Configurable processor

The configurable processor system addresses the challenge of rapid function switching in RF systems by employing a network-on-chip architecture and machine learning-based scheduling, achieving efficient and low-power operation for diverse RF tasks.

WO2026236683A1PCT designated stage Publication Date: 2026-11-12THE ARIZONA BOARD OF REGENTS ON BEHALF OF THE UNIV OF ARIZONA +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027399
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2026-05-08
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing radio frequency systems face challenges in efficiently switching between different processing modes, such as communications and sensing, due to the need for rapid function switching and varying operational requirements.

Method used

A configurable processor system with a network-on-chip architecture, including multiple accelerators and a scheduler that uses machine learning for dynamic task scheduling based on workload and capability, enabling efficient switching between processing functions.

Benefits of technology

The system achieves high throughput and low power consumption, supporting multiple concurrent operations with latency as low as 1 microsecond and energy efficiency, facilitating rapid mode transitions in RF systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027399_12112026_PF_FP_ABST
    Figure US2026027399_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A configurable processor. In some embodiments, a system includes: a plurality of processing circuits, and a network, connected to each of the processing circuits, the processing circuits including: a first accelerator; a second accelerator; and a general-purpose processor, a first processing circuit of the processing circuits being configured to perform a method for scheduling, the method for scheduling including: determining that a computational task is to be performed; and assigning the computational task to a second processing circuit of the processing circuits, the assigning of the computational task including assigning the computational task based on: a capability of the second processing circuit; a workload of the second processing circuit; a capability of a third processing circuit of the processing circuits; and a workload of the third processing circuit.
Need to check novelty before this filing date? Find Prior Art

Description

CONFIGURABLE PROCESSORU.S. GOVERNMENT LICENSE RIGHTS

[0001] This invention was made with government support under FA8650-18-2-7860 awarded by DARPA (Defense Advanced Research Projects Agency). The government has certain rights in the invention.FIELD

[0002] One or more aspects of embodiments according to the present disclosure relate to computing systems, and more particularly to a configurable processor.BACKGROUND

[0003] Radio frequency (RF) systems may have various applications, including communications, sensing applications (such as radar), and position, navigation, and timing (PNT) applications.

[0004] It is with respect to this general technical environment that aspects of the present disclosure are related.SUMMARY

[0005] According to an embodiment of the present disclosure, there is provided a system, including: a plurality of processing circuits, and a network, connected to each of the processing circuits, the processing circuits including: a first accelerator; a second accelerator; and a general-purpose processor, a first processing circuit of the processing circuits being configured to perform a method for scheduling, the method for scheduling including: determining that a computational task is to be performed; and assigning the computational task to a second processing circuit of the processing circuits, the assigning of the computational task including assigning the computational task based on: a capability of the second processing circuit; a workload of the second processing circuit; a capability of a third processing circuit of the processing circuits; and a workload of the third processing circuit.

[0006] In some embodiments, each of the first accelerator and the second accelerator is selected from the group consisting of fast Fourier transform accelerators, forward error correction encoding accelerators, forward error correction decoding accelerators, finite -1- 4899-0920-1066, v. 1impulse response filtering accelerators, correlation accelerators, matrix multiply accelerators, matrix inversion accelerators, machine learning accelerators, eigensystem accelerators, singular value decomposition channel estimation accelerators, filter coefficient calculation accelerators, spatial beam former estimators, space-time beam former estimators, beam former application accelerators, likelihood estimators, symbol modulation accelerators, symbol demodulation accelerators, signal detection accelerators, signal acquisition accelerators, channel compensation accelerators, linear algebra accelerators, and combinations thereof.

[0007] In some embodiments, the second accelerator is different from the first accelerator.

[0008] In some embodiments, the first accelerator is a linear algebra accelerator, configured to perform matrix multiplications, matrix inversions, finite impulse response filtering, and eigenvalue calculations.

[0009] In some embodiments, the linear algebra accelerator includes an array of heterogeneous sets of processing circuits.

[0010] In some embodiments, the network has a bus width of at least 64 bits.

[0011] In some embodiments, the assigning includes selecting a set of instructions for execution, the selecting being based on: the computational task, and the second processing circuit.

[0012] In some embodiments, the system is capable of operating with a throughput of 1 gigabyte per second while consuming less than 10 watts of electrical power.

[0013] In some embodiments, the system is further capable of performing 10 giga operations per second, each operation having two 32-bit operands.

[0014] In some embodiments, the system further includes a radio-frequency (RF) front end, connected to the network, the RF front end being configured to transmit or receive analog RF signals.

[0015] In some embodiments, the RF front end includes a buffer.

[0016] In some embodiments, the processing circuits further include a cache, connected to the network.

[0017] In some embodiments, the system includes a system on a chip including the plurality of processing circuits and the network, wherein the system further includes an off-chip memory.-2- 4899-0920-1066, v. 1

[0018] In some embodiments, the wherein the processing circuits further include one or more interface circuits.

[0019] In some embodiments, the network is configured to transmit data from any one of the processing circuits to any other one of the processing circuits.

[0020] According to an embodiment of the present disclosure, there is provided a method, including performing a method for scheduling, by a first processing circuit of a plurality of processing circuits of a system, the system further including a network connected to each of the processing circuits, the method for scheduling including: determining that a computational task is to be performed; and assigning the computational task to a second processing circuit of the processing circuits, the assigning of the computational task including assigning the computational task based on: a capability of the second processing circuit; a workload of the second processing circuit; a capability of a third processing circuit of the plurality of processing circuits; and a workload of the third processing circuit.

[0021] In some embodiments, the processing circuits include: a first accelerator; and a second accelerator, different from the first accelerator.

[0022] In some embodiments, the first accelerator is a linear algebra accelerator, configured to perform matrix multiplications, matrix inversions, finite impulse response filtering, and eigenvalue calculations.

[0023] In some embodiments, the linear algebra accelerator includes an array of heterogeneous sets of processing circuits.

[0024] In some embodiments, the network has a bus width of at least 64 bits.BRIEF DESCRIPTION OF THE DRAWINGS

[0025] These and other features and advantages of the present disclosure will be appreciated and understood with reference to the specification, claims, and appended drawings wherein:

[0026] FIG. 1 is a system-level block diagram, according to an embodiment of the present disclosure;

[0027] FIG. 2 is a block diagram of a configurable processor, according to an embodiment of the present disclosure;-3- 4899-0920-1066, v. 1

[0028] FIG. 3 is a block diagram of a processing element, according to an embodiment of the present disclosure;

[0029] FIG. 4A is a table of performance parameters, according to an embodiment of the present disclosure;

[0030] FIG. 4B is a table of performance parameters, according to an embodiment of the present disclosure; and

[0031] FIG. 4C is a performance graph, according to an embodiment of the present disclosure.DETAILED DESCRIPTION

[0032] The detailed description set forth below in connection with the appended drawings is intended as a description of exemplary embodiments of a configurable processor provided in accordance with the present disclosure and is not intended to represent the only forms in which the present disclosure may be constructed or utilized. The description sets forth the features of the present disclosure in connection with the illustrated embodiments. It is to be understood, however, that the same or equivalent functions and structures may be accomplished by different embodiments that are also intended to be encompassed within the scope of the disclosure. As denoted elsewhere herein, like element numbers are intended to indicate like elements or features.

[0033] The entire content of US Patent Application No. 18 / 249,851, entitled “RUNTIME TASK SCHEDULING USING IMITATION LEARNING FOR HETEROGENEOUS MANY-CORE SYSTEMS” is incorporated herein by reference.

[0034] In radio-frequency (RF) systems (including microwave, millimeter-wave, and terahertz systems) various intermediate frequency operations of moderate to high complexity may be performed at relatively high throughput, and in different states of such a system, different operations may be required. For example, in a wireless communications system, channel estimation may be performed periodically, and data reception may be performed at other times. These two modes of operations may require different processing of the received radio frequency signal, and it may be advantageous for a processor of the system to be configurable, e.g., capable of switching rapidly between the corresponding modes. Some radio frequency systems may be capable of switching between different functions such as communications and sensing (or of -4- 4899-0920-1066, v. 1performing such functions simultaneously); such systems also may benefit from the use of a processor that is capable of switching rapidly between processing functions.

[0035] FIG. 1 is a system-level diagram of a radio system including a radio frequency (RF) front end, a configurable processor, and a memory and a host, both connected to the configurable processor. As used herein, “RF” refers to electromagnetic radiation having a wavelength longer than 1 mm and shorter than 3 km. Within the RF front end, an antenna provides coupling between (inbound and outbound) free-space electromagnetic waves and guided waves in a transmission line (e.g., a coaxial cable) connecting the antenna to an analog front end. The analog front end includes a power amplifier and a low-noise amplifier for amplifying outbound and inbound signals, respectively. The amplifiers may be connected to the antenna through a transmit-receive (T / R) switch or through a diplexer or a duplexer (e.g., a circulator). The power amplifier and a low-noise amplifier may be connected to a digital to analog converter (DAC) and an analog to digital converter (ADC) respectively, connected to the configurable processor. The host may be or include a processing system running an application (e.g., a communications application, a radar application, or a sensing application) that is a consumer or supplier of data. The host may also provide control functions to the configurable processor.

[0036] FIG. 2 is a block diagram of the configurable processor. The configurable processor includes a plurality of blocks, or circuits, connected together by a network (e.g., network-on-chip (NoC)). Each of the blocks may be a processing circuit, e.g., a general-purpose processor, an interface circuit (providing an interface, e.g., to the RF front end, the memory, or the host) or a memory (e.g., a cache (e.g., a level-3 cache)). The network-on-chip may be, e.g., a commercially available network-on-chip such as an Advanced RISC Machines (ARM) CMN 600. The network-on-chip may include a data channel and a control channel, and it may be configured to operate as a packet-switched network. As such, it may be capable of reconfiguring data paths with low latency, e.g., with a latency of less than 1 microsecond. The network-on-chip may be sufficiently wide (e.g., 256 bits wide) to allow high data throughput between the processing circuits.

[0037] The controller may be or include one or more general purpose processors such as an ARM core, e.g., an ARM Coretex™ A72, A53 or A75, and it may include one or more caches (e.g., a level-1 cache and a level-2 cache). In an embodiment in which -5- 4899-0920-1066, v. 1the controller includes a plurality of cores, the cores may also share a level-3 cache. The controller may be responsible for managing the other processing circuits of the configurable processor, e.g. configuring each accelerator by writing values to the registers of the accelerator. For example, (i) for a fast Fourier transform (FFT) accelerator, it may configure the size of the FFT to be performed, or (ii) for an accelerator (such as a domain-adaptive processor (discussed in further detail below)) which is a stored-program computer, it may arrange for instructions to be stored in the memory or loaded into a cache of the accelerator, and it may write a pointer to an initial instruction to the program counter of the accelerator. The controller may also, in this manner, configure each processor to send output data to another accelerator, to the level-3 cache, to an interface circuit, or to the controller.

[0038] As shown in FIG. 2, the configurable processor may include one or more accelerators. The configurable processor may be implemented in a system on a chip (SoC); the host, memory and RF front end shown in FIG. 1 may be off-chip elements. Three accelerators are shown in FIG. 2; some embodiments include more or fewer accelerators. The accelerators may have access to, and use, the cache illustrated in FIG. 2 (which may be a level-3 cache), and the accelerators may also have a last level cache (LLC) that is shared between them. Each accelerator may be the same as one or more other accelerators (e.g., the configurable processor may include a plurality of FFT accelerators) or different from some or all of the other accelerators. Each accelerator may be selected from the group consisting of fast Fourier transform accelerators, forward error correction (FEC) encoding accelerators, forward error correction decoding accelerators, finite impulse response filtering accelerators, correlation accelerators, matrix multiply accelerators, matrix inversion accelerators, machine learning accelerators, eigensystem accelerators, singular value decomposition channel estimation accelerators, filter coefficient calculation accelerators (e.g., Weiner filters or adaptive equalizers), spatial beam former estimators, space-time beam former estimators, beam former application accelerators, likelihood estimators, symbol modulation accelerators, symbol demodulation accelerators, signal detection accelerators, signal acquisition accelerators, channel compensation accelerators, linear algebra accelerators, and combinations thereof, or it may be a different accelerator such as a domain adaptive processor (DAP) (discussed in further detail below).-6- 4899-0920-1066, v. 1

[0039] Each accelerator may receive data from (i) another accelerator, (ii) the cache, or (iii) the memory, according to configuration settings stored in the accelerator by the controller. Each accelerator may include an internal (level-1) cache, which may be used to store (I) input data to be processed, and (ii) output data that results from processing performed by the accelerator. If the internal cache is full, such data may be temporarily stored in the level-3 cache.

[0040] In some embodiments, a software-generation tool is used, when source code for, e.g., communications, radar, and sensing tasks is generated, that may (i) change the order of performance of computational tasks and (ii) change the destinations to which various such tasks send output data, to improve the efficiency (e.g., the latency, the throughput, or the energy efficiency) of a set of computational tasks. For example, if a communications task uses an FFT accelerator followed by an FEC decoding accelerator, the software-generation tool may adjust the code so that the communications task uses the FFT accelerator, and so that the output of the FFT accelerator is sent directly to the FEC accelerator (instead of, e.g., being stored first in the level-3 cache).

[0041] The configurable processor may include a scheduler, which may be an intelligent scheduler, running on a dedicated processing circuit connected to the network-on-chip as shown in FIG. 2. In some embodiments, the scheduler instead runs in a non-dedicated processing circuit, e.g., in the controller. The scheduler may use machine learning to schedule each of a plurality of computational tasks as the need to perform these computational tasks arises. Scheduling a computational task may entail assigning it to a computational resource, e.g., to a processing circuit, and determining the order in which it will be performed on that computational resource, relative to other computational tasks. For example, if a communications task involves performing an FFT, the scheduler may schedule the task for execution on one of several candidate processing circuits, including an FFT accelerator, a domain adaptive processor, and the controller, depending, e.g., on the current workload of each of the processing circuits and on the capabilities of each of the processing circuits.

[0042] For example, the scheduler may assign the FFT to the FFT accelerator even if the workload of the FFT accelerator is significantly higher than the workload of the controller, because the FFT accelerator may have significantly greater capabilities (e.g., capabilities to perform the task with lower latency or with higher throughput or with lower -7- 4899-0920-1066, v. 1energy consumption than the controller). Conversely, in other circumstances the scheduler may assign the FFT to the controller in spite of the controller’s inferior capabilities if the controller’s workload is sufficiently low compared to that of the FFT accelerator. In some embodiments, instructions for performing a computational task (e.g., the FFT) may be compiled for several different targets (e.g., for the controller and for the domain adaptive processor) and the set of instructions for execution (e.g., the object code) for each of the targets may be stored, e.g., in the memory, so that a set of instructions for execution, for the computational task, is available (and may be selected based on the computational task, and based on the processing circuit selected to perform the computational task) regardless of which processing circuit the processing task is assigned to.

[0043] In some embodiments, the software-generation tool manages resources (power consumption, accelerator usage, etc.) and may perform some level of memory optimization (such as sending data from a first accelerator to a second accelerator directly instead of writing to memory, by the first accelerator, and having the second accelerator read from memory). The intelligent scheduler may decide which operation will be performed on which hardware resource. For example, if a first task uses an FFT and FEC, and a second task uses an FFT, the software-generation tool may arrange for the first task to send data from the FFT accelerator to the FEC accelerator, and the scheduler may decide that the FFT accelerator will be allocated first to the first task, and then to the second task.

[0044] The scheduler may be trained using imitation learning. For example, a first machine learning system, which may be referred to as an “expert system” may be trained to assign computational tasks so as to minimize a cost function, using a variety of computational load scenarios (e.g., various combinations of communications tasks and sensing tasks). The cost function may be selected to take into account the latency, throughput, and energy consumption for the performance of all of the computational tasks in the set of tasks of each training data element. This expert system may be relatively large and slow, and, as such, it may be used to train the scheduler (using imitation learning), which may be a smaller, faster machine learning system, capable of scheduling each computational task quickly. For example, the scheduler may be capable of scheduling each computational task with a latency of between 1 ms and 10 ns (e.g.,-8- 4899-0920-1066, v. 1with a latency of less than 1 microsecond). After being trained in this manner, the scheduler may be configured to perform a method for scheduling, the method for scheduling comprising (i) determining that a first computational task is to be performed and (ii) assigning the computational task to a first processing circuit of the processing circuits (e.g., to the FFT accelerator). The assigning of the computational task may include assigning the computational task based on (i) a capability of the first processing circuit, (ii) a workload of the first processing circuit (iii) a capability of a second processing circuit of the plurality of processing circuits (e.g., of the controller) and (iv) a workload of the second processing circuit.

[0045] In some embodiments, one of the accelerators is a forward error correction decoding accelerator configured to decode any one of a plurality of forward error correction codes that are capable of being decoded using a method involving message passing, and a second one of the accelerators is a forward error correction decoding accelerator configured to decode any one of a plurality of forward error correction codes based on Galois fields. As such, in such an embodiment, any one of a large number of forward error correction codes falling into either of these categories may be efficiently decoded by the configurable processor.

[0046] As shown in both FIG. 1 and FIG. 2, the configurable processor may have an interface to each of the RF front end, the memory, and the host. Each interface may include a buffer or a pair of buffers to store, temporarily, incoming and outgoing data. In some embodiments the three interfaces are implemented in a single interface (e.g., a network interface) providing an interface to each of the RF front end, the memory, and the host. In some embodiments, the RF front end similarly includes a buffer or a pair of buffers to store, temporarily, incoming and outgoing data.

[0047] In some embodiments, the configurable processor includes a processing circuit (e.g., an accelerator) for performing encryption and decryption. In such an embodiment, all of the configurable processor except for any circuitry (not shown) on the decrypted side of the encryption and decryption circuit may be a black portion of the configurable processor, and the remaining circuitry (on the other side of the encryption and decryption circuit (e.g., storage or an interface to a (red) external host)) may be a red section of the configurable processor.-9- 4899-0920-1066, v. 1

[0048] In some embodiments, a configurable processor may be capable of transmitting or receiving multiple waveforms, and multiple communication types simultaneously, with different time varying loads. For example, the configurable processor may support multiple concurrent communication links with different remote transceivers, concurrently with performing sensing (e.g., radar) functions, or analysis of RF radiation emitted by other sources to assess availability of certain spectral regions for various uses; all of these operations may be performed through the single (shared) RF front end.

[0049] As mentioned above, one or more of the accelerators may be a domain adaptive processor. A domain adaptive processor may be or include a heterogeneous processor including a plurality (e.g., an array) of interconnected processing elements (each of which is a heterogeneous set of processing circuits). FIG. 3 is a block diagram of a processing element 305 of the domain adaptive processor. Each processing element 305 includes four sub-processing elements (sub-PEs, or “sub-processors”), each of which may be a processing circuit (discussed in further detail below), including (i) a complex multiply and accumulate (CMAC) circuit 320 for performing mathematical computations such as multiplication and addition on real and complex numbers (ii) a memory circuit 325 (or “data memory” (DMEM)), which may also be referred to as intelligent storage (IS), and which may be configured to store 32-bit complex data (iii) a coordinate rotation digital compute (CORDIC) and division circuit 330 (which may be referred as a CORDIV (CORDIC + DIV) circuit), which may handle CORDIC and division operations, and (iv) a logic circuit 335, which may include gates for performing logical operations. In operation, each of the four sub-processing elements may, during each clock cycle, read a respective instruction word (which may be referred to as a “very long instruction word” (VLIW)) from instruction memory (or “program memory”), and execute it. The domain adaptive processor may be configurable to perform a range of signal processing operations including linear algebra operations such as filtering by a finite impulse response (FIR) filter, matrix multiplications, matrix inversions (e.g., by QR back substitution or Gaussian elimination), eigenvalue and eigenvector calculations, or singular value decompositions. Each of these signal processing operations, when performed by the domain adaptive processor may be readily configured to be performed over a range of matrix and vector sizes. In some embodiments, specialized accelerators -IQ- 4899-0920-1066, v. 1may be present to perform some or all of these signal processing operations instead of (or in addition to) the domain adaptive processor. FIGs. 4A, 4B, and 40 show performance parameters that may be achieved by forward error correction accelerators and (in FIG. 4C) by an FFT accelerator, in some embodiments.

[0050] As used herein, “a portion of” something means “at least some of’ the thing, and as such may mean less than all of, or all of, the thing. As such, “a portion of” a thing includes the entire thing as a special case, i.e., the entire thing is an example of a portion of the thing. As used herein, when a second quantity is “within Y” of a first quantity X, it means that the second quantity is at least X-Y and the second quantity is at most X+Y. As used herein, when a second number is “within Y%” of a first number, it means that the second number is at least (1-Y / 100) times the first number and the second number is at most (1+Y / 100) times the first number. As used herein, the word “or” is inclusive, so that, for example, “A or B” means any one of (i) A, (ii) B, and (iii) A and B.

[0051] Each of the terms “processing circuit” and “means for processing” is used herein to mean any combination of hardware, firmware, and software, employed to process or store data or digital signals. Processing circuit hardware may include, for example, one or more systems on chips (SoCs) or one or more portions of SoCs, application specific integrated circuits (ASICs), general purpose or special purpose central processing units (CPUs), digital signal processors (DSPs), graphics processing units (GPUs), programmable logic devices such as field programmable gate arrays (FPGAs), or memory chips or circuits. In a processing circuit, as used herein, each function is performed either by hardware configured, i.e., hard-wired, to perform that function, or by more general-purpose hardware, such as a CPU, configured to execute instructions stored in a non-transitory storage medium. A processing circuit may be fabricated on a single printed circuit board (PCB) or distributed over several interconnected PCBs. A processing circuit may contain other processing circuits; for example, a processing circuit may include two processing circuits, an FPGA and a CPU, interconnected on a PCB.

[0052] As used herein, when a method (e.g., an adjustment) or a first quantity (e.g., a first variable) is referred to as being “based on” a second quantity (e.g., a second variable) it means that the second quantity is an input to the method or influences the first quantity, e.g., the second quantity may be an input (e.g., the only input, or one of -11- 4899-0920-1066, v. 1several inputs) to a function that calculates the first quantity, or the first quantity may be equal to the second quantity, or the first quantity may be the same as (e.g., stored at the same location or locations in memory as) the second quantity.

[0053] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the inventive concept. As used herein, the terms “substantially,” “about,” and similar terms are used as terms of approximation and not as terms of degree, and are intended to account for the inherent deviations in measured or calculated values that would be recognized by those of ordinary skill in the art.

[0054] Any numerical range recited herein is intended to include all sub-ranges of the same numerical precision subsumed within the recited range. For example, a range of "1.0 to 10.0" or “between 1.0 and 10.0” is intended to include all subranges between (and including) the recited minimum value of 1.0 and the recited maximum value of 10.0, that is, having a minimum value equal to or greater than 1.0 and a maximum value equal to or less than 10.0, such as, for example, 2.4 to 7.6. Similarly, a range described as “within 35% of 10” is intended to include all subranges between (and including) the recited minimum value of 6.5 (i.e. , (1 - 35 / 100) times 10) and the recited maximum value of 13.5 (i.e., (1 + 35 / 100) times 10), that is, having a minimum value equal to or greater than 6.5 and a maximum value equal to or less than 13.5, such as, for example, 7.4 to 10.6. Any maximum numerical limitation recited herein is intended to include all lower numerical limitations subsumed therein and any minimum numerical limitation recited in this specification is intended to include all higher numerical limitations subsumed therein.

[0055] It will be understood that when an element is referred to as being “directly connected” or “directly coupled” to another element, there are no intervening elements present. As used herein, “generally connected” means connected by an electrical path that may contain arbitrary intervening elements, including intervening elements the presence of which qualitatively changes the behavior of the circuit. As used herein, “connected” means (i) “directly connected” or (ii) connected with intervening elements, the intervening elements being ones (e.g., low-value resistors or inductors, or short sections of transmission line) that do not qualitatively affect the behavior of the circuit.

[0056] Although exemplary embodiments of a configurable processor have been specifically described and illustrated herein, many modifications and variations will be -12- 4899-0920-1066, v. 1apparent to those skilled in the art. Accordingly, it is to be understood that a configurable processor constructed according to principles of this disclosure may be embodied other than as specifically described herein. The invention is also defined in the following claims, and equivalents thereof.-13- 4899-0920-1066, v. 1

Claims

WHAT IS CLAIMED IS:

1. A system, comprising:a plurality of processing circuits, anda network, connected to each of the processing circuits,the processing circuits comprising:a first accelerator;a second accelerator; anda general-purpose processor,a first processing circuit of the processing circuits being configured to perform a method for scheduling, the method for scheduling comprising:determining that a computational task is to be performed; and assigning the computational task to a second processing circuit of the processing circuits,the assigning of the computational task comprising assigning the computational task based on:a capability of the second processing circuit;a workload of the second processing circuit;a capability of a third processing circuit of the processing circuits; and a workload of the third processing circuit.

2. The system of claim 1 , wherein each of the first accelerator and the second accelerator is selected from the group consisting of fast Fourier transform accelerators, forward error correction encoding accelerators, forward error correction decoding accelerators, finite impulse response filtering accelerators, correlation accelerators, matrix multiply accelerators, matrix inversion accelerators, machine learning accelerators, eigensystem accelerators, singular value decomposition channel estimation accelerators, filter coefficient calculation accelerators, spatial beam former estimators, space-time beam former estimators, beam former application accelerators, likelihood estimators, symbol modulation accelerators, symbol demodulation accelerators, signal detection accelerators, signal acquisition accelerators, channel compensation accelerators, linear algebra accelerators, and combinations thereof.-14- 4899-0920-1066, v.

13. The system of claim 2, wherein the second accelerator is different from the first accelerator.

4. The system of claim 1 , wherein the first accelerator is a linear algebra accelerator, configured to perform matrix multiplications, matrix inversions, finite impulse response filtering, and eigenvalue calculations.

5. The system of claim 4, wherein the linear algebra accelerator comprises an array of heterogeneous sets of processing circuits.

6. The system of claim 1 , wherein the network has a bus width of at least 64 bits.

7. The system of claim 1 , wherein the assigning comprises selecting a set of instructions for execution, the selecting being based on:the computational task, andthe second processing circuit.

8. The system of claim 1 , wherein the system is capable of operating with a throughput of 1 gigabyte per second while consuming less than 10 watts of electrical power.

9. The system of claim 8, wherein the system is further capable of performing 10 giga operations per second, each operation having two 32-bit operands.

10. The system of claim 1 , further comprising a radio-frequency (RF) front end, connected to the network, the RF front end being configured to transmit or receive analog RF signals.

11. The system of claim 10, wherein the RF front end comprises a buffer.-15- 4899-0920-1066, v.

112. The system of claim 1 , wherein the processing circuits further comprise a cache, connected to the network.

13. The system of claim 12, comprising a system on a chip comprising the plurality of processing circuits and the network, wherein the system further comprises an off-chip memory.

14. The system of claim 13, wherein the wherein the processing circuits further comprise one or more interface circuits.

15. The system of claim 14, wherein the network is configured to transmit data from any one of the processing circuits to any other one of the processing circuits.

16. A method, comprising performing a method for scheduling, by a first processing circuit of a plurality of processing circuits of a system, the system further comprising a network connected to each of the processing circuits, the method for scheduling comprising:determining that a computational task is to be performed; and assigning the computational task to a second processing circuit of the processing circuits,the assigning of the computational task comprising assigning the computational task based on:a capability of the second processing circuit;a workload of the second processing circuit;a capability of a third processing circuit of the plurality of processing circuits; anda workload of the third processing circuit.

17. The method of claim 16, wherein the processing circuits comprise:a first accelerator; anda second accelerator, different from the first accelerator.-16- 4899-0920-1066, v.

118. The method of claim 17, wherein the first accelerator is a linear algebra accelerator, configured to perform matrix multiplications, matrix inversions, finite impulse response filtering, and eigenvalue calculations.

19. The method of claim 18, wherein the linear algebra accelerator comprises an array of heterogeneous sets of processing circuits.

20. The method of claim 16, wherein the network has a bus width of at least 64 bits.-17- 4899-0920-1066, v. 1