For illustrative purposes, specific details and embodiments in which the invention may be practiced are described.

[0024] Unless otherwise indicated, the use of the ordinal adjectives "first", "second", "third", etc., to describe a common object merely means that different instances of similar objects are meant, and is not intended to imply that the objects so described must be in any particular order, be it temporal, spatial, rank, or in any other way.

[0025] It should be understood that other embodiments may be utilized, and structural or logical changes may be made without departing from the scope of the present disclosure. The following detailed description is therefore not to be considered limiting. For purposes of this disclosure, the phrase "A and / or B" means (A), (B), or (A and B). For purposes of this disclosure, the phrase "A, B, and / or C" means (A), (B), (C), (A and B), (A and C), (B and C), or (A, B and C). Reference in this disclosure to "an embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment.The phrases "in one embodiment" or "in an embodiment" do not necessarily all refer to the same embodiment. The phrases "for example," "in an example," or "in some examples" do not necessarily all refer to the same example.

[0026] Fig. 1A shows a block diagram of a microcontroller or microcontroller unit (MCU) 100 according to one or more example embodiments of the present disclosure.

[0027] The MCU 100 or microcontroller 100 includes one or more cores 10, a memory or memory circuit 20, one or more intellectual property (IP) blocks 30, and one or more other applications 40. Other components may be included but are not shown. Connections between the components of the microcontroller 10 may be assumed but are not shown in Fig. 1 is not shown. The MCU 100 can be implemented as part of an SoC.

[0028] As in Fig. 1A, the one or more cores 10 may be processor cores or central processing unit (CPU) cores. The one or more cores 10 may perform one or more operations by executing program instructions or software. The applications 40 may, for example, be in the form of instructions to be executed by one or more cores. The instructions or software described herein may be stored or located on a (non-transitory) computer-readable storage medium within the microcontroller 10 or SoC, or may otherwise be accessible to the cores 10. The MCU 100 may, for example, include a memory circuit 20 including a computer storage medium for storing instructions to be executed by the cores 10.

[0029] An intellectual property (IP) block 30 may refer to a reusable logic, circuit, software, or chip layout element. An IP block or IP may support multiple functions, which in some cases are implemented by one or more devices within the IP block and / or at least partially by one or more processor cores.

[0030] In addition, one or more peripheral devices 60 can be connected to the MCU 100 in an operational manner.

[0031] In general, the power consumption of electronic devices, including digital logic components, includes power dissipation and dynamic components. Power dissipation is generally a function of manufacturing process parameters (threshold voltage, mobility, etc.) and voltage. However, dynamic power depends on the switching activity of the logic gates in the integrated logic circuit. Dynamic power consumption can be described as follows: Pdyn=V2∑i⋅fiαiCi Where: α denotes the activity of a particular node / network of the logic circuit, C stands for the effective capacity, V stands for voltage, and f or feff denotes the effective frequency.

[0032] Modern silicon designs typically use several similar cells, called standard cells, for the entire logic design. For example, many state machines would be implemented with specific types of flip-flops, each consisting of standard logic gates with a specific threshold voltage. For calculating dynamic performance, for example, each of the standard cells can be abstracted. In this case, the total dynamic performance can be described as follows: Pdyn=V2∑p⋅feff,pNpCeff,p Where: N denotes the total number of cells of a particular type (normal threshold voltage, high threshold voltage, NOR gate, etc.)

[0033] The two equations above for calculating dynamic power can be used together for different parts of digital logic. For example, equation (1) can be used to estimate the power consumption of IPs and subsystems with high operating frequency, while equation (2) can be used for other logic. Accordingly, the total dynamic power can be calculated as follows: Pdyn=V2∑i⋅fiαici+V2∑p⋅feff,pNpCeff,p

[0034] Modern microcontrollers or microprocessors can contain several million logic gates. Calculating the dynamic performance for these millions of logic nodes / networks increases the computational complexity. If a finite number of use cases (and thus the chip configuration) is known, the configuration-based dynamic performance for a specific logic cluster can be correlated with the aggregate gate-based dynamic performance for another logic set / cluster. Consequently, the total dynamic performance can be expressed as follows: Pdyn=V2∑p⋅[feff,pNpCeff,p+∑ifi,pαi,pCi,p]+V2∑i[fiαiCi+∑p˙feff,p,iNp,iCeff,p,i]

[0035] Applying this method recursively would significantly reduce the computational effort required to calculate dynamic power. These nested equations can then be generally represented as: vn=∑r⋅Yr∑q⋅xq,r∑k⋅AkWk,q,r

[0036] Equation (5) shows that the dynamic power calculation can be represented as a series of multiplication and addition operations. Accordingly, the dynamic power can be represented as a network of operations, where, for example, each node is a MADD circuit. The inputs A k can be either the activities α or the number of cells, N i The weight set (W, X, Y) would then effectively represent the remaining terms such as capacitance and voltage in equation (4). The activities of specific nodes and the number of logic gates in a predefined cell type are inputted.

[0037] Fig. Figure 1B shows a table 150 with various configurations, e.g., for an MCU or a SOC. Configurations 1 to 4 (CFG1-CFG4) can apply to a specific IP of an MCU / SOC. Configurations CFG1-CFG4 can each perform a fast Fourier transform (FFT) with different resolutions and use 1 MB of memory. Table 150 shows the associated power estimates for each configuration, namely the active or dynamic power (P dyn ), the active load step current, and the active load step duration. In one or more examples of the present disclosure, the power estimation values may be provided by the hardware in real time.

[0038] Fig. 2 shows a block diagram of at least one example of a hardware performance estimator (HPE) 200 according to one or more aspects of the present disclosure.

[0039] The HPE 200 can be implemented in hardware, for example, with hardware components. Furthermore, the hardware components can also be hard-wired. In other words, the HPE 200 can contain hardware components or circuitry configured to perform one or more functions without a processor and without the use of software code or executable instructions.

[0040] The HPE 200 can be implemented as part of the microcontroller / SoC or external to it. The HPE 200 can also be formed as part of an SoC.

[0041] The HPE 200 is configured to produce an output 250 that provides an approximation or estimate of the power for an electronic device, such as a microcontroller or MCU (see, for example, MCU 100 in Fig. 1). In some cases, the output 250 may display or indicate one or more power-related values, such as current, voltage, or power (watts).

[0042] In another example, output 250 may serve as an indicator pointing to one of several possible performance outcomes or a range of possible outcomes, such as an entry in a table or list.

[0043] In addition, in at least one example, the HPE 200 can provide an estimate or approximation of real-time power consumption.

[0044] As in Fig. 2, the HPE 200 includes a scaling circuit 230 and a multiply-add (MADD) circuit or circuitry 240.

[0045] In one or more cases, the scaling circuit 230 may be configured to receive or acquire an input 210 and, in response, generate a scaled, nonlinear output 235 of the input 210. For example, the scaling circuit 230 may apply an activation function or an approximation of an activation function to the received input. Activation functions are mathematical functions that can introduce nonlinearity into various processes. For example, they can transform or map input data such that the input data becomes nonlinear in a meaningful way. Well-known activation functions include the sigmoid function, the hyperbolic tangent function (tanh), the rectified linear unit function (ReLU or Relumax), and the exponential linear unit (ELU).

[0046] For example, the tanh function maps input values to a range between -1 and 1. The ReLU function outputs the input value directly if the input is positive and zero if the input is negative, thereby introducing nonlinearity and sparsity.

[0047] Input 210 to the HPE 200 may be a signal (e.g., a digital signal) containing data values representing state or activity information of one or more components or intellectual property (IP) blocks of the electronic device. For example, various types of signals or data may be used as inputs because they indicate activity or state information about one or more IP blocks or functional components of the MCU 10. Input 210 may, for example, contain one or more of the following signals or the data from one or more of the following signals: - an IP release or deactivation signal, - an IP configuration register or a logical signal (IP activities can be stored or captured as data in configuration registers (CFG)) - a clock configuration register or a logic signal, - an IP performance monitoring signal, - a signal corresponding to bus use, - a hardware state transition signal of a finite state machine or - an energy state signal.

[0048] These signals can contain data indicating activities or states of the IP blocks, components (functional blocks) of an MCU. That is, these signals can contain data or values indicating

[0049] Fig. 3 shows a block diagram of a scaling circuit 300 according to aspects of the present disclosure. The scaling circuit 230 of Fig. 2 can be realized or implemented in a manner similar to the configuration of scaling circuit 300. As shown in the figure, there may be various hardware components within circuit 300 dedicated to performing specific functions. In other words, scaling circuit 300 may include tasks such as performing shift operations, multiplications, addition operations, and summations. Furthermore, scaling circuit 300 includes a comparator circuit dedicated to comparing values.

[0050] In at least one example, the components of scaling circuit 300 may be hard-wired and / or implemented with logic devices without using instructions executed by a processor.

[0051] Scaling circuit 300 may perform an approximation of an activation function, such as the tanh activation function or the Relu activation function, to the received input. That is, scaling circuit 300 may scale or transform received inputs according to an approximation of an activation function.

[0052] The scaling circuit 300 of the example in Fig. 3 can perform the approximation of the activation function to the input 310 piecewise. For example, starting from a received input, several processes or operations can be performed separately or in parallel on the input to generate several outputs. These outputs (outputs 370a-370g in Fig. 3) generated by the individual piecewise operations or processes are summed by a summing circuit 380 to produce an overall or final output 390 for the scaling circuit 300.

[0053] To generate a piecewise output, the scaling circuit 300 may perform any combination of one or more of the following functions or operations: a multiply-add operation, a shift operation, and a comparator operation. Since the input 310 to the scaling circuit may be digital, binary operations such as shift operations may be performed. Shifting a digital value left can effectively be thought of as multiplying by two (2). Similarly, shifting right can be thought of as dividing by 2. In such operations, the least significant bits are assumed to be on the right side, while the most significant digits are found on the left side.

[0054] As in Fig. 3, a piecewise output 370a may be generated or manufactured by performing a multiply-add operation (using a multiply-add circuit 320) using the input 310 (first operand) and using the input 310 after being shifted left twice (using the left shift circuits 325 and 335) (second operand).

[0055] Similarly, output 370b may be generated by performing a multiply-add operation (via MADD circuit 345) from input 310 (first operand) and from input 310 after it has been shifted left three times (using left shift circuits 325, 335, and 340) (second operand).

[0056] Another piecewise output 370c can be generated simply by shifting the input 310 left three times using the left shifters 325, 335, and 340. Another piecewise output 370d can be generated by shifting the input 310 left twice using the left shifters 325 and 335. The piecewise output 370e can be generated by shifting the input 310 right once using the left shifter 325.

[0057] In addition, a piecewise output 370f can only be the input 310 itself, without any operations being performed on it.

[0058] Furthermore, another piecewise output 370g may be generated using a comparator circuit 350. The comparator circuit 350 compares the input 310 with a predefined value or threshold to generate the output 370g.

[0059] The piecewise outputs 370a-370g may then be summed by summing circuit 380 to produce output 390. Depending on the predefined threshold, output 390 may approximate the result of an activation function, such as a hyperbolic tangent (tanh) or a Relu activation function, applied to input 310.

[0060] Fig. 4 shows a graphical representation 400 with a plot 410 of an ideal hyperbolic tangent and a plot 420 representing an approximation to the hyperbolic tangent function, for example, achieved by a circuit such as the scaling circuit 300.

[0061] After the input 210 has been scaled or transformed by the scaling circuit 230, the MADD circuit 240 operates on this scaled or transformed input. The MADD circuit 240 may include one or more stages.

[0062] Fig. 5 shows an example of a single-stage multiply-add (MADD) or multiply-accumulate (MAC) circuit 500. Fig. Figure 6 shows an example of a two-stage MADD circuit 600. Fig. 7 shows a three-stage MADD circuit 700. The Fig. The MADD circuit 240 shown in Figure 2 can be implemented in a manner very similar to the configurations of the MADD circuits 500-700.

[0063] The MADD circuits 500-700 can be configured to perform multiplication and addition / accumulation operations to output an estimate of power consumption. The MADD circuits 500-700 are capable of performing or calculating linear equations to produce an output reflecting an estimate of dynamic power consumption.

[0064] In at least one example, the MADD circuits 500-700 may be hard-wired and / or implemented with logic devices (e.g., gates) without using instructions executed by a processor.

[0065] As previously mentioned, the single-stage MADD circuit 500 is configured to perform the multiplication and accumulation of two sets of numbers. In other words, the MADD circuit 500 can be considered a simple two-input digital circuit for performing a 4-bit multiplication in a single clock cycle.

[0066] The MADD circuit 500 includes an input or input layer 510 coupled to an output or output layer 550 via a single stage 530. In the example of Fig. 5, the input or input layer 510 may have or contain two (different) inputs or input vectors, namely a [0...3] and b [0...3]. The input vector a or input 510a may represent a scaled input, i.e., a scaled version of data indicating the activities and / or the number of active cells or circuits of at least one IP or functional block (e.g., an MCU). This data can be found in the signals described herein. The input vector b or input 510b may be weights that, when multiplied and added with the input vector a, yield a result indicative of a performance estimate. The weights of the input 510b may be derived from training and optimization processes.

[0067] The weights used for the MADD circuits may be stored or provided in any suitable memory or non-volatile storage circuit or device operatively connected to the MADD circuit. For example, the HPE 200 may be Fig. 2 in the MCU 10 of Fig. 1A. The memory circuit 20 may contain the weights. In other cases, the registers or one of the MCUs may contain these weights.

[0068] The single stage 530 of the MADD circuit 500 includes (four) multipliers 540 and (four) adders 545 connected to determine the output 550. The output or output layer 550 may contain values indicative of real-time dynamic power consumption, e.g., the dynamic power of at least one individual component or IP of an MCU or SoC.

[0069] Fig. 6 shows another MADD circuit 600 according to at least one exemplary embodiment of the present disclosure. The MADD circuit 600 includes two stages, a first stage 630a and a second stage 630b. In one example, the MADD circuit 240 of the HPE 200 may be implemented similarly to the MADD circuit 600. The MADD circuit 600 includes, as shown, multipliers 640 and adders 645.

[0070] The MADD circuit 600 includes an input or input layer 610 and an output or output layer 650. For the first stage 630a, the input layer 610 has two inputs or input vectors, namely a [0...3]) or input 610a and b [0...3]) or input 610b.

[0071] The input or input vector 610a may be the scaled input as described herein. For example, the input 610a may be a scaled version of the input indicating the activities and / or the number of active cells of an aspect of an MCU, e.g., an IP or functional block. The input vector 610b is weights or weighted values. These weights or weighted values may also be trained or optimized. The first stage 630a generates a first output, output 620.

[0072] Another input, input vector c, or input 610c, is provided for the second stage 630b. The input 610 may be a different set of weights or weighted values. These weights or weighted values may also be optimized or trained according to aspects of the present disclosure.

[0073] The first input 610a and the second input 610b are used for the first stage, 630a. The first output or output 620 of the first stage 630a and the input 610c can be used as inputs for the second stage 630b. The output 650 of the second stage 630b is also the output for the MADD circuit 600. Thus, the output 650 can correspond to the dynamic power consumption for at least one component or IP of the microcontroller (e.g., MCU 100) or an SoC. The MADD circuit 600 can generate the output 650 in two clock cycles.

[0074] In short, the MADD circuit 600 comprises two sequential stages for implementing the three-input multiplication, which can be generated or achieved in two clock cycles. Extending this concept, a multi-stage MADD circuit could be implemented to process a four-input multiplication, which can be achieved in three clock cycles. This is shown in Fig. 7 with a circuit designated 700. This circuit 700 can be considered to be essentially identical to the other MADD circuits 500 and 600, with the only difference being that it includes three stages (730a-730c).

[0075] Assuming that each stage n is represented by the equation X × kn + Cn, according to at least one example, a three-stage network or the three-stage MADD circuit 700 may be represented by the equation: ((X*k1+c1)k2+c2)*k3+c3=X*k1*k2*k3+C1*k2*k3+C2*k3+C3 where: - X represents the number of instances of the IP (scaled by the scaling circuit), kn represents the average power consumed by the IP (predefined values), - Cn stands for the leakage contribution of the IP.

[0076] In one or more examples, the values for k1, k2, k3 and C1, C2 and C3 are different values for the same IP.

[0077] Input values X are digital values that can represent the following signals (after scaling): - Hardware IP logic signals for enabling / disabling (example: Core_en signal; ip_inst_en signal) - IP configuration registers / logic signals (example: Lockstep_en signal) - Clock configuration registers / logic signals (example: PLL, DLL and divider configuration signals) - IP performance monitoring signals (example: IPC for computing cores, baud rate equivalents for communication models) - Bus usage equivalence signals (example: bus connection usage duty cycle rates) - State transition signals of the hardware finite state machine (example: init_to_active signal for an FSM) - Device and / or IP power supply signals (example: Sx signal)

[0078] This simple three-stage MADD circuit 700 can perform the sum of 4-input multiplication, 3-input multiplication, and 2-input multiplication. MADD circuit 700 includes inputs 710a-710d, intermediate outputs 720a and 720b, and final output 750.

[0079] Fig. Figure 8 shows a diagram of an HPE Network 800 or aspects thereof. The HPE Network 800 can be configured to estimate the dynamic power consumption of an MCU or SoC, for example, in real time. More specifically, the HPE Network 800 can be configured similarly to the HPE Network 200 in Fig. 2. In addition, the HPE Network 800 can provide instant or real-time estimation of dynamic power consumption for functional blocks or IP components of an MCU or SoC 850. The HPE 800 can be implemented in an MCU, including the MCUs described here.

[0080] In the HPE network 800, each node 810a-810N may represent at least one HPE or HPE unit, which may correspond to a particular IP block. Each node 810a-810N may represent an HPE unit and therefore include a scaling circuit and a MADD circuit, such as in Fig. 2. The MADD circuits of the HPE 810a-810N units can be implemented as described here, with the exception that each MADD circuit can be tailored in terms of its weight to a specific corresponding IP of the MCU / SoC 850.

[0081] The HPE network 800 can be implemented as a hardware network (hardwired) with hardware nodes in the form of the HPE units described here (e.g., HPE 200). The MADD circuits of nodes 810a-810N can be implemented in single- or multi-stage form; see, for example, the MADD circuits 500, 600, or 700.

[0082] In some examples, nodes 810a-810N may only represent MADD circuits, as nodes 810a-810N may share one or more common scaling circuits (e.g., such as scaling circuit 220 of Fig. 2). In such a case, a single scaling circuit can scale inputs, e.g., from the MCU or SoC 850, and provide the scaling inputs to each of the nodes 810a-810n. The HPE network 800 can include circuitry to forward the scaled inputs to the appropriate node. The output of each node can represent an estimate of the dynamic power consumption (e.g., current, voltage, or power in watts) for the corresponding or associated IP, e.g., an electronic device (MCU / SoC 850).

[0083] As long as the HPE supports the number of gates (multipliers, adders, etc.) required for a higher number of MADD stages for all IPs, the power estimation for the entire SoC can be performed on the order of a few clock cycles. Thus, the entire HPE 800 network for estimating the power consumption of an MCU or SoC can be implemented within 10 to 16 clock cycles, since the power estimation from different MADD circuits can be performed for all IPs or components simultaneously or in parallel.

[0084] Various embodiments herein show and describe a hardware-based power estimator (HPE) that can estimate the SoC's power consumption within a few clock cycles, thereby enabling faster dynamic power management.

[0085] The various HPEs described here can be used in both the pre- and post-silicon phases using simple multiplication and addition circuits that can estimate dynamic power consumption quickly and in real time.

[0086] In the post-silicon phase, an HPE (implemented as part of the silicon) can be used to estimate the power consumption of an application / IP block. Since the HPE can estimate power consumption with very low latency, faster transients / changes in power consumption can be estimated. This estimated power can be compared to the actual power consumption measured with a power supply (powering the silicon / DUT). A learning algorithm is used to minimize the least-squares (LMS) loss function and arrive at optimized weights for the HPE, as shown in Fig. 9 shown.

[0087] Various embodiments relate to training a power estimator based on the configuration of individual IPs, subsystems, SoCs, or the overall system. Learning algorithms, LMS (least mean squared) multivariate curve fitting (for specific subsystems or IPs) and neural networks (for other IPs and the entire system-on-chip (SoC)), are used to obtain coefficients for this estimator (HPE). The models for the learning algorithms (LAs) can be trained, for example, with pre-silicon power measurements (using simulators such as PrimePower®) or post-silicon power measurements (using power supplies). The models can then be validated with an independent set of application code / software. Furthermore, the HPE implementation can be scaled to accommodate a higher number of IP instances between different products (or different architectures).In addition, HPE considers both the leakage and dynamic components of silicon power consumption, thus accounting for PVTF fluctuations.

[0088] Returning to equations (4) and (5), a simple linear equation for each node would work well. Therefore, the following function can be used to train the model for HPE: y=max(0,x)

[0089] For a simple linear regression, it can be shown that the minimum set of training data required to arrive at a potential set of weights (W, X, Y) is as follows: (n+m)5 / 3+(n+m)+(n+m)1 / 3.

[0090] However, training to arrive at a suitable weighted set could be performed using a loss function to maximize accuracy. For example, a loss function representing the least mean squared (LMS) error of the power estimate for a set of leaf cells could be used. Since Equation (8) is a multivariate, nonlinear polynomial, the training dataset is typically 5 to 10 times larger than the required number, and a modified gradient descent algorithm is used to optimize the weights.

[0091] Fig. 9 shows an example flowchart and environment 900 illustrating an example post-silicon training process for determining the weights or weighted set of values to be used for the HPEs described herein or portions thereof (e.g., MADD circuits).

[0092] For post-silicon training, a power estimator (PET) or HPE 930 is already implemented as part of the silicon or MCU 915. This allows the HPE 930 to estimate the power consumption of an IP block or application of an MCU, such as any IP, referred to in this example as IPX 920.

[0093] The MCU or SoC 915 may be provided with inputs, e.g., code or input pattern 905, that effect the operation or function of the at least one IPx 920. The corresponding parameters generated by the IPx 920, e.g., the number of active cells, may be captured and used as input, e.g., as signals described herein, to the HPE 930. In this way, the HPE 930 may generate an output, e.g., an estimated power consumption. This may be in the form of an estimated current I est, e.g., 935. The HPE 930 can already be configured or set with initial values for the weighted set 945 (Wi, Xj, Yl), which can be stored in the MCU's registers. Here, too, the HPE 910 can provide a very low-latency power consumption estimate. Relatively fast or faster transients or changes in power consumption can be estimated dynamically.

[0094] Fig. Figure 9 shows the estimated power 935 compared to the actual power consumption 940, which is measured using a power supply 940. The power supply 940 can power the device under test (DUT), e.g., the SoC or the MCU 910.

[0095] The difference between the measured power 945 and the estimated power 935 from the HPE 930 can be used as input for a learning algorithm (LA) 950. The learning algorithm can be implemented in the form of instructions (e.g., stored on a non-transitory computer-readable medium) and executed by one or more processors, e.g., on a separate computing device. The LA 950 uses the current weights 955 and the differences in power measurements between the directly measured and estimated power consumption to determine optimized weights 955 used by the HPE 930. In particular, the LA 950 can be configured to determine an optimized set of weights 955. Any suitable (machine) learning algorithms or techniques can be used to find optimized weights 955.In one example, the LA 950 can use a least-square-mean (LMS) loss function and arrive at optimized weights for the HPE 930.

[0096] Fig. 10 shows an example flowchart and environment 1000 illustrating an example pre-silicon training process for determining weights or weighted value sets for the HPEs described herein. In a pre-silicon training process, a SoC / MCU 1010 and its components, e.g., IPx 1020 and HPE 1030, are not physically realized and implemented. Instead, the SoC / MCU 1010 may be represented as an abstract data form for simulation data, e.g., data used for RTL simulations or similar types of simulations.

[0097] During simulations, various inputs, such as codes or patterns, can be fed into the simulated SoC / MCU 1010, causing one or more operations or tasks from the IPx 1010 to be implemented. As with post-silicon training, the simulated HPE 1010 can provide an estimate of power consumption. This means that parameters generated by the IPx 1020, such as the number of active cells, can be captured and used as input to the HPE 1030. Again, the power consumption can be provided in the form of an estimated current I est , e.g., 1035. Similarly, the HPE 1030 can be simulated with initial values for the weighted set 1055 (Wi, Xj, Yl), which is used for the HPE / PET 1030 and which can be updated for a later simulation using the weights determined by the LA 1050.

[0098] Again, the estimated power consumption 1035 can be compared to the power consumption 1045 generated by the simulated power estimator 1040, which can also be a simulated power supply for the SoC / MCU 1010.

[0099] As in the post-silicon phase, the difference between the simulated measured power 1040 and the simulated estimated power 1035 from the HPE 1030 can be used by a learning algorithm (LA) 1050 to find an optimized weight value.

[0100] Again, the LA 1050 can be implemented as instructions (e.g., stored on a non-transitory computer-readable medium) and executed by one or more processors, e.g., on a separate computing device. The LA 1050 receives the differences in the power measurements as well as the current set of weights 1045. Based on such inputs, the LA 1050 can be configured to determine the optimized set of weights 1055. The LA 1050 can use a least-squares (LMS) loss function to determine optimized weights for the HPE 1030. The LA can be applied repeatedly or iteratively to multiple simulations to update or find the best or optimized set of weights that produces the least error in the power consumption induced by the simulated HPE 1030 power consumption measurement.Because the HPE is not physically realized, the HPE 1030 itself can be repeatedly optimized or its design or configuration updated accordingly to determine a suitable result.

[0101] It should be noted that not all application code used in this training set needs to be functional (in terms of functionality). In a pre-silicon analysis, a vector-driven approach could be used to achieve higher coverage of the internal nodes. Therefore, a set of patterns (or codes) is used to estimate logic activity through simulation or emulation. A Fast Signal Database (FSDB) captures all node / signal activity and serves as input for industry-standard power estimation tools (e.g., PrimePower™ from Synopsys®, Voltus™ from Cadence®, etc.). As discussed later, the power consumption (I meas) with the estimated power consumption I est of a realized or implemented HPE. A learning algorithm can be used to minimize the LMS loss function to obtain optimized weights for the HPE.

[0102] Without loss of generality, the HPEs described here can be considered hardware accelerators for accurate performance estimation. Furthermore, HPE can be used in both the pre- and post-silicon phases to improve overall system performance in the following ways: - Dynamic power management systems: In specific applications, the HPE can be used to estimate power with minimal latency (few clock cycles). This feature of the HPE can be used by various power management systems such as Dynamic Voltage Control (DVC), Dynamic Voltage and Frequency Scaling (DVFS), Envelope Tracking, etc. - Predictive software-based power optimization: An application can be designed so that the power estimated with HPE is available to the processor within a few clock cycles before the actual code is implemented. In this case, the core can prepare itself by telling the voltage control loop that it needs to sink or sink more power. - Architecture exploration: HPE RTL implementations can be used for architecture comparison analysis. For example, different bus / bridge topologies can be evaluated using a previously trained HPE RTL modeling slave IPs. - On-chip and system diagnostics and troubleshooting: HPE real-time data can be used to diagnose problems within an SoC. For example, HPE trace data can be used to evaluate previous activity within the SoC prior to an unintended alarm / reset / event. Similarly, HPE trace data from various silicon components within a given system can be combined to determine an overall system state at a specific point in time.

[0103] Fig. 11 illustrates a method 1100 consistent with aspects of the present disclosure. The method 1100 includes, at 1110, receiving input comprising data values representing state information or activity information from one or more intellectual property (IP) components or blocks of the electronic device.

[0104] At 1120, the method 1100 includes generating a scaled nonlinear output using a scaling circuit by applying an approximation of an activation function to the input.

[0105] At 1130, method 1100 includes generating an output indicative of a mathematical performance estimate of the electronic device using the generated scaled nonlinear output using multiply-add (MADD) circuitry.

[0106] Fig. 12 illustrates a method 1200 consistent with aspects of the present disclosure. The method 1200 includes, at 1210, providing input from a version of the electronic device to a prototype of the hardware performance estimator.

[0107] At 1220, the method 1200 includes determining an output of the prototype hardware performance estimator based on the provided input.

[0108] At 1230, method 1200 includes determining a power measurement of the electronic device corresponding to the mathematical power estimate indicated by the output of the prototype hardware power estimator.

[0109] At 1240, the method 1200 includes applying a learning algorithm to the difference between the determined output of the prototype hardware power estimator and the determined power measurement to derive optimized weight values for the MADD circuitry of the prototype hardware power estimator.

[0110] The HPEs described here can be used in conjunction with other software and debugging tools to predict power consumption for specific application code, helping to efficiently optimize the application code. This provides design and architecture teams with early information for planning silicon parameters.

[0111] The following examples relate to further aspects of the present disclosure:

[0112] Example 1 is a hardware performance estimator (HPE) configured to provide a power estimate for an electronic device, the hardware performance estimator comprising: a scaling circuit configured to receive an input and generate a scaled nonlinear output of the input, the scaling circuit applying an approximation of an activation function to the input; and multiply-add (MADD) circuitry configured to generate an output indicative of a mathematical performance estimate of the electronic device using the output of the scaling, the input to the scaling circuit comprising data values representing state information or activity information from one or more intellectual property (IP) components or blocks of the electronic device.

[0113] Example 2 is the subject of Example, wherein the scaling circuit may include a plurality of hardware components configured to perform the approximation of the activation function to the received input.

[0114] Example 3 is the subject matter of Example 2, wherein the scaling circuit may be configured to piecewise apply the activation function approximation to generate a plurality of piecewise outputs in parallel, and wherein the scaling circuit may be further configured to sum the generated piecewise outputs and generate the output of the scaling circuit.

[0115] Example 4 is the subject matter of Example 3, wherein to generate at least one piecewise output, the scaling circuit may be configured to perform a multiply-add operation on the input.

[0116] Example 5 is the subject matter of Example 3, wherein to generate at least one piecewise output, the scaling circuit is configured to apply one or more shift operations to the input.

[0117] Example 6 is the subject matter of Example 3, wherein to generate at least one piecewise output, the scaling circuit may be configured to apply at least one or more shift operations and at least one multiply-add operation to the input.

[0118] Example 7 is the subject matter of Example 3, wherein to generate at least one piecewise output, the scaling circuit may be configured to apply a comparator with predefined thresholds to the input.

[0119] Example 8 is the subject matter of any of Examples 1 to 7, wherein the scaling circuit configured to apply an approximation to an activation function may include the scaling circuit to apply an approximation to a hyperbolic tangent function.

[0120] Example 9 is the subject matter of any of Examples 1 to 8, wherein the scaling circuit configured to apply an approximation to an activation function may include the scaling circuit to apply an approximation to a rectified linear unit function.

[0121] Example 10 is the subject matter of any of Examples 1 to 9, wherein the MADD circuitry may comprise a plurality of MADD circuits arranged in stages such that the output of one stage is input to a subsequent stage.

[0122] Example 11 is the subject matter of Example 10, wherein each stage may include a MADD circuit configured to calculate an accumulated sum of product values between a first input and a second input to the stage, wherein the first input comprises an output from a previous stage or an output from the scaling circuit, and wherein the second input comprises predefined weight values.

[0123] Example 12 is the subject matter of Example 11, wherein the predefined weight values may correspond to one or more electrical parameters and / or states of the electronic device.

[0124] Example 13 is the subject matter of Example 11, wherein the predefined weight values may correspond to one or more electrical parameters and / or states of a component of the electronic device.

[0125] Example 14 is the subject matter of Example 11, wherein the predefined weight values may correspond to one or more parameters and / or states of a hardware accelerator of the electronic device.

[0126] Example 15 is the subject matter of Example 11, wherein the predefined weight values may correspond to one or more parameters and / or states of an IP of the electronic device.

[0127] Example 16 is the subject of Example 11, where the predefined weight values are values that can be determined from a training process.

[0128] Example 17 is the subject matter of Example 16, wherein the training process may include a process comprising: providing a simulated input from a simulated electronic hardware device to a prototype of the hardware performance estimator, determining a simulated performance measurement, and applying a learning algorithm to the difference between the performance estimate from the prototype of the hardware performance estimator and the simulated performance measurement to derive the predefined weight values.

[0129] Example 18 is the subject matter of any of Examples 10 to 17, wherein the MADD circuitry may include three stages or fewer such that an output is generated from an input in three clock cycles or fewer.

[0130] Example 19 is the subject matter of any of Examples 1 to 18, wherein the hardware power estimator may be configured to provide the power estimate for the electronic device in real time.

[0131] Example 20 is the subject matter of any of Examples 1 to 20, wherein the electronic device may be a microcontroller chip.

[0132] Example 21 is the subject matter of any of Examples 1 to 20, wherein the MADD circuitry is configured to generate an output signal indicative of a mathematical power estimate according to the following equation: Pdyn=V2∑i⋅fiαiCi where: α stands for the activity of a particular node / network, C stands for the effective capacity, f stands for the effective frequency, V stands for voltage.

[0133] Example 22 is the subject matter of any of Examples 1 to 21, wherein the input of the scaling circuit may include an IP enable or disable signal.

[0134] Example 23 is the subject matter of any of Examples 1 to 22, wherein the input to the scaling circuit may include an IP configuration register or a logic signal.

[0135] Example 24 is the subject matter of any of Examples 1 to 23, wherein the input of the scaling circuit may include a clock configuration register or a logic signal.

[0136] Example 25 is the subject matter of any of Examples 1 to 24, wherein the input to the scaling circuit may include an IP power monitoring signal.

[0137] Example 26 is the subject matter of any of Examples 1 to 25, wherein the input to the scaling circuit may include a bus utilization equivalent signal.

[0138] Example 27 is the subject matter of any of Examples 1 to 26, wherein the input to the scaling circuit may include a hardware finite state machine state transition signal.

[0139] Example 28 is the subject matter of any of Examples 1 to 27, wherein the input to the scaling circuit may include a power state signal.

[0140] Example 29 is a microcontroller chip that can include the hardware power estimator of any of claims 1 to 29, wherein the electronic device is the microcontroller chip.

[0141] Example 30 is the subject matter of Example 20, wherein the microcontroller may include a plurality of functional blocks or IP blocks, at least some of the functional blocks having at least one state variable associated therewith; and at least some of the functional blocks being coupled to the scaling circuitry such that at least some of the state variables act as said input to the scaling circuitry.

[0142] Example 1A is a method for hardware performance estimation for an electronic device, the method comprising obtaining an input comprising data values representing state information or activity information from one or more intellectual property (IP) components or blocks of the electronic device; generating a scaled nonlinear output using scaling circuitry by applying an approximation of an activation function to the input; and generating an output indicative of a mathematical performance estimation of the electronic device using the generated scaled nonlinear output using multiply-add (MADD) circuitry.

[0143] Example 2A is the subject matter of Example 1A, wherein the scaling circuit may include a plurality of hardware components configured to perform the approximation of the activation function to the received input.

[0144] Example 3A is the subject matter of Example 1A or 2A, wherein applying the activation function to the input may comprise generating a plurality of piecewise outputs in parallel by piecewise applying the activation function to the input and summing the generated piecewise outputs to generate said nonlinear output.

[0145] Example 4A is the subject matter of Example 3A, wherein generating at least one piecewise output may include the scaling circuit performing a multiply-add operation on the input.

[0146] Example 5A is the subject matter of Example 3A or 4A, wherein generating at least one piecewise output comprises the scaling circuit applying one or more shift operations to the input.

[0147] Example 6A is the subject matter of any of Examples 3A to 5A, wherein generating at least one piecewise output may include the scaling circuit applying at least one or more shift operations and at least one multiply-add operation to the input.

[0148] Example 7A is the subject matter of any of Examples 3A to 6A, wherein generating at least one piecewise output may include the scaling circuit applying a comparator with predefined thresholds to the input.

[0149] Example 8A is the subject matter of any of Examples 1A to 8A, wherein the scaling circuitry that applies the approximation of an activation function to the input may include the scaling circuitry that applies an approximation of a hyperbolic tangent function to the input.

[0150] Example 9A is the subject matter of any of Examples 1A to 8A, wherein the scaling circuitry applying an approximation of the activation function to the input may include the scaling circuitry applying an approximation of a rectified linear unit function to the input.

[0151] Example 10A is the subject matter of any of Examples 1A to 10A, wherein the MADD circuitry may include a plurality of MADD circuits arranged in stages such that the output of one stage is input to a subsequent stage.

[0152] Example 11A is the subject matter of Example 10A, wherein each stage comprises a MADD circuit that is configurable to calculate an accumulated sum of product values between a first input and a second input to the stage, wherein the first input comprises an output from a previous stage or an output from the scaling circuit, and wherein the second input comprises predefined weight values.

[0153] Example 12A is the subject matter of Example 11A, wherein the predefined weight values may correspond to one or more electrical parameters and / or states of the electronic device.

[0154] Example 13A is the subject matter of Example 11A or 12A, wherein the predefined weight values may correspond to one or more electrical parameters and / or states of a component of the electronic device.

[0155] Example 14A is the subject matter of any of Examples 11A to 13A, wherein the predefined weight values may correspond to one or more parameters and / or states of a hardware accelerator of the electronic device.

[0156] Example 15A is the subject matter of any of Examples 12A to 15A, wherein the predefined weight values may correspond to one or more parameters and / or states of an IP of the electronic device.

[0157] Example 16A is the subject matter of any of Examples 11A to 15A, wherein the predefined weight values may be values that may be determined from a training process.

[0158] Example 17A is the subject matter of any of Examples 10A to 16A, wherein the MADD circuitry may comprise three stages or fewer, and wherein generating the output using the MADD circuitry comprises generating the output from the scaled nonlinear output in three clock cycles or fewer.

[0159] Example 18A is the subject matter of any of Examples 1A to 18A, wherein the output indicative of a mathematical performance estimate of the electronic device is generated in real time.

[0160] Example 1B is methods for training a hardware power estimator configured to provide a power estimate for an electronic device, the hardware power estimator including a scaling circuit configured to receive an input and generate a scaled nonlinear output of the input, the scaling circuit applying an approximation of an activation function to the input, and multiply-add (MADD) circuitry configured to generate an output indicative of a mathematical power estimate of the electronic device using the output of the scaling, wherein the training may comprise: providing an input from a version of the electronic device to a prototype of the hardware power estimator,Determining an output from the prototype hardware power estimator based on the provided input; determining a power measurement of the electronic device corresponding to the mathematical power estimate indicated by the output of the prototype hardware power estimator; and applying a learning algorithm to the difference between the determined output from the prototype hardware power estimator and the determined power measurement to derive optimized weighting values for the MADD circuitry of the prototype hardware power estimator.

[0161] Example 2B is the subject of Example 1B, where the prototype of the hardware performance estimator may be a simulated version of the hardware performance estimator.

[0162] Example 3B is the subject matter of Example 2B, wherein the electronic device is a simulated electronic device that is configurable to provide its output as a simulated input to the simulated version of the hardware power estimator, wherein determining the power measurement comprises performing the power measurement of the electronic device in a simulation.

[0163] Example 4B is the subject matter of Example 1B, where the prototype of the hardware performance estimator may be a physical version of the hardware performance estimator.

[0164] Example 5B is the subject matter of Example 4B, wherein the electronic device is a physical electronic device that can be configured to provide its output as an input to the physical version of the hardware performance estimator, wherein determining the performance measurement comprises performing performance measurements of the physical version of the electronic device.

[0165] It should be noted that one or more of the features of each of the above examples may be suitably combined with any of the other examples or with the embodiments disclosed herein.

[0166] The foregoing description has been given by way of example only, and it will be appreciated by those skilled in the art that changes may be made without departing from the broader spirit or scope of the invention as set forth in the claims. The specification and drawings are therefore to be considered illustrative rather than restrictive.

[0167] The scope of the disclosure is therefore indicated by the appended claims, and all changes which come within the meaning and equivalence of the claims are therefore intended to be embraced.

[0168] It is understood that implementations of methods described in detail herein are demonstrative in nature and are thus understood to be capable of being implemented in a corresponding device. Likewise, it is understood that implementations of devices described in detail herein are understood to be capable of being implemented as a corresponding method. Likewise, it is understood that implementations of devices described in detail herein are understood to be capable of being implemented as a corresponding method. It is thus understood that a device corresponding to a method described in detail herein may include one or more components configured to perform any aspect of the associated method.

[0169] All acronyms defined in the above description shall additionally apply in all claims contained herein.

Claims

[1] A hardware power estimator configured to provide a power estimate for an electronic device, the hardware power estimator comprising: a scaling circuit configured to receive an input and generate a scaled nonlinear output of the input, the scaling circuit applying an approximation of an activation function to the input; and a multiply-add circuitry (MADD) configured to generate an output signal indicative of a mathematical performance estimate of the electronic device using the output signal of the scaling, wherein the input to the scaling circuit comprises data values representing state information or activity information of one or more intellectual property (IP) components or blocks of the electronic device. [2] The hardware power estimator of claim 1, wherein the scaling circuit comprises a plurality of hardware components configured to perform the approximation of the activation function to the received input. [3] Hardware performance estimator according to claim 2, wherein the scaling circuit is configured to apply the activation function approximation piecewise to generate a plurality of piecewise outputs in parallel, and wherein the scaling circuit is further configured to sum the generated piecewise outputs and generate the output of the scaling circuit. [4] The hardware power estimator of claim 3, wherein the scaling circuit for generating at least one piecewise output is configured to perform a multiply-add operation on the input. [5] The hardware power estimator of claim 3, wherein the scaling circuit is configured to generate at least one piecewise output by applying one or more shift operations to the input. [6] The hardware power estimator of claim 3, wherein to generate at least one piecewise output, the scaling circuit is configured to apply at least one or more shift operations and at least one multiply-add operation to the input. [7] The hardware power estimator of claim 3, wherein the scaling circuit for generating at least one piecewise output is configured to apply a comparator with predefined thresholds to the input. [8] The hardware performance estimator of any one of claims 1 to 7, wherein the scaling circuit configured to apply an approximation to an activation function comprises the scaling circuit applying an approximation to a hyperbolic tangent function. [9] The hardware power estimator of any one of claims 1 to 7, wherein the scaling circuit configured to apply an approximation to an activation function comprises the scaling circuit applying an approximation to a rectified linear unit function. [10] A hardware power estimator according to any one of claims 1 to 9, wherein the MADD circuitry comprises a plurality of MADD circuits arranged in stages such that the output of one stage is input to a subsequent stage. [11] Hardware performance estimator according to claim 10, wherein each stage comprises a MADD circuit configured to calculate an accumulated sum of product values between a first input and a second input of the stage, wherein the first input comprises an output signal from a previous stage or an output signal from the scaling circuit, and where the second input includes predefined weight values. [12] Hardware performance estimator according to claim 11, wherein the predefined weight values correspond to one or more electrical parameters and / or states of the electronic device, wherein the predefined weight values correspond to one or more electrical parameters and / or states of a component of the electronic device, wherein the predefined weight values correspond to one or more parameters and / or states of a hardware accelerator of the electronic device, and / or wherein the predefined weight values correspond to one or more parameters and / or states of an IP of the electronic device. [13] A hardware power estimator according to claim 11 or 12, wherein the MADD circuitry comprises three stages or less and thus produces an output from an input in 3 clock cycles or less. [14] The hardware power estimator of any one of claims 1 to 13, wherein the hardware power estimator is configured to provide the power estimate for the electronic device in real time. [15] The hardware power estimator of any one of claims 1 to 14, wherein the input to the scaling circuit comprises an IP configuration register or logic signal, a clock configuration register or logic signal, an IP power monitor signal, a bus utilization equivalence signal, a hardware finite state machine state transition signal, and / or a power state signal. [16] A method for hardware performance estimation for an electronic device, the method comprising: Receiving an input with data values representing state information or activity information of one or more intellectual property (IP) components or blocks of the electronic device; Generating a scaled nonlinear output using a scaling circuit by applying an approximation of an activation function to the input; and Generating an output signal indicative of a mathematical performance estimate of the electronic device using the generated scaled nonlinear output signal using multiply-add (MADD) circuitry. [17] Method according to claim 16, wherein the scaling circuit comprises a plurality of hardware components configured to perform the approximation of the activation function to the received input, where the application of the activation function to the input comprises: parallel generation of a plurality of piecewise outputs by piecewise application of the activation function to the input, and Summing the generated piecewise outputs to produce said nonlinear output. [18] Method according to claim 17, wherein generating at least one piecewise output consists in the scaling circuit performing a multiply-add operation on the input. wherein generating at least one piecewise output comprises the scaling circuit applying one or more shift operations to the input, and wherein generating at least one piecewise output comprises the scaling circuit applying at least one or more shift operations and at least one multiply-add operation to the input. [19] A method for training a hardware power estimator configured to provide a power estimate for an electronic device, the hardware power estimator comprising a scaling circuit configured to receive an input and generate a scaled non-linear output of the input, the scaling circuit applying an approximation of an activation function to the input, and multiply-add (MADD) circuitry configured to generate an output indicative of a mathematical power estimate of the electronic device using the output of the scaling, the training comprising: Providing an input from a version of the electronic device to a prototype of the hardware performance estimator, determining an output of the prototype hardware performance estimator based on the provided input; Determining a power measurement of the electronic device according to the mathematical power estimate indicated by the output of the prototype hardware power estimator; and Applying a learning algorithm to the difference between the determined output of the hardware power estimator prototype and the determined power measurement to derive optimized weight values for the MADD circuitry of the hardware power estimator prototype. [20] Method according to claim 19, where the prototype of the hardware performance estimator is a simulated version of the hardware performance estimator, wherein the electronic device is a simulated electronic device configured to provide its output as a simulated input to the simulated version of the hardware performance estimator, and wherein determining the power measurement comprises performing the power measurement of the electronic device in a simulation.

Citation Information

Patent Citations

  • ADAPTIVE DIGITAL ON-CHIP POWER ASSESSOR

    DE112020004583T5

  • Dynamic weight calculation in a digital power estimation and management system

    US20130024713A1