SYSTEM AND METHOD FOR MULTI-CHIP ISING MACHINE ARCHITECTURE
The scalable Ising machine architecture with reconfigurable chips and interconnects addresses the capacity limitations of existing machines, enabling efficient solution of larger problems with significant speedup and energy efficiency.
Patent Information
- Application Number
- JP2024529216
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-11-22
- Filing Date
- 2022-11-22
- Publication Date
- 2024-11-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing Ising machines have a fixed problem-solving capacity, limiting their ability to handle larger problems effectively, and divide-and-conquer strategies using these machines result in significant performance degradation due to glue computations and reprogramming overheads.
A scalable Ising machine architecture comprising multiple chips with reconfigurable resistor connections and interconnects, allowing independent or collective operation, and a method for computing the Hamiltonian of coupled spin systems by segmenting the problem into subsystems and calculating them across multiple chips.
Enables efficient solution of larger problems with high performance, achieving speedups of over 6 orders of magnitude compared to sequential solvers and 2000x faster than state-of-the-art compute accelerators, while maintaining energy efficiency.
Smart Images

Figure 2024541079000001_ABST
Abstract
Description
[Technical field]
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Application No. 63 / 281,944, filed November 22, 2021, which is incorporated by reference in its entirety.
[0002] STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT This invention was made with Government support under HR00112090012 awarded by the Defense Advanced Research Projects Agency. The United States Government has certain rights in this invention. [Background technology]
[0003] Nature has inspired many problem-solving techniques for decades. Recently, researchers have increasingly turned to using nature to solve problems directly. Ising machines are a good example, with numerous research prototypes and many design concepts. Ising machines can map out families of NP-complete problems and derive competitive solutions at speeds much faster than traditional algorithms, in some cases at a fraction of the energy cost of a von Neumann computer.
[0004] However, the problem-solving capacity of a physical Ising machine is often fixed. Without support, larger problems cannot be solved at all. The state of the art is a software-based divide-and-conquer strategy: problems with more than N spin variables are transformed into a series of subproblems with N or fewer spins and fed into the machine. As far as the Ising machine is concerned, the problem also fits into the capacity of the machine. As will be explained further herein, with this type of support, the advantage of using a high-performance Ising machine quickly disappears. The degradation is caused by fundamentals and cannot be improved by software optimization.
[0005] We disclose a machine architecture that is fundamentally scalable. In other words, each machine can act independently to solve simple problems, or collectively in groups to solve larger problems. In the latter mode, the machine is explicitly aware of external spins and implements both intra-chip and inter-chip spin coupling. Summary of the Invention
[0006] In one aspect, a scalable Ising machine system includes a plurality of chips, each chip including a plurality of N nodes, each node including a capacitor, a positive terminal, and a negative terminal; a plurality of N×M connection units arranged in N rows and M columns, each connection unit including a set of reconfigurable resistive connections, each connection unit being configurable to connect a pair of the N nodes via the reconfigurable resistive connections; and a plurality of interconnects, each chip of the plurality of chips communicatively connected to every other chip of the plurality of chips via at least one interconnect.
[0007] In one embodiment, the plurality of chips are arranged in a two-dimensional array. In one embodiment, the plurality of chips are arranged in a three-dimensional array. In one embodiment, the plurality of chips are arranged in at least one square array. In one embodiment, at least one interconnect of the plurality of interconnects comprises a wireless data connection. In one embodiment, each chip further comprises a data buffer configured to digitally store state information of at least a subset of the N nodes. In one embodiment, N=M.
[0008] In one embodiment, each connection unit includes two positive terminals, each connected to a positive terminal of a different node in the plurality of nodes, and two negative terminals, each connected to a negative terminal of a different node in the plurality of nodes. In one embodiment, at least one interconnect of the plurality of interconnects includes a switch configured to connect or disconnect the interconnect. In one embodiment, the at least one chip further includes a reconfigurable connection fabric for connecting the nodes. In one embodiment, each chip further includes a buffer memory, a processor, and a non-transitory computer-readable medium having instructions stored thereon that, when executed by the processor, store a node state in the buffer memory and retrieve a node state from the buffer memory.
[0009] In one embodiment, the instructions further include transmitting the node states in sequence from one chip to the next chip to execute a larger task in batch mode. In one embodiment, each buffer memory is sufficient to store a buffered copy of at least a subset of states in the scalable Ising machine system. In one embodiment, each buffer memory is sufficient to store a buffered copy of all states in the scalable Ising machine system.
[0010] In one aspect, a method for computing a Hamiltonian of a coupled spin system includes providing a scalable Ising machine system having a plurality of chips, each chip having a plurality of N nodes, each node having a capacitor, a positive terminal, and a negative terminal, where a charge on the capacitor represents spin, and a plurality of N×M connection units arranged in N rows and M columns, each connection unit having a set of reconfigurable resistive connections, each connection unit being configurable to connect a pair of the N nodes via the reconfigurable resistive connections; connecting the plurality of chips to each other via a set of interconnects; segmenting the coupled spin system into a set of subsystems, configuring each chip of the plurality of chips with a subsystem of the set of subsystems; and computing a Hamiltonian of the coupled spin system by computing all the subsystems.
[0011] In one embodiment, the method includes computing the subsystems at least partially in sequence. In one embodiment, the method includes computing the subsystems concurrently. In one embodiment, the method includes storing states of at least a subset of the nodes in a buffer memory. In one embodiment, the method includes transmitting the subset of the node states from one chip to another chip. In one embodiment, the method includes storing states of all the nodes in a buffer memory on each chip of a plurality of chips.
[0012] The above and other objects and features will become apparent from the following description and by reference to the accompanying drawings, which are included to provide an understanding of the invention and are incorporated in and constitute a part of this specification, and in which like numerals represent like elements. [Brief description of the drawings]
[0013] [Figure 1] 1 is an exemplary computing device. [Diagram 2] A and B are graphs of speedup versus graph size. [Diagram 3] FIG. 2 is a diagram of an exemplary multi-node chip. [Figure 4] FIG. 2 is a schematic diagram of a coupling unit. [Figure 5A] FIG. 1 is a diagram of an exemplary multi-chip architecture. [Figure 5B] FIG. 1 is a diagram of an exemplary multi-chip architecture. [Figure 6] FIG. 13 is an exemplary diagram of a matrix mapping. [Figure 7] 1 is an exemplary multi-chip architecture. [Figure 8] FIG. 1 illustrates an example reconfigurable multi-node architecture. [Figure 9] FIG. 1 is a diagram of an exemplary multi-chip architecture arranged in a three-dimensional array. [Figure 10] A and B are graphs of the surprise over ignorance energy for different epoch sizes. [Figure 11] FIG. 1 illustrates an example batch processing architecture. [Figure 12] 1 is a graph of execution times for various architectures. [Figure 13] 1 is a graph of execution times for various architectures. [Figure 14A] 1 is a graph of flips and bit changes over execution time. [Figure 14B] 13 is a graph of average flips and bit changes over different epoch sizes. [Figure 15] A and B are graphs of the computation results over execution time. [Figure 16A] 13 is a graph of induced spin flip over run time. [Figure 16B] 1 is a graph of the average percentage of induced spin flips over epoch time. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0014] It should be understood that the drawings and description of the present invention have been simplified to illustrate elements relevant to a clear understanding of the present invention, but for clarity, many other elements found in related systems and methods have been omitted. Those skilled in the art may recognize that other elements and / or steps are desirable and / or necessary in implementing the present invention. However, because such elements and steps are well known in the art and do not facilitate a better understanding of the present invention, descriptions of such elements and steps are not provided herein. The disclosure herein covers all variations and modifications to such elements and methods known to those skilled in the art.
[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, exemplary methods and materials are described below.
[0016] As used herein, each of the following terms has the meaning associated with it in this section.
[0017] As used herein, the articles "a" and "an" refer to one or to more than one (i.e., to at least one) of the grammatical object of the article. By way of example, "an element" means one or more elements.
[0018] As used herein, "about" when referring to a measurable value, such as an amount, a temporal duration, and the like, is meant to encompass variations of ±20%, ±10%, ±5%, ±1%, and ±0.1% from the particular value, as such variations are appropriate.
[0019] Throughout this disclosure, various aspects of the invention can be expressed in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the invention. Thus, the description of a range should be considered to have specifically disclosed all possible narrower ranges and individual numerical values within that range. For example, the description of a range such as 1-6 should be considered to have specifically disclosed the narrower ranges such as 1-3, 1-4, 1-5, 2-4, 2-6, 3-6, etc., as well as the individual numerical values contained within that range, such as 1, 2, 2.7, 3, 4, 5, 5.3, 6, and all whole and partial increments therebetween. This applies regardless of the breadth of the range.
[0020] In some aspects of the present invention, software executing the instructions provided herein may be stored on a non-transitory computer readable medium, which when executed on a processor performs some or all of the steps of the present invention.
[0021] Aspects of the invention relate to algorithms executed in computer software. Although certain embodiments may be described as being written in a particular programming language or executed on a particular operating system or computing platform, it is understood that the systems and methods of the invention are not limited to any particular computing language, platform, or combination thereof. Software executing the algorithms described herein may be written in, compiled, or interpreted in any programming language known in the art, including, but not limited to, C, C++, C#, Objective-C, Java, JavaScript, MATLAB, Python, PHP, Perl, Ruby, or Visual Basic. Furthermore, it is understood that elements of the invention may be executed on any acceptable computing platform, including, but not limited to, a server, a cloud instance, a workstation, a thin client, a mobile device, an embedded microcontroller, a television, or any other suitable computing device known in the art.
[0022] Some of the present invention are described as software running on a computing device. Although the software described herein may be disclosed as running on one particular computing device (e.g., a dedicated server or workstation), it is understood in the art that software is inherently portable, and that most software running on a dedicated server may also be run, for the purposes of the present invention, on any of a wide range of devices, including desktop or mobile devices, laptops, tablets, smartphones, watches, wearable electronics or other wireless digital / cellular phones, televisions, cloud instances, embedded microcontrollers, thin client devices, or any other suitable computing devices known in the art.
[0023] Similarly, portions of the invention are described as communicating over various wireless or wired computer networks. For purposes of the present invention, the terms "network," "networked," and "networking" are understood to encompass wired Ethernet, fiber optic connections, wireless connections including any of the various 802.11 standards, cellular WAN infrastructures such as 3G, 4G / LTE, or 5G networks, Bluetooth, Bluetooth Low Energy (BLE), or Zigbee communications links, or any other manner in which one electronic device can communicate with another electronic device. In some embodiments, elements of the networking portion of the invention may be implemented via a Virtual Private Network (VPN).
[0024] 1 and the following discussion are intended to provide a brief, general description of a suitable computing environment in which the invention may be implemented. Although the invention has been described above in the general context of program modules executing in combination with application programs running on a computer's operating system, those skilled in the art will recognize that the invention may also be implemented in combination with other program modules.
[0025] Generally, program modules include routines, programs, components, data structures, and other types of structures that perform particular tasks or implement particular abstract data types. Moreover, those skilled in the art will appreciate that the invention may be practiced with other computer system configurations, including handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframe computers, etc. The invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.
[0026] 1 illustrates an exemplary computer architecture of a computer 100 for implementing various embodiments of the present invention. The computer architecture illustrated in FIG. 1 illustrates a conventional personal computer including a central processing unit 150 ("CPU"), a system memory 105 including random access memory 110 ("RAM") and read only memory ("ROM") 115, and a system bus 135 coupling the system memory 105 to the CPU 150. A basic input / output system containing basic routines that help to transfer information between elements within the computer, such as during start-up, is stored in the ROM 115. The computer 100 further includes an operating system 125, applications / programs 130, and a storage device 120 for storing data.
[0027] Storage device 120 is connected to CPU 150 through a storage controller (not shown) that is connected to bus 135. Storage device 120 and its associated computer-readable media provide non-volatile storage for computer 100. Although the descriptions of computer-readable media contained herein refer to storage devices such as hard disks or CD-ROM drives, those skilled in the art should understand that computer-readable media can be any available media that can be accessed by computer 100.
[0028] By way of example, and not limitation, computer readable media may include computer storage media. Computer storage media includes volatile and nonvolatile, removable and fixed media implemented in any method or technology for storing information such as computer readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EPROM, EEPROM, flash memory or other solid-state memory technology, CD-ROM, DVD or other optical storage devices, magnetic cassettes, magnetic tapes, magnetic disk storage devices or other magnetic storage devices, or any other medium that can be used to store desired information that can be accessed by a computer.
[0029] According to various embodiments of the present invention, computer 100 may operate in a networked environment using logical connections to remote computers through a network 140, such as a TCP / IP network such as the Internet or an intranet. Computer 100 may connect to network 140 through a network interface unit 145 connected to bus 135. It should be appreciated that network interface unit 145 may also be utilized to connect to other types of networks and remote computer systems.
[0030] Computer 100 may also include an input / output controller 155 for receiving and processing input from a number of input / output devices 160, including a keyboard, a mouse, a touch screen, a camera, a microphone, a controller, a joystick, or other types of input devices. Similarly, input / output controller 155 may provide output to a display screen, a printer, speakers, or other types of output devices. Computer 100 may be connected to input / output devices 160 via a wired connection, including, but not limited to, optical fiber, Ethernet, or copper wire, or via wireless means, including, but not limited to, Wi-Fi, Bluetooth, Near Field Communication (NFC), infrared, or other suitable wired or wireless connection.
[0031] As briefly mentioned above, a number of program modules and data files may be stored in the storage device 120 and / or RAM 110 of the computer 100, which includes an operating system 125 suitable for controlling the operation of a networked computer. The storage device 120 and RAM 110 may also store one or more applications / programs 130. In particular, the storage device 120 and RAM 110 may store applications / programs 130 for providing a variety of functions to a user. For example, the applications / programs 130 may include many types of programs, such as word processing applications, spreadsheet applications, desktop publishing applications, database applications, gaming applications, Internet browsing applications, email applications, messaging applications, and the like. According to one embodiment of the present invention, the applications / programs 130 include a multi-function software application for providing word processing functionality, slide presentation functionality, spreadsheet functionality, database functionality, and the like.
[0032] In some embodiments, computer 100 may include various sensors 165 for monitoring the environment around and within computer 100. These sensors 165 may include global positioning system (GPS) sensors, photosensitive sensors, gyroscopes, magnetometers, thermometers, proximity sensors, accelerometers, microphones, biometric sensors, barometers, humidity sensors, radiation sensors, or any other suitable sensors.
[0033] Disclosed herein, in various embodiments, are 3D integrated multi-chip Ising machines, a macro-chip version in which multiple "chiplets" are integrated on a carrier, a system based on a universal digital interconnection of several reconfigurable Ising machines, and various optimization techniques for the interconnected Ising machines.
[0034] Examples of computing tasks performed in nature include solving differential equations, performing random sampling, etc. Some of these natural processes are already being exploited; for example, transistors, which can be turned on and off, are the basis of most computers today. However, this is different from exploiting nature's computational power at a higher level, for example to solve an entire problem. In fact, some very powerful algorithms have been inspired by nature (S. Kirkpatrick, et al., Science, vol. 220, 1983; G. Zames, et al., Information Technology Journal, 1981; and DH Ackley, et al., Cognitive science, 1985).
[0035] It is not difficult to imagine that a given problem can be solved more quickly and efficiently when the computational substrate is based on nature than by mapping it onto a von Neumann architecture. One area where rapid progress has been made in such efforts is the Ising machine.
[0036] Simply put, an Ising machine seeks low energy states for a coupled spin system. Some problems (indeed, all NP-complete problems) can be expressed as an equivalent optimization problem in Ising formulas, as detailed further below. Existing Ising machines, most of which are prototypes or concepts, have already shown promise for good performance and energy efficiency for certain problems.
[0037] In some cases, when the size of the problem exceeds the capacity of the machine, it may become impossible to map the problem to a particular piece of hardware. Intuitively, some form of divide-and-conquer would allow a larger problem to be broken down into smaller subproblems that can be mapped to multiple instances of a given piece of hardware, and thus benefit from further Ising machine acceleration. In one existing example, the algorithm employed by D-Wave (M. Booth, et al, Technical Report, 2017) shows that the effective speedup of a system employing such a divide-and-conquer strategy decreases rapidly as the problem size increases. For example, a 500-node machine can reach a speedup of 600,000 over a von Neumann solver (simulated annealing), but using the same machine to solve a 520-node problem achieves only a speedup of 250.
[0038] This disclosure includes a discussion of the problems presented by simple divide-and-conquer strategies. It is shown that such strategies are essentially limited in performance by "glue" computation. There is therefore a need in the art for a machine designed from the ground up to eliminate the need for such glue. In some embodiments, a hardware design is disclosed herein that can achieve this goal. Finally, experimental data is presented that shows that this design can indeed scale to larger problems while maintaining high performance, achieving over 6 orders of magnitude speedup compared to sequential solvers and over 2000x speedup compared to state-of-the-art computational accelerators.
[0039] Principle of the Ising machine The Ising model is used to describe the Hamiltonian of a coupled spin system. Spin has one degree of freedom and can take one of two values (+1, -1). The energy of the system is determined by the pairwise coupling of the spins (J ij =J ji ) and the interaction of some external field (μ) with each spin (h i ) and the resulting Hamiltonian is:
number
[0040] Given such an equation, we can state the minimization problem of which state of the system ([σ1,σ2,...]) has the lowest energy. A physical system with such a Hamiltonian will naturally tend towards low-energy states. It is as if nature is always trying to solve the minimization problem, but this is never easy.
[0041] Indeed, the cardinality of the state space grows exponentially with the number of spins, and the optimization problem is NP-complete, which can be easily transformed into a generalized max-cut problem, which is part of the original list of NP-complete problems (RM Karp, Springer US, 1972, pp. 85-103).
[0042] Therefore, the physical system of spins somehow has programmable coupling parameters (J in Eq. 1). ij and μh i), this can be used as a dedicated computer to solve optimization problems that can be expressed in Ising equations (Eq. 1). In fact, all problems in the Karp NP-complete set have their Ising equations derived (A. Lucas, Frontiers in Physics, 2014). Also, if a problem already has a QUBO (quadratic unconstrained binary optimization) equation, mapping it to an Ising equation is as simple as substituting bits for spins, i.e., σ i =2b i The result is -1.
[0043] Due to the broad class of problems that can be mapped to Ising equations, building nature-based computing systems to solve these problems has attracted considerable attention. Roughly speaking, the design of an Ising machine involves four steps:
[0044] (1) Identify the physical variable that represents spin, whether it be a quantum bit (R. Harris, et al., Phys. Rev. B, Jul 2010); the phase of an optical pulse (T. Inagaki, et al., Science, 2016); or the polarity of a capacitor (R. Afoakwa, et al., 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2021).
[0045] (2) Identify the coupling mechanism and how to program the coefficients.
[0046] (3) Demonstrate problem-solving capabilities both through its theory of operation (revealing nature's "invisible hand") and through satisfactory practical results.
[0047] (4) Demonstrate superior machine metrics (time to solution, energy consumption, and / or build costs).
[0048] Importantly, different approaches may offer different fundamental tradeoffs. It is therefore premature to evaluate general approaches based on observed prototypic cases. Nevertheless, disclosed herein is a rough characterization that can help those skilled in the art get a basic sense of the landscape, so long as the caveats are properly understood. This characterization is by no means comprehensive. In particular, for conceptual clarity, we treat the numerous designs that use GPUs, FPGAs, or ASICs to accelerate von Neumann-type algorithms (simulated annealing or its variants) as accelerated simulated annealers, rather than as physical Ising machines.
[0049] Quantum annealing (QA) differs from adiabatic quantum computing (AQC) in that it relaxes the adiabaticity requirement (S. Boixo, et al., Nature physics, 2014). Although QA technically includes AQC as a subset, the current D-Wave system is not adiabatic. In other words, it has no theoretical guarantee that the ground state will be reached. Without the ground state guarantee, the Ising physics of qubits has no other known advantages over alternatives. And it can be argued that using quantum devices to represent spin is probably suboptimal. First, these devices are much more sensitive to noise and therefore require cryogenic operating conditions that consume most of the 25kW operating power. Second, coupling qubits can be difficult compared to coupling other spin forms, which explains why current machines use local coupling networks. As a result, for common graph topologies, the number of nodes required for these locally coupled machines grows quadratically, with the nominal 2000 nodes on the D-Wave 2000q corresponding to only about 64 effective nodes (R. Hamerly, et al., Science Advances, 2019; R. Hamerly, et al., D-Wave 2000Q arXiv, 2018).
[0050] Coherent Ising machines (CIMs) can be considered as a second-generation design that addresses several issues. In (T. Inagaki, et al., Science, 2016), all 2000 nodes can be interconnected, making it clearly the most powerful physical Ising machine to date. An exemplary CIM can operate at room temperature and consumes only about 200 W of power because it uses special light pulses that act as spins. However, the pulses must be contained in a 1 km long optical fiber, and the system requires strict temperature stability, making it difficult to maintain a stable operating state for many spins. Efforts to scale beyond 2000 nodes have not yet been successful.
[0051] The working principle of the CIM can be observed by the Kuramoto model (Y. Takeda, et al., Quantum Science and Technology, Nov 2017), so theoretically the same goal can be achieved using other oscillators. This has led to the development of a number of electronic oscillator-based Ising machines (OIMs), which can be considered the third generation. These systems use LC tanks for the spin and (programmable) resistors as coupling units. Technically, the phase of the oscillator corresponds to a two-degree-of-freedom spin across the XY plane (instead of one degree of freedom up or down as in the Ising model). As a result, additional mechanisms are required to solve the Ising equation problem by imposing constraints, such as subharmonic injection locking (SHIL) (K. Cho, et al., International conference on artificial neural networks. Springer, 2011). These electronic oscillator-based Ising machines are a significant improvement over previous designs in terms of mechanical metrics. Indeed, their exact power consumption and operating speed can range over orders of magnitude, depending on the exact inductances and capacitances chosen. However, it is not difficult to achieve desktop sizes with power consumption of around 1-10W, which is a significant improvement over cabinet-sized machines with power consumption of 200W-25kW. However, in on-chip integration, inductors often cause practical challenges. For example, they are area intensive and have undesirable parasitics such as lower quality factor and increased phase noise, all of which cause practical challenges in maintaining frequency uniformity and phase synchrony among thousands of on-chip oscillators. Another electronic design with a different architecture is the BISM (Bistable Resistively-coupled Ising Machine) (R. Afoakwa, et al., 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2021).Additional information regarding BRIM can be found in PCT Application No. PCT / US2021 / 070402, filed April 16, 2021, the entire contents of which are incorporated herein by reference.
[0052] In BRIM, spins are implemented as capacitors whose voltages are controlled by a feedback circuit that is made bistable. The design is CMOS compatible and uses voltages (rather than phases) to represent spins, allowing for easy interfacing to additional architectural support for computational tasks. Therefore, certain embodiments disclosed herein use a baseline substrate similar to BRIM. It is noted that the same principles herein can be directly applied to all Ising machines with different amounts of glue logic.
[0053] For most current (physical) Ising machines, significant speedup and energy gains can be expected compared to von Neumann computing when the problem fits into the hardware. However, there is little discussion of what happens when the problem exceeds the capacity of the machine. It is understandable to assume that the machine can still accelerate in proportion to the fraction of the problem that can be mapped. As will be discussed further herein, in reality, little or no gain is expected for problems that exceed the capacity of the machine.
[0054] We first discuss the approach taken by D-Wave's system. As it is the only commercially available Ising machine platform, its solutions are state-of-the-art and the baseline for any comparison. We then start from basic principles and discuss the details of the divide-and-conquer strategy, followed by the issue that the subproblems to be solved are not strictly independent of each other, making parallelization difficult.
[0055] D-Wave's tools (M. Booth, et al., Technical Report, 2017) allow any Ising machine to be used to solve problems that exceed its hardware capacity. To check the performance of such a system (the algorithm is reproduced in the appendix for the reader's convenience), the BRIM model is used as the Ising machine. Although the general strategy should work for any Ising machine, BRIM offers many practical advantages to the disclosed work. First, it provides all-to-all connectivity. This means that a machine with n nodes can map any graph of n nodes. Many machines provide a large number of nominal nodes, but only neighborhood connections (see R. Harris, et al., Phys. Rev. B, Jul 2010; M. Yamaoka, et al., 2015 IEEE International Solid-State Circuits Conference-(ISSCC) Digest of Technical Papers, Feb 2015; and T. Takemoto, et al., IEEE International Solid-State Circuits Conference, February 2019). A typical n-node graph requires O(n 2 ) join parameters. Thus, to map such a graph, a locally join machine requires O(n 2 ) node is required.
[0056] Second, although architectural support for scalable Ising machines is explored later, BRIM's CMOS compatibility offers significant design flexibility.
[0057] In Figures 2A and 2B, the speedup of a 500-node Ising machine is shown as the problem size grows beyond the capacity of the machine. With reference to Figures 2A and 2B, the speedup of two dnc algorithms (D-Wave and others disclosed herein) using a 500-spin BRIM as well as a supporting sequential computer is shown. Figure 2A shows the speedup for all graphs tested. Figure 2B zooms in on the segment from graph size 500 to 520.
[0058] The details of the measurements are ignored as they do not affect the qualitative lesson. Two things can be seen from this figure. First, the speedup of the Ising machine improves when the problem gets larger but still fits within the machine. Clearly, a larger hardware capacity is desirable. However, the figure also shows a second, and perhaps more important, point: the speedup crashes sharply as soon as the problem exceeds the hardware capacity (Fig. 2B). It is not surprising that the speedup drops off. What is surprising is how much and how quickly. Naturally, part of this is due to the specific implementation of the D-Wave tool. Thus, an alternative was made, but the result (shown in Fig. 2A) is only a slight improvement. The sharp drop in Fig. 2A is due to the underlying reasons discussed next.
[0059] Users interact with the D-Wave system using the Solver API (SAPI) over the network. Jobs are submitted to the SAPI server queue. The jobs are then assigned to workers, which, once executed on a classical processor, are responsible for submitting instructions to the quantum processor, receiving results from the quantum processor, post-processing the results if necessary, and sending the results back to the user.
[0060] The general idea of the algorithm (see Algorithm 1 below) is straightforward: if some part of the state remains fixed, the original QUBO problem is transformed into a smaller subproblem (subQubo in line 15) that can be fed into the solver (line 18). Repeating this action for different parts of the state vector (lines 15 to 21) constitutes one pass of the algorithm. To obtain better results, multiple passes are performed (a while loop is started in line 14). Algorithm 2 below shows the disclosed and improved approach, which is more efficient. Algorithm 1 D-WAVE divide and conquer algorithm 1: Input: QUBO instance 2: #V best, lowest value found to date 3: #Qbest, solution bit vector corresponding to the lowest value so far 4: #index, indices of the bits in the solution 5: 6: # Get initial estimate of minimum value and backbone 7: Qtmp ← random 0 / 1 vector 8: (Vbest, Qbest) ← TabuSearch(QUBO, Qtmp) 9: index ← OrderByImpact(QUBO, Qbest) 10: passCount ← 0 11: Qtmp ← Qbest 12: total ← fraction * size(QUBO) 13: 14: while passCount < numRepeats do 15: for i = 0;i < total;i = i + subQuboSize do 16: # select subQubo with other variables clamped 17: subQubo ← Clamp(QUBO;Qtmp;index[i : i +subQuboSize - 1]) 18: (subV, subQ) ← DWaveSearch(subQubo) 19: # project onto full solution 20: Qtmp[index[i : i + subQuboSize - 1]] ← subQ 21: End for 22: (V, Qnew) ← TabuSearch(QUBO, Qtmp) 23: index ← OrderByImpact(QUBO, Qnew) 24: if V < Vbest then 25: Vbest ← V;Qbest ← Qnew 26: passCount ← 0 27: else if V == Vbest then 28: Qbest ← Qnew 29: passCount++ 30: else 31: passCount++ 32: end if 33: Qtmp ← Qnew 34: end while 35: Output: Vbest;Qbest Algorithm 2: An alternative divide-and-conquer algorithm 1: Input: Graph 2: #V ← random 0 / 1 spin vector 3: # numSolvers ← number of solvers 4: # numRepeats ← number of times to repeat 5: # epoch ← Epoch times for each solver 6: # ratio ← ratio in which the graph is to be partitioned 7: 8: (subG, subV) ← RandPartition(Graph, V, ratio) 9: ene ← IsingEnergy(subG, subV) 10: for r = 0;r < numRepeats;++r do 11: for i = 0;i < numSolvers;++i do 12: # Launch solvers (Can be parallel) 13: Solver(subG[i];subV [i];epoch[i];ene[i]) 14: end for 15: Synchronize(subG;subV;ene) 16: end for 17: 18: V ← copy(subV) 19: FinalEne ← IsingEnergy(Graph;V ) 20: Output: V;FinalEne
[0061] We now explain the principle of divide and conquer in Ising optimization problems. The problem of minimizing Equation 1 above is best described as navigating a (high-dimensional) energy landscape to find the lowest valley. The idea is to keep some dimensions fixed (e.g. longitude) and navigate along the remaining dimensions looking for a better spot. (Many solvers can be explained with this analogy.) This is the essence of the divide and conquer strategy. This point (and the problem) is clearly and explicitly stated below. Here, the concept of matrices is more useful. Equation 1 can be rewritten as: H=--σ T Jσ--μh T σ formula 2 where σ = [σ1, ..σ n ] T , J = │J ij │ n×n , and h=[h1,..,h n ] T where J is a symmetric matrix with zeros on the diagonal. If the n-node problem is split into two subproblems of k nodes and nk nodes, equation 2 can be rewritten as:
number
number
[0062] This rewriting produces a larger square matrix with upper and lower submatrices J u and J. l (both squares), and "cross terms" (J × and its transpose), the effect of the cross terms can be decomposed into the original bias (h u and h l ) and the new bias (respectively g u and g l ) From this perspective, an n-node Ising optimization problem can always be decomposed into k and nk-node subproblems, and by transitivity into a combination of subproblems of any size.
[0063] Equation 3 not only shows the principle of decomposition, but also clearly shows the problem of decomposition. In the original problem, J and h are parameters and do not change. After decomposition, the bias of the upper partition (g u) is a function of the state of the lower partition. This means that the two partitions are not independent. In other words, strictly speaking, the subproblems need to be solved in sequence: if the search changes the current state of the upper partition, the parameters of the lower partition need to be updated to reflect that change before starting the search in the lower partition. Also, partitioning does not reduce the total workload. Therefore, it is not surprising that there is no parallel version of canonical simulated annealing.
[0064] If you are trying to solve a problem that is larger than the capacity of the machine, the issue may seem irrelevant. After all, if the larger problem can be decomposed into two parts (say A and B), where A fits into the Ising machine, and a speedup can be expected from the processing of A, even if the processing of A and B cannot overlap. The argument is correct. But in practice, there are several subtle issues that have serious consequences. Two related ones are discussed below.
[0065] First, as already shown, decomposition requires reprogramming because the formulation of the subproblems constantly changes. For many Ising machines, reprogramming is a costly operation that may take longer than solving the problem. To take a perhaps extreme example, a D-Wave has a programming time of 11.7 ms, while a typical run has a total of 240 μs for the remaining steps. Note that a typical (if not universal) usage pattern for these Ising machines is to program once and anneal many times (e.g., 50 times) from different initial conditions to obtain the best result. With such a usage pattern, the long programming time is amortized over many annealing runs. A decomposed problem may require many reprogramming runs within a single annealing run.
[0066] Second, even if the costs of reprogramming are somehow addressed, Amdahl's law must still be considered. Using the concrete example of BRIM (Figure 2A), the speedup of solving a 500-node problem on a sequential computer is about 10 5 Consider using the disclosed algorithm to decompose the 510-node problem into a 500-node subproblem that is mapped to hardware, as well as the remainder of the problem, the glue computation, left in software. The software workload, measured in instructions, is about 0.13% of the software workload of the original 510-node problem. Much of the remaining software workload is "glue" (a new bias, g u =μh u +J × σ l and g l =μh l +J × T σ u ), which differs from the original solver. As a result, Amdahl's Law does not apply directly. Nevertheless, it is possible to estimate an upper limit on the speedup to be around 700x.
[0067] While some of the simplified analysis disclosed aims to explain the core of the issue from first principles without including all the nuances, the more important points are quite clear and are summarized below.
[0068] In principle, the problem formulation clearly allows for the decomposition of larger problems, but the smaller component problems are not independent. As a result, relying on von Neumann computing to glue together multiple Ising machines is a fundamentally flawed strategy, as it severely limits the acceleration of problems just beyond its capacity. These machines need to be used in tandem and designed from the ground up to address the decomposition bottleneck.
[0069] General analysis of scaling of Ising machines The core of an Ising machine contains two types of components: nodes and coupling units. As already discussed above, coupling units have coupling strength J ij as input to an optimization problem, and the dynamical system evolves under some form of controlled annealing before reading out the state of each of the individual spins as a solution to the problem. i σ i is σ i Add a fixed spin (σ n+1 =+1)
number
[0070] Unless the problem has some particular topology, any spin can be coupled to any other spin. Thus, there are many more coupling parameters (O(N)) than spins (O(N)). 2 )). However, most existing Ising machines use machine architectures in which only nearby spins are allowed to couple, resulting in a system containing O(N) coupled units and O(N) spins. Using special software tools, we first transformed the original problem into a form that obeys the constraints imposed by the machine architecture. Thus, roughly speaking, these O(N) coupled units are
number
[0071] Recently, several electronic Ising machines have been proposed (T. Wang et al., 2019; T. Wang, et al., Proceedings of the 56th Annual Design Automation Conference 2019; J. Chou, et al., Scientific Reports, 2019). Although the main working principles are similar, there are significant technical differences at a deeper level. The baseline is the BRIM, where an array of N bistable nodes is interconnected by an array of NxN resistive coupling units. Referring to Figures 3 and 4, a block diagram of the BRIM is shown showing the node capacitors (305), coupling resistors (303, 304), and parallel / anti-parallel connections. The coupling units are programmed by an array of DACs before the system starts annealing. The value of the coupling resistor between nodes i and j is 1 / J ij In other words, stronger coupling means lower resistance. The sign of the coupling strength can be implemented either in the parallel connections (303, 402) or in the anti-parallel connections (304, 403). Specifically, the coupling parameter J ij is negative, the two nodes are connected in an anti-parallel (304) fashion (e.g., the anode plate of 302a is connected to the cathode plate of 302b, and vice versa). This encourages the two nodes to be of opposite polarity, so the coupling (J ij σ i σ j ) contributes to a decrease in the total energy. Conversely, J ij is positive, plates of the same polarity are coupled through a resistor (e.g., the positive plate of 302a is connected to the positive plate of 302c, and vice versa).
[0072] These all-electron Ising machines can be analyzed as dynamical systems, and Lyapunov stability analysis shows in a more theoretical way why they tend towards low-energy states. However, for the purposes of this disclosure, a more intuitive discussion using an example situation will suffice. If the system is in a particular state and has one spin (e.g., σ k=1) is assumed to be a "mistake". That is, the spin is "flipped" (σ k =+1), energy is improved or decreased. This means that
number
[0073] J jk is represented by the coupling resistor between nodes j and k, and σ j Let us consider its expression (V j , the voltage at node j), the term
number
number
[0074] Macrochip Architecture Referring to FIG. 5A, a block diagram of a "macrochip" k×k array (k=2) of BRIM chips is shown. Switches allow each chip to either operate independently or work together as a larger Ising machine. As shown in the inset, the coupling units of the top right chip are connected either to nodes on the chip itself or to wires connecting them to the same row of coupling units in adjacent chips. In the context of FIG. 5A, the CU is the coupling unit and the PU is the programming unit.
[0075] k 2 The chips (502, 503, 504, 505) are connected to each other to obtain (kN) 2 A larger machine can be formed with kN connection units 513. The wires of a row of a coupling unit are coupled to corresponding wires of the same row from the left and / or right adjacent chips. Similarly, wires of the same column from the top and bottom neighbors are coupled. If we ignore the packaging issues of the individual chips, we can simply view the entire circuit area as one (larger) "macrochip" Ising machine with kN nodes.
[0076] Referring to FIG. 5B, an alternative schematic diagram of a "macrochip" is shown, also having four chips 502-505, but in this embodiment having only one node 510 per row, and each chip having one programming unit 512 to control all the coupling units on the chip.
[0077] Given an Ising machine with N nodes, the disclosed system can solve multiple smaller problems simultaneously, as long as the sum of the number of nodes from each problem does not exceed N. This can be seen in the diagram shown in Figure 6. If the shaded regions of the coupling matrix are kept all zero, the matrix is effectively split into several smaller sub-matrices. This is clearly not resource efficient, and it would take kN nodes to create a macrochip. 2 chips are needed. If this macrochip is used to solve smaller problems of size N, then only k such problems can be accommodated. In fact, in that case, only the combined array of chips on the diagonal of the k×k array is used.
[0078] It is not difficult to avoid such waste. By isolating one chip from the rest of the microchip, it can apparently function as an independent Ising machine. In some embodiments, switches may be inserted at the nodes that allow the i-th row and column to be connected either to the i-th node on the chip, or to the corresponding row or column from an adjacent chip. With this support, the chip can either participate in a larger microchip arrangement or operate independently. In fact, the smallest independent unit need not be a single chip, but a module of a size selected by the designer. Referring again to Figures 5A and 5B, the entire diagram can be alternatively treated as one chip, with each block (502-505) being a module. The chip can be treated as one large machine, or as k 2 Each chip can operate either as a single small independent machine or as a single small independent machine. In other embodiments, different systems may be introduced for reconfigurability. For example, referring again to FIG. 5A and FIG. 5B, three types of units in each chip may be reconfigurable: nodes 510, diagonal couplers 511, and interface pins 513. In independent operation, each chip is isolated from the other chips and operates like a single chip Ising machine, with the nodes in normal mode, the pins disconnected from the rows and columns of wires, and the diagonal couplers in crossover mode (wires in row i and column i are connected to diagonal coupler (i,i)). In collective operation, the corresponding rows and columns of wires are connected to adjacent chips via pins 513, all diagonal couplers except the main diagonal of the entire macro chip are switched to normal coupler mode (the main diagonal remains in crossover mode), and only the nodes of the leftmost chip are in normal mode, while other nodes of other chips are in bypass mode. In some arrangements, a subset of chips function in a collective and the rest function independently.
[0079] Although the design of this macrochip is conceptually simple, there are many issues regarding its implementation. The main concern is the chip interface. Depending on whether the chip is integrated through a PCB or an interposer, the interface between the chip and the substrate can be an engineering challenge. The interface carries fast-changing analog signals between multiple chips, which certainly does not make the system behavior easy to analyze. For this reason, in some embodiments, the devices disclosed herein may include a fully digital interface. In a sense, multiple chips as well as digital interconnects are used to create a multi-processor Ising machine.
[0080] As contemplated herein, the digital interconnect between the multiple chips may take a variety of forms, using any transceiver known in the art, including, for example, but not limited to, SPI, I2C, Ethernet, or PCI Express (PCIe), or other bus communication standards. The use of a digital interconnect may require the use of temporary or non-transitory memory to store information received from the digital interconnect or waiting to be transferred across the digital interconnect. In some embodiments, each chip of the multiple chips may include one or more buffers, for example, divided into N regions for storing data associated with N nodes.
[0081] Multiprocessor Architecture By implementing all coupling coefficients in physical units, the macrochip disclosed herein fundamentally avoids any glue calculations to support multichip operation. This important feature is maintained while the multiprocessor architecture addresses the macrochip interface issue.
[0082] This design is shown in Figure 7. At the top 701 is the logical system layout, i.e., N 27 shows an N-node 702 with N-combination units 721. Physically, it can be divided into, for example, four chips (703, 704, 705, 706), each chip containing a complete set of nodes. Some of these nodes (shown in orange) are just "shadow copies" of real nodes on some other chips. For example, node 3 (N3) on chip C1 may be just a buffer in some embodiments. Real node 3 is on chip C2 (704). When node 3 changes polarity (e.g., to -1), C2 communicates this information to the other chips via the digital fabric (not shown). All other chips use buffers to maintain the -1 value of node 3 until C2 informs them of their further changes. In other words, compared to the real nodes, the shadow copies are both approximate in time (a little delayed) and value (always at the voltage rail).
[0083] In this design, a single chip logic structure captures a long slice of the entire connectivity matrix. Yet this logic structure is implemented based on a typical square baseline chip architecture. The difference is that the logic structure of this disclosure is built from modular reconfigurable arrays. Referring to FIG. 8, a diagram of a reconfigurable chip made from 4x4 modules, each with n nodes and nxn connectivity arrays, is shown. The nodes can operate in three different modes: normal nodes (blue), shadow copies (orange), or completely bypassed (green). These modules can be configured in three ways: (1) 4nx4n: the chip is an independent machine with 4n nodes. These 4n nodes are in the first column (blue). The remaining nodes of the module are bypassed (green). (2) 2nx8n: this arrangement allows four chips to be connected into an 8nx8n system. In the current chip, only the 2n nodes are normal (blue) and the 6n nodes are shadow copies (orange). These are connected with wires to become two columns of logic arrays with 8 arrays each. (3) nx16n: As in the previous example, this chip is used with 15 chips to form a multiprocessor with 16n nodes. While the illustrated example of the reconfigurable logic structure is shown as a 4x4 array, it is understood that the multi-module structure may have any suitable geometric shape, including but not limited to 2x2, 3x3, 4x4, 5x5, 6x6, 7x7, 8x8, etc. In some embodiments, the multi-module structure may be arranged in three dimensions, for example, 2x2x2, 3x3x3, 4x4x4, 5x5x5, 6x6x6, etc. While a square or cubic arrangement may be convenient, in some embodiments, the array may have one or more dimensions different from the others. For example, a two-dimensional array may be arranged as XxYxZ, where each of X, Y, and Z is selected from 2, 3, 4, 5, 6, 7, 8, 9, 10, or any other integer.
[0084] As shown in Figure 8, a single chip can be a square array of kxk modules (k=4 in this example). Each module consists of an array of n configurable nodes and nxn connection units. The general idea is that these modules can be stitched together in different ways for different purposes. For example, Figure 8 shows the same group of 16 modules connected together in three configurations: 4x4; 8x2; and 16x1, by varying the interconnections between the modules. Thus, the chip can be used as a single machine with 4n nodes, be part of a 4-chip multiprocessor with 8n nodes, or be part of a 16-chip multiprocessor with 16n nodes.
[0085] Taking a 2nx8n arrangement as a specific example, when combined with three other chips of the same arrangement, the system forms a complete 8nx8n connectivity matrix. In Figure 8 (bottom right) the desired logical configuration of 16 modules is shown. Of these modules, only two (providing the 2n nodes) are configured as normal nodes (modules 1 and 2 in blue), six are configured as shadow copies (3, 4, and 9-12 in orange), and the rest are configured as pass-throughs (green). Not only are the node arrangements different, but the wire connections also need to be changed. For example, modules 1 and 9 need to be connected in such a way that modules 1-4 and 9-12 can function as one column of 8 modules high.
[0086] The basic idea is that when a spin changes polarity, one chip needs to communicate with all other chips, and all those other chips need to update their shadow copies. The communication demand is, to a first approximation, f s for Nlog(N), where N is the total number of spins in the system, and f sis the frequency of spin flips. Considering a specific example of the baseline Ising substrate of this disclosure, on average one spin / node flips every 10-20 ns, depending on the problem being solved. Assuming the same spin flip frequency, if sixteen 8,000-spin chips were used to form a multiprocessor Ising machine, the entire system would have 32,000 spins.
number
[0087] Note that such communication is necessary for any multi-threaded von Neumann solver, the difference being that compared to state-of-the-art physical Ising machines, von Neumann solvers are orders of magnitude slower, and therefore have correspondingly lower bandwidth demands.
[0088] Considering this significant inherent bandwidth demand, several technological solutions immediately come to mind. Optical communication and 3D integration are both attractive options. Indeed, 3D integration is a very convenient solution for the proposed architecture. Figure 9 shows an example of a 4-layer 3D IC. The nodes and their shadow copies are conveniently placed on top of each other, so they can be easily connected with through-silicon vias (TSVs). In fact, due to the short distances of the TSVs, the shadow nodes are no longer necessary in the architecture. In some embodiments, the shadow nodes may still improve the driving capability.
[0089] Finally, in some embodiments, the physical speed of the Ising machine may be slowed down so that communication demand matches the supply of the fabric. In the case of BRIM, this may be achieved by a combination of (at least) two methods. First, the RC constant of the machine may be increased (the coupling resistors used may be larger to slow down the charging). For example, in some embodiments, the coupling resistors may have resistance values between 5 kΩ and 50 kΩ, or between 10 kΩ and 40 kΩ, or between 30 kΩ and 35 kΩ, or about 31 kΩ. In some embodiments, larger coupling resistors (e.g., at least 100 kΩ, at least 200 kΩ, at least 500 kΩ, or between 100 kΩ and 1 MΩ) may be used to increase the associated time constants and slow down the charging.
[0090] Second, the system can be halted completely, for example to wait for congestion to end. No matter what combination of these mechanisms is used, the math is simple: to reduce bandwidth demand by 2x, you need to slow down the machine by 2x. Other methods, discussed further below, may be used to reduce bandwidth demand without a corresponding decrease in performance.
[0091] The parallel operation of multiple Ising machines (solving the same problem) can be loosely described as a combination of each machine independently running a local search and exchanging information about the state of the spins with each other. A surprisingly important design parameter is the latency to communicate spin changes to others. Since a multiprocessor functions most closely to a single large Ising machine, it seems the natural choice to transmit any changes immediately. However, there is also an advantage to waiting: during a time window, the spin may flip multiple times. Some latency avoids wasting bandwidth on unnecessary updates. In this respect, the longer the latency, the more bandwidth can be saved. In practice, however, the latency impacts the solution quality, as explained further below.
[0092] In parallel operation, all solvers are to some extent "ignorant" of the true state of the spins mapped to the other solvers. Take a four-solver system as an example: at any point in time, the spin state of the system is S g =[A,B,C,D] T where each letter represents the spin vector of each machine. Thus, due to the above waits as well as communication delays, the first solver's belief of the current state is S1 = [A,B t ,C t ,D t ] T In the local search, we basically try to optimize the energy E(S1) of this believed state. A low E(S1) does not necessarily mean that the energy E(S g ) is also lower. Now suppose that a solver is provided with the true state of another solver and recalculates the energy. This difference is called the energy surprise (E サプライズ =E(S1)E(S g )). Thus, a positive value means that the current state has lower (better) energy than the solver believed before the update, or in other words, it is a good surprise. Figures 10A and 10B show empirical observations of energy surprises.
[0093] With reference to Figures 10A and 10B, the degree of ignorance and the corresponding energy surprises for different epoch sizes are shown. The graph has 8000 nodes, and when partitioned and mapped to 8 solvers, each solver solves 1000 nodes. Figure 10A shows the communication for small, medium, and large epochs. Figure 10B shows a zoomed-in segment near the origin.
[0094] This particular experiment is obtained by splitting an 8000-node problem into eight subproblems and solving each of them with a Simulated Annealing (SA) solver. After initialization, each solver uses the state of the other 7000 nodes to calculate a bias. Then, they perform a local search of a certain time (called an epoch) before communicating their state to each other. At the epoch boundaries, the external spin can indicate the amount of ignorance measured by the fraction of spins that are not up-to-date. We also calculate the energy surprise. The figure shows the results for each epoch from 20 runs.
[0095] Longer epoch times result in more spin changes from other solvers. As a result, any single solver is in a higher degree of ignorance of the external state, leading to a higher degree of misjudgment and a larger magnitude of surprise. If the epochs are longer than a certain value, the energy surprise is highly correlated with the degree of ignorance. In this regime, we can say that the parallel solver is doing a clearly worse job (also reflected in the very poor final solution quality, not shown in the graph). So far, this message is consistent with the previous analysis that the decomposed subproblems are not independent of each other. However, if the epoch time falls below a certain threshold, the situation seems to undergo a phase change (where the energy surprise is no longer uniformly negative). In any case, the magnitude of surprise becomes lower. In other words, despite some ignorance, the solver is still able to find a reasonable solution. In fact, in some cases, it is better than what the solution would have been believed to be in ignorance. Indeed, the overall solution quality is no worse (and in fact, statistically better) than running the solvers in sequence (and thus without any ignorance).
[0096] Thus, in some embodiments, multiple solvers can run in parallel as long as they keep each other notified "quickly enough", which means that short epoch times are advantageous, and generally means higher communication demands.
[0097] Another important aspect of the design concerns the spin flips introduced into the system to prevent it from getting stuck in a local minimum. (These are referred to herein as induced spin flips.) These spin flips are generally applied probabilistically, similar to the accepted proposal in the Metropolis algorithm (W.K. Hastings, Biometrika, April 1970). In practical implementations, the randomness is often of a deterministic pseudorandom nature. As a result, if the pseudorandom number generators (PRNGs) are properly synchronized on each chip, it can be guaranteed that each chip will generate the same output everywhere at the same time. In this way, induced spin flips may be applied without explicit communication. In other words, instead of randomly selecting, for example, node 3 to flip and sending an explicit message from the chip containing node 3 to the other chips providing the information of the flip, the PRNGs on all chips simultaneously induce a flip in node 3 and update the node's capacitor or shadow register at approximately the same time.
[0098] Batch mode operation Careful design of parallel operations can achieve significant bandwidth savings (in some embodiments, around 1.5x), but an entirely different mode of operation, the batch mode, allows for even more substantial savings (around 5x). This mode takes advantage of the fact that a common, though not universal, mode of using the annealer is to do batch runs with different initial states and then obtain the best solution from that batch. Batch runs of the same setup, if known to exist, may be staggered in a very simple way to reduce the communication required.
[0099] The key idea is illustrated in Figure 11. Vertically, a single job (from one initial state, say Job 1) is distributed across multiple solvers (on chip 1 in epoch 1, chip 2 in epoch 2, etc.) just like in the parallel mode. Thus, each chip is still only annealing a part of the problem. But the solvers now work in sequence. At the end of epoch 1, chip 1 passes its updated spin state to the other chips (indicated in the figure by the first quarter of all chips' spins turning dark red). Then chip 2 selects Job 1 and continues exploring the second quarter of the spins.
[0100] Viewed horizontally, the four chips work on different jobs (indicated by different colors) for each epoch. In the synchronization phase, they exchange updated states and then start annealing on another job. The main advantage of this approach is that the duration of each epoch can be significantly longer without any ignorance occurring. As already discussed, with longer epochs, the total communication bandwidth required can be much less than the bandwidth required to communicate every single event of a spin flip.
[0101] To take advantage of parallelism, in batch mode, n different runs (from different initial states) can be run simultaneously across n solvers in some embodiments. As a result, the system as a whole needs to carry n copies of the state instead of just one in parallel mode. Supporting this requires a modest increase in storage ((nxN) bits per solver) to keep track of the state of the various runs.
[0102] Finally, it is tempting to think that a good way to run the batch mode would be to have every machine run independently, just like in a von Neumann system. This would obviously be less efficient for a multi-processor Ising machine, and if the entire problem were to be solved on a single machine, it would require a context switch with new parameters at the end of each epoch. The data volume would be O(bN 2) bits, and b is the bit width of the combining weight. In contrast, in the disclosed batch mode, the amount of data is O(N). EXAMPLES
[0103] Experimental Example The present invention will be described in more detail with reference to the following experimental examples.These examples are provided for illustrative purposes only and are not intended to be limiting unless otherwise specified.Therefore, the present invention should not be construed as being limited to the following examples in any way, but should be construed as embracing any and all variations that become evident as a result of the teachings provided herein.
[0104] Without further elaboration, it is believed that one of ordinary skill in the art can, using the preceding description and the following examples, make and use the systems and methods of the present invention. Accordingly, the following examples specifically illustrate exemplary embodiments of the present invention, and are not to be construed as limiting in any way the remainder of the disclosure.
[0105] Experimental Method As the development of Ising machines is still in its early stages, access to the physical system is difficult. Therefore, most comparisons disclosed herein are performed by combining modeling using results reported in the literature and measuring the time of simulated annealing (SA). In all experiments, SA is performed natively, but the evolution of the dynamic system in BRIM is modeled by solving differential equations using a fourth-order Runge-Kutta method. When comparing with reported results, the present experiments are limited by the type of benchmarks used in the literature for direct comparison. There is no easy way to compare the solution quality of various problems, so cross-benchmark comparisons are fraught with pitfalls. Fortunately, several benchmarks (K-graphs) are in common use. One such graph used for comparison is known as K16384 (see Kosuke Tatsumura, et al., Nature Electronics (01Mar2021)) and contains 16,384 spins with all-to-all connections. Simulation of dynamic systems using differential equations can be orders of magnitude slower than cycle-level microprocessor simulations. Simulating 1μs dynamics on K16384 takes about 3 days on a very powerful server, so it was only used for direct performance comparisons. Smaller K-graphs (e.g., K2000, Takahiro Inagaki, et al., Science (2016)) are used for some additional analyses.
[0106] In general, the execution time of SA is closest to the standard performance benchmark, but in practice there are some rather nuanced points. First, the fastest known version, Isakov's algorithm (S. Isakov, et al., Computer Physics Communications, Jul 2015), was chosen. Second, an optimization was applied using a dense matrix representation, which takes advantage of fully connected graphs such as K-graphs to improve performance. Finally, the researchers tuned the annealing schedule for these specific graphs, which turned out to have a large impact on execution time. It is possible that performance could be improved by similarly tuning the disclosed hardware annealing schedule. Unfortunately, such tuning was not yet possible in these experiments, as simulation costs were prohibitively high.
[0107] Single Solver Baseline To get a sense of the landscape of physical Ising machines and digital accelerators for simulated annealing, in Figure 12 we show results for several uniprocessor Ising machines running the K2000: the BRIM chip (simulated), simulated annealing (measured), as well as reported results for STATICA (K. Yamamoto, et al., IEEE Journal of Solid-State Circuits, 2021); CIM (T. Inagaki, et al., Science, 2016); and two variants of the Simulated Branching Machine (SBM) (H. Goto, et al., Science Advances, 2021). Many other machines are excluded from this comparison because they simply cannot be mapped into a graph. Despite its seemingly modest size (2000 nodes), the K2000 is a fully connected graph, whereas a locally connected machine would require millions of nodes.
[0108] Referring to Figure 12, the performance of the K2000 graphs on various machines is shown as solution cut value on the y-axis (higher is better) and run time on the x-axis. For every timescale, the machine ran 100 runs. For machines where results for multiple timescales were available, the average result for each timescale is shown as a dashed line, with the range shown as a shaded area. Results for only one timescale are shown as a bar graph, with dots indicating the range and average of the cut values.
[0109] For this graph, BRIM was able to reach the best known solution of 33,337 in 11 μs. The only other machine that was able to reach a similar solution quality was the dSBM, which took 2 ms, about 180x slower. Even if the lower solution quality was acceptable, the single-chip BRIM is still about two orders of magnitude faster than the fastest annealer. The current CIM implementation uses a computational subsystem to simulate the coupling between spins; therefore, it is not strictly a physical Ising machine, but a hybrid machine. One can also hypothesize that performance could be improved if the designers could work out the physical full coupling mechanism (which may not be easy or even possible).
[0110] In summary, a properly designed physical Ising machine can be six orders of magnitude faster than a classical Simulated Annealer (SA) and about two orders of magnitude faster than state-of-the-art computational annealers. The only drawback of a physical Ising machine over a computational annealer is that the latter can be easily scaled to solve larger problems. Below we show how the proposed multiprocessor architecture addresses this issue. Here we focus on the comparison with SA and SBM only, since the latter are currently the fastest systems known.
[0111] High-level comparison The disclosed multiprocessor BRIM (mBRIM) architecture is compared with the SBM using the larger K16384 graph as a benchmark. This allows a direct comparison of the solution quality and performance with results reported in the literature. A four-chip multiprocessor was considered. Each chip was a BRIM-type electronic Ising machine with 8192 nodes. Such chips have a smaller die size (approximately 80 mm in 45 nm technology) than the single FPGA used in the SBM. 2 ) and consumes significantly less power (less than 10W). Three realizations of this multiprocessor were used as proxies for various implementation choices:
[0112] (1) mBRIM 3D : 3D integrated version where communication is essentially instantaneous and there are no bandwidth limitations;
[0113] (2) mBRIM HB : A system with high communication bandwidth. Each chip has three dedicated channels of 250GB / s each. Thus, the total bandwidth is close to that of HBM.
[0114] (3) mBRIM LB : Low bandwidth systems (mBRIM HB (One-fourth of).
[0115] Fig. 13 shows the best solution quality and time obtained by various mBRIMs, an 8-FPGA implementation of SBM, and SA. For clarity, the graph shows only the best quality runs. The best performing mBRIM (mBRIM 3D When compared to SBM in parallel mode, mBRIM achieves a much better solution quality (793,423 - 799,292 vs. SBM's best result of about 792,000) and is about 2200x faster (1.1 μs vs. 2.47 ms). LB ) is over 700x faster than SBM and achieves higher solution quality.
[0116] Next, we look at the impact of bandwidth limitations. As already discussed above, if the communication bandwidth between chips is insufficient, in some embodiments it is possible to address this by slowing down the Ising machine. The impact, of course, is that we have to wait longer to get a solution. mBRIM HB and mBRIM LB Both of these are due to congestion-induced outages, and mBRIM 3D In these bandwidth-limited situations, the batch mode operation disclosed is a reasonably effective tool to improve execution speed. Specifically, in batch mode, the same amount of annealing can be done with mBRIM. HB and mBRIM LB In batch mode, mBRIM can finish 2.8x and 7x faster, respectively. HB mBRIM 3D About 2x slower than mBRIM LB is 1.4x slower, but the solution quality drops to 792,728.
[0117] Finally, we compared mBRIM with SA. To obtain the same solution quality, mBRIM required approximately 4.5 × 10 6 This is shown to be faster than the 1.3×10 6 Note that there is an insane difference (~140x) in the performance of SA due to tuning the annealing schedule.
[0118] In-depth analysis : It can be useful to understand from first principles how different types of solvers work. Whatever solver is used, achieving a good solution requires sufficient exploration of the high-dimensional energy landscape. As an example, for an 800-node graph, simulated annealing (SA) and BRIM explored 148K and 115K distinct states, respectively, to reach comparable solution quality. In BRIM, a spin flip occurs on average every 20ps.
[0119] In (sequential) SA, each individual spin flip was computationally realized, the energies of alternative configurations (with specific spins flipped) were calculated, and the new state was probabilistically accepted based on those energies. Roughly speaking, running SA counted 140,000 instructions executed per spin flip.
[0120] Simulated bifurcation (SB) is a completely new computational approach. It can be thought of as simulating a dynamical system. Thanks to its algorithmic design, it is easily parallelized. It can therefore be faster despite having a similar workload. The non-trivial part of SA is also massively parallel. However, no efforts have been seen to implement parallel SA in custom hardware. Nevertheless, accelerating SB to the level of BRIM would require about 1000x more computational throughput, or about 2 petaOps per second. Therefore, it is clear why physical Ising machines are more attractive, even when compared to the best computational accelerators.
[0121] As already discussed above, the degree of ignorance of the global state can have a large impact on the solution quality. Thus, in parallel runs, it was necessary to have each solver update all the other solvers frequently. In batch mode, this is no longer an issue, since different solvers of the same run are essentially executed in sequence and only need to communicate the cumulative state changes (called bit changes to distinguish them from spin flips) between the start and end of the same epoch. If the spin flips 4 times within an epoch and ends up in the same state as the start of the epoch, there is no need to communicate anything. In other words, there are 4 spin flips but 0 bit changes. Intuitively, the longer the epoch, the more spin flips there are, but no bit changes. Figures 14A and 14B quantitatively confirm this intuition. We measured the number of spin flips during the epoch and counted the number of bit changes. For an epoch of 3.3 ns, both the numbers and their ratios as the annealing progresses are shown in Figure 14A. FIG. 14B shows the ratio as a function of different epoch sizes.
[0122] Referring to Figures 14A and 14B, Figure 14A shows the evolution of flips and bit changes over time for a 4-chip BRIM with a fixed epoch of 3.3 ns. The left vertical axis corresponds to flips (solid blue line) and bit changes (dashed blue line). The right vertical axis corresponds to the ratio of flips to bit changes shown in red. Figure 14B shows the correlation between the average ratio of flips to bit changes and epoch size. This ratio increases approximately linearly with increasing epoch size.
[0123] The number of spin flips during an epoch was measured and the number of bit changes was counted. In Fig. 14A, both the numbers and their ratios are shown as the annealing progresses. If the epoch size is fixed, the ratios after the initial period are rather stable. Fig. 14B shows the ratios as a function of different epoch sizes. Naturally, the longer the epoch, the higher the ratio. As shown, when using an epoch size of about 3 ns, the traffic demand can be reduced by about 4-5x compared to when using sub-nanosecond epochs. As can be seen in Fig. 15A and Fig. 15B, where the solution quality is shown as a function of epoch size, increasing the epoch size reduces the solution quality. The best solution quality was achieved in parallel mode with a small epoch size. If the bandwidth is sufficient, this is the best mode to use. In a bandwidth-limited system, the dynamic system needs to slow down. In this case, a traffic reduction of 4-5x means that the dynamic system can run about 4-5x faster. Longer epochs would be needed to achieve the traffic reduction. In such cases, parallel mode cannot tolerate longer epochs and solution quality degrades quickly and significantly. Batch mode, on the other hand, is much more tolerant of longer epochs, as solution quality degrades, but only slightly. Therefore, in bandwidth-limited systems, batch mode is very useful as it does not sacrifice execution speed while still maintaining high solution quality.
[0124] Finally, we investigated the bandwidth reduction when tuning the induced spin flip. Figure 16A shows the amount of bit changes and induced spin flips with time evolution. The fraction of bit changes that are induced spin flips is also plotted. Naturally, this value is a function of epoch size. Figure 16B shows the average fraction over different epoch sizes. Clearly, a non-trivial amount of communication (30-38%) can be saved by optimizing the tuning of the induced flip. In bandwidth-constrained systems, a corresponding improvement in run time (around 1.5x) can be expected.
[0125] Figure 16A shows the evolution of induced spin flips and bit changes for a 4-chip BRIM with a fixed epoch of 3.3 ns. The left vertical axis corresponds to induced spin flips (solid blue line) and bit changes (dashed blue line). The right vertical axis corresponds to the fraction of bit changes that are induced spin flips, shown in red. Figure 16B shows the average fraction of induced spin flips versus epoch duration.
[0126] Comparison with other parallel processes Finally, we note that in parallel processing, communication between distributed agents is clearly a common component and performance bottleneck. Thus, in the search for a solution for Ising machines, some wheels may have been reinvented. For example, using shadow copies is necessary for the disclosed system, as is keeping copies (ghosts) of non-local neighbors is for parallel algorithms. Also, techniques to reduce communication while limiting the impact on performance have been explored in various contexts, namely sending low-precision data (possibly only 1 bit), using lossy compression, reducing the number of elements transmitted, or even skipping rounds. Compared to these contexts, specifically in the case of Ising machines, two important differences can be highlighted:
[0127] (1) Problem Scale: Ising machines are dynamical systems and can evolve very rapidly. Thus, without optimization, the raw communication demands become enormous. For example, the four disclosed BRIM chips (each about 80 mm in 45 nm technology) 2 ) requires approximately 4TB / s.
[0128] (2) Design flexibility: The freedom to orchestrate the evolutionary process of the underlying Ising machine can be exploited, allowing optimizations such as batch mode and tuning of induced spin flips. As a result, there is no additional logic such as compression, and the communication demand can be reduced to just 218GB / s (a 20x reduction) while maintaining solution quality.
[0129] conclusion A physical Ising machine can solve optimization problems of the Ising equations extremely fast and energy-efficiently, even compared to dedicated (von Neumann) accelerators. However, existing Ising machines have a fixed capacity. When a divide-and-conquer strategy is used, the benefits of using an Ising machine diminish rapidly even when the problem is slightly larger than the machine capacity. The device disclosed herein is essentially designed to cooperate with other machines to solve larger problems. This disclosure presents the design and optimization of a multi-processor Ising machine architecture. Experimental results related to the design can be summarized in several key points:
[0130] (1) Even under conservative conditions (low chip count, low chip area and power consumption), multiprocessors can achieve speedups of approximately 2200x compared to state-of-the-art computational accelerators.
[0131] (2) The proposed multiprocessor architecture also allows the physical Ising machine to be scaled up to solve larger problems. Although it is difficult to compare the speedup for different problems, it is safe to say that the performance advantage of a multiprocessor BRIM over its von Neumann counterpart is as significant as that of a single-chip BRIM.
[0132] (3) Because physical Ising machines are so fast, communication bandwidth can become a potential bottleneck and the machine needs to slow down accordingly. In these cases, batch mode operation can reduce communication demands by about 4-5x, which means that processing throughput can be increased by about 4-5x.
[0133] The disclosures of any and all patents, patent applications, and publications cited herein are incorporated herein by reference in their entirety. Although the present invention has been disclosed with reference to specific embodiments, it will be apparent to those skilled in the art that other embodiments and variations of the present invention may be devised without departing from the true spirit and scope of the present invention. It is intended that the appended claims be construed to include all such embodiments and equivalent variations.
[0134] References The following publications are incorporated herein by reference in their entireties:
[0135] Martin Abadi,Ashish Agarwal,Paul Barham,Eugene Brevdo,Zhifeng Chen,Craig Citro,Gregory S.Corrado,Andy Davis,Jeffrey Dean,Matthieu Devin,Sanjay Ghemawat,Ian J.Goodfellow,Andrew Harp,Geoffrey Irving,Michael Isard,Yangqing Jia,Rafal Jozefowicz,Lukasz Kaiser,Manjunath Kudlur,Josh Levenberg,Dan Mane,Rajat Monga,Sherry Moore,Derek Gordon Murray,Chris Olah,Mike Schuster,Jonathon Shlens,Benoit Steiner,Ilya Sutskever,Kunal Talwar,Paul A.Tucker,Vincent Vanhoucke,Vijay Vasudevan,Fernanda B.Viegas,Oriol Vinyals,Pete Warden,Martin Wattenberg,Martin Wicke,Yuan Yu,and Xiaoqiang Zheng.2016.TensorFlow:Large-Scale Machine Learning on Heterogeneous Distributed Systems.CoRRabs / 1603.04467(2016).arXiv:1603.04467 http: / / arxiv.org / abs / 1603.04467
[0136] David H Ackley,Geoffrey E Hinton,and Terrence J Sejnowski.1985.A learning algorithm for Boltzmann machines.Cognitive science9,1(1985),147-169.
[0137] Richard Afoakwa,Yiqiao Zhang,Uday Kumar Reddy Vengalam,Zeljko Ignjatovic,and Michael Huang.2021.BRIM:Bistable Resistively-Coupled Ising Machine.2021 IEEE International Symposium on High-Performance Computer Architecture(HPCA)(2021),749-760.https: / / doi.org / 10.1109 / HPCA51647.2021.00068
[0138] Dan Alistarh,Torsten Hoefler,Mikael Johansson,Sarit Khirirat,Nikola Konstantinov,and Cedric Renggli.2018.The Convergence of Sparsified Gradient Methods.In Proceedings of the 32nd International Conference on Neural Information Processing Systems(Montreal,Canada)(NIPS’18).Curran Associates Inc.,Red Hook,NY,USA,5977-5987.
[0139] Dan Alistarh,Jerry Li,Ryota Tomioka,and Milan Vojnovic.2016.QSGD:Randomized Quantization for Communication-Optimal Stochastic Gradient Descent.CoRRabs / 1610.02132(2016).arXiv:1610.02132 http: / / arxiv.org / abs / 1610.02132
[0140] Rami Barends,Alireza Shabani,Lucas Lamata,Julian Kelly,Antonio Mezzacapo,Urtzi Las Heras,Ryan Babbush,Austin G Fowler,Brooks Campbell,Yu Chen,et al.2016.Digitized adiabatic quantum computing with a superconducting circuit.Nature534,7606(2016),222-226.
[0141] William J.Barry,Mark T.Jones,and Paul E.Plassmann.1998. Parallel adaptive mesh refinement techniques for plasticity problems.Advances in Engineering Software29,3(1998),217-225.https: / / doi.org / 10.1016 / S0965-9978(98)00040-4
[0142] Debraj Basu,Deepesh Data,Can Karakus,and Suhas Diggavi.2019.Qsparse-LocalSGD:Distributed SGD with Quantization,Sparsification,and Local Computations.Curran Associates Inc.,Red Hook,NY,USA.
[0143] Natalia G Berloff,Matteo Silva,Kirill Kalinin,Alexis Askitopoulos,Julian D Topfer,Pasquale Cilibrizzi,Wolfgang Langbein,and Pavlos G Lagoudakis.2017.Realizing the classical XY Hamiltonian in polariton simulators.Nature materials16,11(2017),1120-1126.
[0144] Jeremy Bernstein,Yu-Xiang Wang,Kamyar Azizzadenesheli,and Anima Anandkumar.2018.signSGD:compressed optimisation for non-convex problems.CoRR abs / 1802.04434(2018).arXiv:1802.04434 http: / / arxiv.org / abs / 1802.04434
[0145] Fabian Bohm,Guy Verschaffelt,and Guy Van der Sande.2019.A poor man’s coherent Ising machine based on opto-electronic feedback systems for solving optimization problems.Nature Communications10,1(2019),3538.https: / / doi.org / 10.1038 / s41467-019-11484-3
[0146] Sergio Boixo,Troels F Ronnow,Sergei V Isakov,Zhihui Wang,David Wecker,Daniel A Lidar,John M Martinis,and Matthias Troyer.2014. Evidence for quantum annealing with more than one hundred qubits.Nature physics 10,3(2014),218-224.
[0147] Michael Booth,Steven P.Reinhardt,and Aidan Roy.2017.Partitioning Optimization Problems for Hybrid Classical / Quantum Execution.Technical Report (2017).https: / / docs.ocean.dwavesys.com / projects / qbsolv / en / latest / _downloads / bd15a2d8f32e587e9e5997ce9d5512cc / qbsolv_techReport.pdf
[0148] Paul I Bunyk,Emile M Hoskinson,Mark W Johnson,Elena Tolkacheva,Fabio Altomare,Andrew J Berkley,Richard Harris,Jeremy P Hilton,Trevor Lanting,Anthony J Przybysz,et al.2014.Architectural considerations in the design of a superconducting quantum annealing processor.IEEE Transactions on Applied Superconductivity24,4(2014),1-10.
[0149] KyungHyun Cho,Alexander Ilin,and Tapani Raiko.2011.Improved learning of Gaussian-Bernoulli restricted Boltzmann machines.In International conference on artificial neural networks.Springer,10-17.
[0150] Jeffrey Chou,Suraj Bramhavar,Siddhartha Ghosh,and William Herzog.2019.Analog Coupled Oscillator Based Weighted Ising Machine.Scientific Reports9,1(2019),14786.https: / / doi.org / 10.1038 / s41598-019-49699-5
[0151] Chase Cook,Hengyang Zhao,Takashi Sato,Masayuki Hiromoto,and Sheldon X.D.Tan.2019.GPU Based Parallel Ising Computing for Combinatorial Optimization Problems in VLSI Physical Design.arXiv:1807.10750[physics.comp-ph]
[0152] D-WAVE.2014.minorminer.https: / / github.com / dwavesystems / minorminer
[0153] D-WAVE.2022.Operation and Timing.https: / / docs.dwavesys.com / docs / latest / c_qpu_timing.html
[0154] Zachary DeVito,Niels Joubert,Francisco Palacios,Stephen Oakley,Montserrat Medina,Mike Barrientos,Erich Elsen,Frank Ham,Alex Aiken,Karthik Duraisamy,Eric Darve,Juan Alonso,and Pat Hanrahan.2011.Liszt:A domain specific language for building portable mesh-based PDE solvers.In SC’11:Proceedings of 2011 International Conference for High Performance Computing,Networking,Storage and Analysis.1-12.https: / / doi.org / 10.1145 / 2063384.2063396
[0155] Melih Elibol,Lihua Lei,and Michael I.Jordan.2020.Variance Reduction with Sparse Gradients.CoRR abs / 2001.09623(2020).arXiv:2001.09623 https: / / arxiv.org / abs / 2001.09623
[0156] Hayato Goto, Kotaro Endo, Masaru Suzuki, Yoshisato Sakai, Taro Kanao, Yohei Hamakawa, Ryo Hidaka, Masaya Yamasaki, and Kosuke Tatsumura Advances7.6(2021),eabe7953.https: / / doi.org / 10.1126 / sciadv.abe7953 arXiv:https: / / www.science.org / doi / pdf / 10.1126 / sciadv.abe7953
[0157] Hidenori Gyoten, Masayuki Hiromoto, and Takashi Sato. Area Efficient Annealing Processor for Icing Model without Random Number Generator.IEICE Transactions on Information and Systems E101.D,2(2018),314-323
[0158] Ryan Hamerly, Takahiro Inagaki, Peter L McMahon, David Venturelli, Alireza Marandi, Tatsuhiro Onodera, Edwin Ng, Carsten Langrock, Kensuke Inaba, et al 2000Q.D-Wave 2000Q arXiv(2018).
[0159] Ryan Hamerly,Takahiro Inagaki,Peter L McMahon,Davide Venturelli,Alireza Marandi,Tatsuhiro Onodera,Edwin Ng,Carsten Langrock,Kensuke Inaba,Toshimori Honjo,et al.2019.Experimental investigation of performance differences between coherent Ising machines and a quantum annealer.Science advances5,5(2019),eaau0823.
[0160] R Hamerly,A Sludds,L Bernstein,M Prabhu,C Roques-Carmes,J Carolan,Y Yamamoto,M Soljacic,and D Englund.2019.Towards Large-Scale Photonic Neural-Network Accelerators.2019 IEEE International Electron Devices Meeting (IEDM).IEEE,22-8.
[0161] R.Harris,M.W.Johnson,T.Lanting,A.J.Berkley,J.Johansson,P.Bunyk,E.Tolkacheva,E.Ladizinsky,N.Ladizinsky,T.Oh,F.Cioata,I.Perminov,P.Spear,C.Enderud,C.Rich,S.Uchaikin,M.C.Thom,E.M.Chapple,J.Wang,B.Wilson,M.H.S.Amin,N.Dickson,K.Karimi,B.Macready,C.J.S.Truncik,and G.Rose.2010.Experimental investigation of an eight-qubit unit cell in a superconducting optimization processor.Phys.Rev.B82(Jul2010),024511.Issue2.https: / / doi.org / 10.1103 / PhysRevB.82.024511
[0162] Takahiro Inagaki,Yoshitaka Haribara,Koji Igarashi,Tomohiro Sonobe,Shuhei Tamate,Toshimori Honjo,Alireza Marandi,Peter L McMahon,Takeshi Umeki,Koji Enbutsu,et al.2016. A coherent Ising machine for 2000-node optimization problems.Science354,6312(2016),603-606.
[0163] S.V.Isakov,I.N.Zintchenko,T.F.Ronnow,and M.Troyer.2015.Optimised simulated annealing for Ising spin glasses.Computer Physics Communications192(Jul2015),265-271.https: / / doi.org / 10.1016 / j.cpc.2015.02.015
[0164] Richard M.Karp.1972.Reducibility among Combinatorial Problems.Springer US,Boston,MA,85-103.https: / / doi.org / 10.1007 / 978-1-4684-2001-2_9
[0165] Kihwan Kim,M-S Chang,Simcha Korenblit,Rajibul Islam,Emily E Edwards,James K Freericks,G-D Lin,L-M Duan,and Christopher Monroe.2010.Quantum simulation of frustrated Ising spins with trapped ions.Nature465,7298(2010),590-593.
[0166] Andrew D King,Juan Carrasquilla,Jack Raymond,Isil Ozfidan,Evgeny Andriyash,Andrew Berkley,Mauricio Reis,Trevor Lanting,Richard Harris,Fabio Altomare,et al.2018.Observation of topological phenomena in a programmable lattice of 1,800qubits.Nature 560,7719(2018),456-460.
[0167] S.Kirkpatrick,C.D.Gelatt,and M.P.Vecchi.1983.Optimization by Simulated Annealing.Science220,4598(1983),671-680.https: / / doi.org / 10.1126 / science.220.4598.671 arXiv:https: / / science.sciencemag.org / content / 220 / 4598 / 671.full.pdf
[0168] Orion S.Lawlor,Sayantan Chakravorty,Terry L.Wilmarth,Nilesh Choudhury,Isaac Dooley,Gengbin Zheng,and Laxmikant V. Kale.2006.ParFUM:A Parallel Framework for Unstructured Meshes for Scalable Dynamic Physics Applications.Eng.with Comput.22,3(dec2006),215-235.
[0169] Tao Lin,Sebastian U.Stich,and Martin Jaggi.2018.Don’t Use Large Mini-Batches,Use Local SGD.CoRR abs / 1808.07217(2018).arXiv:1808.07217 http: / / arxiv.org / abs / 1808.07217
[0170] Andrew Lucas.2014.Ising formulations of many NP problems.Frontiers in Physics2(2014),5.
[0171] Peter L.McMahon,Alireza Marandi,Yoshitaka Haribara,Ryan Hamerly,Carsten Langrock,Shuhei Tomato,Takahiro Inagaki,Hiroki Takesue,Shoko Utsunomiya,Kazuyuki Aihara,Robert L.Byer,MMFejer,Hideo Mabuchi,and Yoshihisa Yamamoto.2016.A fully Science354,6312(2016),614-617.https: / / doi.org / 10.1126 / science.aah5178 arXiv:https: / / science.sciencemag.org / content / 354 / 6312 / 614.full.pdf
[0172] Chris Mellor.2021.DRAM,It Stacks Up:SK Hynix Rolls Out 819 GB / s HBM3 Tech.
[0173] A Parallel Ghosting Algorithm for the Flexible Distributed Mesh Database.Scientific Programming 21(01Jan1900),654971.https: / / doi.org / 10.3233 / SPR-130361
[0174] Saavan Patel,Lili Chen,Philip Canoza,and Sayeef Salahuddin.2020.Ising Model Optimization Problems on a FPGA Accelerated Restricted Boltzmann Machine.arXiv:2008.04436[cs.AR]
[0175] D Pierangeli,G Marcucci,and C Conti.2019. Large-scale photonic Ising machine by spatial light modulation.Physical review letters122,21(2019),213902.
[0176] Frank Seide,Hao Fu,Jasha Droppo,Gang Li,and Dong Yu.2014.1-Bit Stochastic Gradient Descent and Application to Data-Parallel Distributed Training of Speech DNNs.In Interspeech 2014(interspeech 2014 ed.).https: / / www.microsoft.com / en-us / research / publication / 1-bit-stochastic-gradient-descent-and-application-to-data-parallel-distributed-training-of-speech-dnns /
[0177] Sebastian U.Stich.2019.Local SGD Converges Fast and Communicates Little.In International Conference on Learning Representations.https: / / openreview.net / forum?id=S1g2JnRcFX
[0178] Sebastian U.Stich,Jean-Baptiste Cordonnier,and Martin Jaggi.2018.Sparsified SGD with Memory.In Proceedings of the 32nd International Conference on Neural Information Processing Systems(Montreal,Canada)(NIPS’18).Curran Associates Inc.,Red Hook,NY,USA,4452-4463.
[0179] Kenta Takata,Alireza Marandi,Ryan Hamerly,Yoshitaka Haribara,Daiki Maruo,Shuhei Tamate,Hiromasa Sakaguchi,Shoko Utsunomiya,and Yoshihisa Yamamoto.2016.A 16-bit Coherent Ising Machine for One-Dimensional Ring and Cubic Graph Problems.Scientific Reports6,1(2016),34089.https: / / doi.org / 10.1038 / srep34089
[0180] Y Takeda,S Tamate,Y Yamamoto,H Takesue,T Inagaki,and S Utsunomiya.2017.Boltzmann sampling for an XY model using a non-degenerate optical parametric oscillator network.Quantum Science and Technology3,1(nov2017),014004.https: / / doi.org / 10.1088 / 2058-9565 / aa923b
[0181] T.Takemoto,M.Hayashi,C.Yoshimura,and M.Yamaoka.2019.2.6A 2 by 30k-Spin Multichip Scalable Annealing Processor Based on a Processing-In-Memory Approach for Solving Large-Scale Combinatorial Optimization Problems.In IEEE International Solid-State Circuits Conference.
[0182] Kosuke Tatsumura,Alexander R.Dixon,and Hayato Goto.2019.FPGA-Based Simulated Bifurcation Machine.In 2019 29th International Conference on Field Programmable Logic and Applications(FPL).59-66.https: / / doi.org / 10.1109 / FPL.2019.00019
[0183] Kosuke Tatsumura,Masaya Yamasaki,and Hayato Goto.2021.Scaling out Ising machines using a multi-chip architecture for simulated bifurcation.Nature Electronics4,3(01Mar2021),208-217.https: / / doi.org / 10.1038 / s41928-021-00546-4
[0184] Hongyi Wang,Scott Sievert,Zachary Charles,Shengchao Liu,Stephen Wright,and Dimitris Papailiopoulos.2018.ATOMO:Communication-Efficient Learning via Atomic Sparsification.In Proceedings of the 32nd International Conference on Neural Information Processing Systems(Montreal,Canada)(NIPS’18).Curran Associates Inc.,Red Hook,NY,USA,9872-9883.
[0185] Tianshi Wang and Jaijeet Roychowdhury.2019.OIM:Oscillator-Based Ising Machines for Solving Combinatorial Optimisation Problems.arXiv:1903.07163[cs.ET]
[0186] Tianshi Wang,Leon Wu,and Jaijeet Roychowdhury.2019.New Computational Results and Hardware Prototypes for Oscillator-Based Ising Machines.In Proceedings of the 56th Annual Design Automation Conference 2019(Las Vegas,NV,USA)(DAC’19).Association for Computing Machinery,New York,NY,USA,Article239,2pages.https: / / doi.org / 10.1145 / 3316781.3322473
[0187] Jianqiao Wangni,Jialei Wang,Ji Liu,and Tong Zhang.2018.Gradient Sparsification for Communication-Efficient Distributed Optimization.In Advances in Neural Information Processing Systems,S.Bengio,H.Wallach,H.Larochelle,K.Grauman,N.Cesa-Bianchi,and R.Garnett(Eds.),Vol.31.Curran Associates,Inc.https: / / proceedings.neurips.cc / paper / 2018 / file / 3328bdf9a4b9504b9398284244fe97c2-Paper.pdf
[0188] Kasho Yamamoto,Kazushi Kawamura,Kota Ando,Normann Mertig,Takashi Takemoto,Masanao Yamaoka,Hiroshi Teramoto,Akira Sakai,Shinya Takamaeda-Yamazaki,and Masato Motomura.2021.STATICA:A 512-Spin 0.25M-Weight Annealing Processor With an All-Spin-Updates-at-Once Architecture for Combinatorial Optimization With Complete Spin-Spin Interactions. IEEE Journal of Solid-State Circuits 56,1(2021),165-178.https: / / doi.org / 10.1109 / JSSC.2020.3027702
[0189] Yoshihisa Yamamoto,Kazuyuki Aihara,Timothee Leleu,Ken-ichi Kawarabayashi,Satoshi Kako,Martin Fejer,Kyo Inoue,and Hiroki Takesue.2017.Coherent Ising machines-Optical neural networks operating at the quantum limit.npj Quantum Information3,1(2017),1-15.
[0190] Masanao Yamaoka,Chihiro Yoshimura,Masato Hayashi,Takuya Okuyama,Hidetaka Aoki,and Hiroyuki Mizuno.2015. A 20k-spin Ising chip to solve combinatorial optimization problems with CMOS annealing.IEEE Journal of Solid-State Circuits51,1(2015),303-309.
[0191] M.Yamaoka,C.Yoshimura,M.Hayashi,T.Okuyama,H.Aoki,and H.Mizuno.2015.24.3 20k-spin Ising chip for combinational optimization problem with CMOS annealing.In 2015 IEEE International Solid-State Circuits Conference-(ISSCC)Digest of Technical Papers.1-3.https: / / doi.org / 10.1109 / ISSCC.2015.7063111
[0192] Chihiro Yoshimura,Masato Hayashi,Takuya Okuyama,and Masanao Yamaoka.2017.Implementation and Evaluation of FPGA-based Annealing Processor for Ising Model by use of Resource Sharing. International Journal of Networking and Computing7,2(2017),154-172.http: / / www.ijnc.org / index.php / ijnc / article / view / 148
[0193] G Zames,NM Ajlouni,NM Ajlouni,NM Ajlouni,JH Holland,WD Hills,and DE Goldberg.1981. Genetic algorithms in search,optimization and machine learning.Information Technology Journal3,1(1981),301-302.
[0194] W.K.Hastings,Monte Carlo sampling methods using Markov chains and their applications,Biometrika,Volume 57,Issue1,April1970,Pages97-109,https: / / doi.org / 10.1093 / biomet / 57.1.97
Claims
1. A scalable Ising machine system including multiple chips, comprising: Each chip is a plurality of N nodes, each node having a capacitor, a positive terminal, and a negative terminal; a plurality of N×M connection units arranged in N rows and M columns, each connection unit having a set of reconfigurable resistive connections, each connection unit being configurable to connect a pair of the N nodes via the reconfigurable resistive connections; a plurality of interconnects, each chip of the plurality of chips being communicatively coupled to every other chip of the plurality of chips via at least one interconnect; The scalable Ising machine system.
2. The scalable Ising machine system of claim 1 , wherein the plurality of chips are arranged in a two-dimensional array.
3. The scalable Ising machine system of claim 1 , wherein the plurality of chips are arranged in a three-dimensional array.
4. 13. A scalable Ising machine system according to any preceding claim, wherein the plurality of chips are arranged in at least one square array.
5. 2. A scalable Ising machine system according to any preceding claim, wherein at least one interconnect of the plurality of interconnects comprises a wireless data connection.
6. 13. A scalable Ising machine system according to any preceding claim, wherein each chip further comprises a data buffer configured to digitally store state information of at least a subset of the N nodes.
7. 13. A scalable Ising machine system according to any preceding claim, wherein N=M.
8. 13. A scalable Ising machine system according to any preceding claim, wherein each connection unit comprises two positive terminals, each connected to the positive terminals of a different node in the plurality of nodes, and two negative terminals, each connected to the negative terminals of a different node of the plurality of nodes.
9. 2. A scalable Ising machine system according to any preceding claim, wherein at least one interconnect of the plurality of interconnects includes a switch configured to connect or disconnect the interconnect.
10. 13. A scalable Ising machine system according to any preceding claim, wherein at least one chip further comprises a reconfigurable connectivity fabric for connecting said nodes.
11. 13. The scalable Ising machine system of any preceding claim, wherein each chip further includes a buffer memory, a processor, and a non-transitory computer-readable medium having instructions stored thereon that, when executed by the processor, store node states in the buffer memory and retrieve node states from the buffer memory.
12. 10. A scalable Ising machine system according to any preceding claim, wherein the instructions further comprise transmitting node states in sequence from one chip to the next chip for executing larger tasks in batch mode.
13. 10. A scalable Ising machine system according to any preceding claim, wherein each buffer memory is sufficient to store a buffered copy of at least a subset of the states in the scalable Ising machine system.
14. 10. A scalable Ising machine system according to any preceding claim, wherein each buffer memory is sufficient to store a buffered copy of all the states in the scalable Ising machine system.
15. 1. A method for calculating a Hamiltonian for a coupled spin system, comprising the steps of: To provide a scalable Ising machine system including a plurality of chips, each chip comprising: a plurality of N nodes, each node having a capacitor, a positive terminal, and a negative terminal, the charge on the capacitor representing a spin; a plurality of N×M connection units arranged in N rows and M columns, each connection unit having a set of reconfigurable resistive connections, each connection unit being configurable to connect a pair of the N nodes via the reconfigurable resistive connections; providing, connecting the plurality of chips to one another via a set of interconnects; segmenting the coupled spin system into a set of subsystems, each chip of the plurality of chips being constructed with a subsystem of the set of subsystems; calculating the Hamiltonian of the coupled spin system by computing all of the subsystems; The method comprising:
16. The method of claim 15 , comprising computing the subsystems at least partially in sequence.
17. The method of claim 15 , comprising computing the subsystems simultaneously.
18. 10. A method according to any preceding claim, further comprising storing the states of at least a subset of the nodes in a buffer memory.
19. 10. The method of any preceding claim, further comprising transmitting a subset of the node states from one chip to another.
20. 10. A method according to any preceding claim, further comprising storing the state of all said nodes in a buffer memory on each chip of said plurality of chips.
