A fully connected Ising model reconfigurable processing circuit supporting multiple algorithms

By designing a fully connected Ising model reconfigurable processing circuit that supports multiple algorithms, the problem that existing processors cannot support multiple combinatorial optimization algorithms is solved, and a combinatorial optimization solution with high convergence speed and low complexity is achieved.

CN117951427BActive Publication Date: 2026-11-17SOUTH CHINA UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410296289.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-15
Publication Date
2026-11-17
Estimated Expiration
2044-03-15

AI Technical Summary

Technical Problem

Existing processors struggle to effectively support a variety of combinatorial optimization algorithms, especially the PMT algorithm, which limits the scope of combinatorial optimization problems that can be solved.

Method used

Design a fully connected Ising model reconfigurable processing circuit that supports multiple algorithms, including eight replicas, a global controller, a temperature exchange module, a random number generator, and a temperature generator. Different algorithms are implemented through different computation modes, and the computation process is optimized by utilizing the configuration of the global controller and spin index storage.

Benefits of technology

It achieves high convergence speed and high scalability, enabling it to quickly solve a wider range of combinatorial optimization problems, reducing computational complexity and hardware resource consumption, and improving computational efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117951427B_ABST
    Figure CN117951427B_ABST
Patent Text Reader

Abstract

This invention discloses a fully connected Ising model reconfigurable processing circuit supporting multiple algorithms, relating to the field of combinatorial optimization technology. It addresses the limitations of existing technologies in handling combinatorial optimization problems by proposing this solution. It includes eight replicas, a global controller, a temperature exchange module, a random number generator, and a temperature generator, implementing different algorithms through different computation modes. Its advantages include support for multiple algorithms, high convergence speed, and the ability to quickly solve more types of combinatorial optimization problems: different computation modes are implemented by configuring the mode signal S in different global controllers, allowing multiple algorithms to be implemented in a single architecture; in PMT mode, by storing the flipped spin index and using it for the next local field update, it is not necessary to traverse all spin states, greatly reducing the time for each spin update state, and the single-step update time decreases with increasing iterations because the number of flipped spins decreases as the number of iterations increases.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of combinatorial optimization technology, and in particular to a fully connected Ising model reconfigurable processing circuit that supports multiple algorithms. Background Technology

[0002] Combinatorial optimization problems are prevalent in everyday life, such as the Traveling Salesman Problem, the Maximum Cut Problem, the Packing Problem, and the Knapsack Problem. The Ising model, based on numerous spin connections, attempts to simulate the complexities of ferromagnetic materials. Many combinatorial optimization problems can be mapped to the Ising model, and solutions can be obtained by solving these problems. However, most combinatorial optimization problems are nondeterministic, polynomial-time difficult problems, which are challenging to solve using the traditional von Neumann architecture.

[0003] Quantum annealing architectures offer exceptional accuracy and speed, but at the cost of extreme operating conditions and limited scalability. With advancements in semiconductor manufacturing technology, several CMOS Ising machines based on analog annealing have been developed. These are widely accepted due to their speed, acceptable accuracy, and lower cost. However, most of them only support sparsely connected Ising models, lacking versatility and practicality.

[0004] Fully connected Ising models can more accurately describe the complex interactions in a system and adapt to various data and problems, but they lead to huge computational and storage requirements when the number of nodes is large. Fortunately, recent research has proposed some hardware-friendly algorithms that balance accuracy and speed, suitable for high-speed solutions to fully connected Ising models. These are simulated annealing (SA), parallel tempering (PT), and concurrent dynamic tempering (PMT), each with its own advantages and characteristics when solving the same or different problems.

[0005] However, currently no processor can adequately accelerate the above algorithms, especially the PMT algorithm. Furthermore, research shows that there is no single universal algorithm that can solve all common combinatorial optimization problems in real life. Therefore, a processor that supports multiple algorithms is needed. However, no processor can currently support all these algorithms simultaneously, resulting in a limited range of combinatorial optimization problems it can solve. Summary of the Invention

[0006] The purpose of this invention is to provide a fully connected Ising model reconfigurable processing circuit that supports multiple algorithms, so as to solve the problems existing in the prior art.

[0007] The fully connected Ising model reconfigurable processing circuit supporting multiple algorithms described in this invention includes: eight replicas, a global controller, a temperature exchange module, a random number generator, and a temperature generator, which implement different algorithms through different calculation modes.

[0008] The fully connected Ising model reconfigurable processing circuit supporting multiple algorithms described in this invention has the advantages of supporting multiple algorithms, having a high convergence speed, and being able to quickly solve a wider range of combinatorial optimization problems: by configuring different computation modes in different global controllers, multiple algorithms are implemented in one architecture to solve a wider range of combinatorial optimization problems; in PMT mode, by storing the flipped spin index and using it for the next local field update, it is not necessary to traverse all spin states, which greatly reduces the time for updating each spin state, and because the number of flipped spins decreases as the number of iterations increases, the single-step update time decreases with the increase of iterations, which greatly improves the convergence speed. Attached Figure Description

[0009] Figure 1 This is a diagram of the overall architecture of the present invention.

[0010] Figure 2 This is a schematic diagram of the global controller of the present invention.

[0011] Figure 3 This is a schematic diagram of the processing unit of the present invention.

[0012] Figure 4 This is a schematic diagram of the PMT updater and SA updater of the present invention. Detailed Implementation

[0013] like Figures 1 to 4 As shown, the fully connected Ising model reconfigurable processing circuit supporting multiple algorithms described in this invention includes eight replicas, a global controller, a temperature exchange module, a random number generator, and a temperature generator, and implements different algorithms through different calculation modes.

[0014] Each of the eight replicas consists of a 4×32 pulsating array supporting 1024 left-handed and 1024 right-handed spins, a selector controlling the PMT updater and SA updater, an energy calculator, and a flipped spin register. The pulsating array is functionally reconfigurable, supporting different algorithms depending on the configuration; the selector selects the operation of the PMT updater or the SA updater according to functional requirements; the energy calculator calculates the total energy of the system; and the flipped spin register stores the coefficients of the flipped spins.

[0015] The global controller consists of four modules: an I / O module, an iteration step counter, a configuration allocator, and a finite state machine. The I / O module is responsible for exchanging information between the user and the processor architecture; the iteration step counter counts the current iteration step; the configuration allocator configures global parameters and processes based on the finite state machine state and user input configuration; and the finite state machine performs state transitions based on the iteration step count and configuration. The global controller comprises seven operating states: Idle (S0), Left Spin Initialization (S1), Right Spin Initialization (S2), Left Spin Update (S3), Right Spin Update (S4), Replica Exchange (S5), and Complete (S6). In state S0, all components are in standby mode. When the user's mode signal S is not 00, the Idle state switches to either S1 or S2 based on the value of S. State S1 exists only in PMT mode. It calculates and stores the local fields of all left spins, updates all left spin states, and then enters state S2. In state S2, if in PMT mode, it calculates and stores the local fields of all right spins, updates all right spin states, stores the flipped right spin indices, and then enters state S3. If in PT or SA mode, it calculates and stores the local fields of all right spins and then enters state S4. State S3 exists only in PMT mode. It only updates the flipped right spin indices based on the existing left spin local fields. The process involves calculating compensation for the change in the local field of the left spin caused by the right spin rotation, updating all left spin states, storing the index of the flipped left spin, and then entering state S4. In state S4, if it is PMT mode, compensation is calculated only for the change in the local field of the right spin caused by the flipped left spin, based on the original local field of the right spin. Then, all right spin states are updated, the index of the flipped right spin is stored, and the replica energy is calculated, entering state S5. If the iteration counter meets the preset number of exchange steps at this point, state S6 is entered. If it is PT mode, a right spin update operation is performed, i.e., a right spin is selected based on the index generated by the random number generator, the spin is flipped according to the flipping criterion, and all right spin local fields are updated. After reaching the preset number of update steps, state S5 is entered. If the preset number of exchange steps is met at this point, state S6 is entered. If it is SA mode, a right spin update operation is performed, and after reaching the preset number of update steps, state S6 is entered. In S5 state, replica swapping is performed according to the backfire replica swapping criteria. After the swap, if the device is in PMT mode, it returns to S3 state; if it is in PT mode, it returns to S4 state.

[0016] The reconfigurable pulsating array is used to calculate spin energy and update spin states. It includes 4×32 processing units (PEs), one SA updater, and a PMT updater and local energy accumulator configured in a row of four PEs. The PE is the main computational component, consisting of a three-stage pipeline. It includes 16 spins (8 left spins and 8 right spins), a memory for 8×1024 2-bit spin weight coefficients, and local field memories corresponding to the 16 spins (8 left spins and 8 right spins), as well as selectors, shifters, adders, registers, XOR gates, and NOT gates. The processing unit (PE) consists of a three-stage pipeline: the first stage pipeline obtains the address i_addr corresponding to the weight of the i-th spin by flipping the indices i and j of the spin register; the second stage pipeline reads the weight from the i_addr of the previous stage. ij The 64 bits of the corresponding memory address J i_addr Then, based on the index j of the previous level, read the corresponding 2-bit spin weighting coefficient J. ij And select the addend through configuration information; J was selected during initialization. ij For the 0th position, when updating the state, select the J position shifted one position to the left. ij The corresponding spin σ j With J ij If the XOR result of the sign bit is 1, the result is directly passed to the next stage; otherwise, the result is passed to the next stage through bitwise NOT and addition operations using a NOT gate. The third-stage pipeline then modifies the result from the previous stage with the result in the local field F. i F read from memory i Add them together, calculate and store the new F. i .

[0017] The PMT updater consists of an ω memory, registers, selectors, NOT gates, adders, a sign bit deselector, and an equality checker. Taking state S3 as an example, the input index i reads the corresponding ω from the ω memory. i , and enter F i Random number γ, corresponding right spin σ R and left spin σ L After the selection operation, the result enters the adder, and the sign of the adder's output determines the new spin σ. new The state; σ new By using the above σ L The comparison determines whether the spin has been flipped; the i corresponding to the flipped spin is output as an index to the flipped spin register; the SA updater consists of a shifter, multiplexer, AND gate, NOT gate, adder, sign bit extractor, counter, NOT gate, and register; the SA updater outputs the row and column information of the input index i to the systolic array through addressing, and the systolic array returns the F corresponding to the index. iThe sign bit of the adder result is used to determine whether the spin should be flipped. If it is flipped, an input data set is generated and fed into the pulsating array for F... i The update process uses a counter to determine if the update is complete. Once the update is complete, an enable signal is generated to continue receiving the selected index and the update step count is incremented by 1. If the index is not flipped, an enable signal is generated directly and the update step count is incremented by 1.

[0018] The pulsating array computation mode is controlled by signal S in the global controller, which controls the algorithm mode. There are three computation modes: PMT, PT, and SA. In the PMT computation mode, during states S1-S4, the data group (i,j,σ) generated by the toggling spin register is transmitted hop-by-hop to all PEs in a row of the subarray. The PEs then store the results in the corresponding F. i The memory transmits the data in the opposite direction to the input data to the PMT updater of each row, generating a flipped spin index, which is then stored in the flipped spin register, replacing j in the input data group of the next state. If the replica energy is also calculated simultaneously in states S2 and S4, and the preset number of swap steps has not been reached after the end of the state, it jumps to state S5. In state S5, replica swapping is performed according to the replica energy and swapping criteria, and then it jumps back to state S3. In the PT calculation mode, after state S2, it enters state S4, where a spin index is generated by random numbers, and the corresponding F of the pulsating array is found through addressing. i The SA updater determines whether a flip has occurred. If not, it continues to generate and read the next spin index. If a flip has occurred, it inputs the data set into the pulsating array to update each F. i After the value is updated, the next spin index is read. If the preset update step number is reached but the preset exchange step number is not reached, the replica energy is calculated and the process jumps to the S5 state. In the S5 state, replicas are exchanged according to the replica energy and exchange criteria, and the process jumps back to the S4 state. The SA calculation mode is similar to the PT calculation mode, except that the process jumps to the value completion state after the preset update step number is reached.

[0019] The fully connected Ising model reconfigurable processing circuit supporting multiple algorithms described in this invention has at least the following technical advantages:

[0020] (1) Supports multiple algorithms and has a high convergence speed, enabling it to quickly solve a wider range of combinatorial optimization problems: By configuring different global controllers in S to implement different computation modes, multiple algorithms are implemented in one architecture to solve a wider range of combinatorial optimization problems; In PMT mode, by storing the flipped spin index and using it for the next local field update, it is not necessary to traverse all spin states, which greatly reduces the time for each spin update state. Furthermore, as the number of flipped spins decreases with the increase of iteration steps, the single-step update time decreases with the increase of iterations, which greatly improves the convergence speed.

[0021] (2) It has high scalability. Each processing unit can handle 8 spins. Each row of the array consists of 4 processing units. By simply increasing the number of rows, it can handle larger-scale combinatorial optimization problems.

[0022] (3) Low complexity and low hardware resource cost. The architecture adopts a special design: In the pulsating array, each unit processes multiple spins and each row of the array shares a PMT updater, which reduces resource consumption while ensuring computing speed; In the PMT updater, a random number generator based on read-only memory is used instead of multipliers to generate random numbers for temperature coefficients. In the SA updater, adders, shifters and selectors are used instead of multipliers to generate toggle signals, which greatly reduces the resource consumption caused by multipliers; In the processing unit, a distributed storage and near-memory computing structure is adopted, which simplifies the structure of the processing unit, reduces the communication traffic load, reduces the computational complexity, and improves the computational efficiency. At the same time, the processing unit uses a fully pipelined structure, combined with a selector determined by configuration to select addends to avoid multiple adders, which greatly reduces hardware overhead.

[0023] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

Claims

1. A fully connected Ising model reconfigurable processing circuit supporting multiple algorithms, characterized in that, include: Eight replicas, a global controller, a temperature exchange module, a random number generator, and a temperature generator, implementing different algorithms through different computation modes; Each of the eight replicas consists of a 4×32 pulsating array supporting 1024 left spins and 1024 right spins, a selector controlling the PMT updater and SA updater, an energy calculator, and a flip-spin register. The pulsating array supports reconfigurable functions and supports different algorithms depending on the configuration; the selector selects the PMT updater or SA updater to run according to functional requirements; the energy calculator calculates the total energy of the system; the register stores the inverted spin index after the update; The global controller includes an I / O module, an iteration step counter, a configuration allocator, and a finite state machine. The I / O module is responsible for exchanging information between the user and the processor architecture. The iteration step counter is responsible for calculating the current iteration step of the system. The configuration allocator configures global parameters and processes based on the finite state machine state and user input configuration. The finite state machine performs state transitions based on the iteration step and configuration. The global controller has seven working states: idle state S0, left spin initialization state S1, right spin initialization state S2, left spin update state S3, right spin update state S4, replica exchange state S5, and completion state S6. In state S0, all components are in standby mode. When the user's mode signal S is not 00, the idle state is switched to state S1 or state S2 according to the value of S. The S1 state exists only in PMT mode. It calculates and stores the local field of all left spins, updates all left spin states, and then enters the S2 state. In state S2, if in PMT mode, calculate and store the local fields of all right spins, update all right spin states, store the flipped right spin indices, and then enter state S3; if in PT or SA mode, calculate and store the local fields of all right spins, and then enter state S4. The S3 state only exists in PMT mode. Based on the original left spin local field, it only compensates for the change in the left spin local field caused by the flipped right spin, updates all left spin states, stores the index of the flipped left spin, and then enters the S4 state. In state S4, if it is PMT mode, the compensation calculation is performed only on the local field change of the right spin caused by the flipped left spin, based on the original right spin local field. Then, all right spin states are updated, the index of the flipped right spin is stored and the copy energy is calculated. If the iteration counter does not meet the preset number of swap steps at this time, it enters state S5; if the iteration counter meets the preset number of swap steps at this time, it enters state S6. If it is in PT mode, a right spin update operation is performed. That is, after generating an index according to the random number generator, a right spin is selected according to the generated index, a spin flip operation is performed according to the flip criterion, and all right spin local fields are updated. After reaching the preset number of update steps, if the preset number of exchange steps is not met at this time, it enters the S5 state; if the preset number of exchange steps is met at this time, it enters the S6 state. If it is in SA mode, a right spin update operation is performed. After reaching the preset number of update steps, it enters the S6 state. In S5 state, replica swapping is performed according to the backfire replica swapping criteria. After the swap, if the system is in PMT mode, it returns to S3 state; if the system is in PT mode, it returns to S4 state. The pulsating array is used to calculate spin energy and update spin states. It contains 4×32 processing units (PEs), one SA updater, and a row of 4 PEs configured with one PMT updater and a local energy accumulator. The processing unit (PE) consists of a three-stage pipeline: the first stage pipeline obtains the address i_addr of the weight corresponding to the i-th spin by flipping the indices i and j of the spin register input; the second stage pipeline reads j from the i_addr of the previous stage. ij The corresponding 64-bit address J of the memory i_addr Then, based on the index j of the previous level, read the corresponding 2-bit spin weighting coefficient J. ij The addend is selected through configuration information, and the spin weighting coefficient J is selected during initialization. ij For the 0th bit, when updating the state, the spin weighting coefficient J shifted one bit to the left is selected. ij The corresponding spin σ j With spin weighting coefficient J ij If the XOR result of the sign bit is 1, the result is directly passed to the next stage; otherwise, the result is inverted bit by bit through the NOT gate, then added and passed to the next stage. The third stage pipeline adds the result of the previous stage to the corresponding local field read from the local field memory, calculates and stores the new local field.