Closed-loop regulation calculation system and equipment

Abstract

The invention discloses a closed-loop regulation calculation system and equipment, and relates to the technical field of neural networks, the closed-loop regulation calculation system comprises a digital time coding module, a digital control oscillation module, a phase accumulation module, an energy detection module and an annealing control module; the digital time coding module is used for coding the received spin data into time pulse data and sending the time pulse data to the phase accumulation module; the digital control oscillation module is used for converting the received weight data into frequency signal data and sending the frequency signal data to the phase accumulation module; the phase accumulation module is used for calculating the received time pulse data and frequency signal data to obtain a calculation result and sending the calculation result to the energy detection module; and the energy detection module is used for evaluating a system energy value of the closed-loop regulation calculation system according to a calculation result, and updating spinning data through the annealing control module. According to the invention, the technical problems of high complexity and poor precision of a full-connection Ising model during implementation are solved.

Description

Technical Field

[0001] The present application relates to the field of neural network technology, and in particular to a closed-loop regulation computing system and device. Background Art

[0002] The Ising model, a universal modeling framework, has garnered widespread attention in recent years for solving numerous complex problems, including combinatorial optimization, neural network training, and graph computing. The Ising machine, the hardware solver for this model, aims to efficiently approximate the optimization problem by exploring the coupling relationships between spin variables to find the minimum energy state of the system. However, in a fully connected Ising model, coupling relationships exist between every node. While this fully connected nature theoretically offers enhanced modeling capabilities and convergence performance, it also presents significant implementation challenges.

[0003] Specifically, traditional digital circuits must handle a large number of coupling terms when implementing fully connected structures. This leads to a quadratic increase in hardware resource overhead and system size, resulting in a sharp increase in power consumption and area. When faced with large-scale problems, such architectures are prone to poor scalability, poor solution accuracy, and complex structures.

[0004] The above information disclosed in this Background section is only for understanding the background of the present invention and therefore it may contain information that does not constitute prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a closed-loop regulation calculation system and equipment, aiming to solve the technical problems of high complexity and poor accuracy in the implementation of the current fully connected Ising model.

[0006] To achieve the above objectives, the present application provides a closed-loop regulation computing system, comprising a digital time encoding module, a digitally controlled oscillation module, a phase accumulation module, an energy detection module, and an annealing control module, wherein the digital time encoding module and the digitally controlled oscillation module are respectively connected to the phase accumulation module, the phase accumulation module is connected to the energy detection module, and the energy detection module is connected to the annealing control module;

[0007] The digital time encoding module is used to encode the received spin data into time pulse data and send it to the phase accumulation module;

[0008] The digitally controlled oscillation module is used to convert the received weight data into frequency signal data and send it to the phase accumulation module;

[0009] The phase accumulation module is configured to calculate the received time pulse data and frequency signal data to obtain a calculation result, and send the calculation result to the energy detection module.

[0010] The energy detection module is configured to evaluate a system energy value of the closed-loop regulation calculation system according to the calculation result, and update spin data through the annealing control module.

[0011] In an embodiment, the phase accumulation module includes a plurality of multiply-accumulate calculation units connected in series. In the process of calculating the time pulse data and the frequency signal data, the phase accumulation module determines a starting accumulation position from each of the multiply-accumulate calculation units according to a bit weight level corresponding to the frequency signal data, and performs phase accumulation on a phase increment mapped by the time pulse data and the frequency signal data based on the starting accumulation position.

[0012] In an embodiment, each of the multiply-accumulate calculation units in the phase accumulation module is activated in sequence based on a preset clock period, and in the calculation process of the phase accumulation module, a calculation result is output by one multiply-accumulate calculation unit in each clock period.

[0013] In an embodiment, the closed-loop regulation calculation system further includes a spin data storage array and a weight data storage array. The spin data storage array is connected to the digital time encoding module and the annealing control module respectively, and the weight data storage array is connected to the digital control oscillation module.

[0014] The annealing control module is configured to update spin data in the spin data storage array according to the system energy value after receiving the system energy value sent by the energy detection module.

[0015] In an embodiment, the annealing control module includes at least a closed-loop temperature control unit, a random disturbance source connected to the closed-loop temperature control unit, and a state update processing unit.

[0016] The closed-loop temperature control unit is configured to receive the system energy value sent by the system energy detection unit, and determine a system control temperature value according to the system energy value.

[0017] The random disturbance source is configured to generate a disturbance random number, and the disturbance random number and the system control temperature value are input to the state update processing unit.

[0018] The state update processing unit is configured to determine whether to update spin data according to the system control temperature value, and determine a spin data update value according to the system control temperature value and the random disturbance number when it is determined to update the spin data.

[0019] In one embodiment, the state update processing unit is further used to determine whether the energy decreases after the spin data is updated based on the system control temperature value. If so, determine to update the spin data; if not, calculate the update probability based on the energy change amplitude after the spin data is updated and the system control temperature value, and determine whether to update the spin data based on the update probability.

[0020] In one embodiment, the closed-loop regulation computing system is further used to initialize and generate multiple copies, and operate each of the copies at multiple temperatures;

[0021] The energy detection module is used to detect the total energy value corresponding to each of the replicas and send it to the annealing control module;

[0022] The annealing control module is used to generate an updated energy value based on the total energy value and the random perturbation number and determine the energy change amplitude of each replica. When the energy change amplitude is lower than 0, the updated energy value is determined as the new total energy value; when the energy change amplitude is greater than or equal to 0, whether to determine the updated energy value as the new total energy value is determined according to the update probability.

[0023] In one embodiment, the annealing control module is further used to determine whether the number of iterative updates of the current energy value is greater than or equal to a first preset number after determining a new total energy value or determining not to update the energy value; if not, determine whether the energy change amplitude is greater than a preset change threshold; if greater, adjust the temperature value of the replica based on a first cooling rate; if not, adjust the temperature value of the replica based on a second cooling rate, the first cooling rate is greater than the second cooling rate, and return to execute the step of generating an updated energy value according to the total energy value and the random disturbance number and determining the energy change amplitude of each replica; if not, output the current replica state.

[0024] In one embodiment, the annealing control module is further configured to, after the number of iterative updates of the energy value is greater than a first preset number, exchange the states corresponding to each of the replicas according to the update probability, and then determine whether the current total number of exchanges is greater than or equal to a second preset number; if not, return to the step of generating an updated energy value based on the total energy value and the random perturbation number and determining the energy change amplitude of each of the replicas; if not, output the spin data of the optimal replica among the replicas.

[0025] In addition, to achieve the above-mentioned purpose, the present application also provides a closed-loop regulation computing device, which includes the closed-loop regulation computing system as described above.

[0026] The present application provides a closed-loop regulation computing system, which includes a digital time coding module, a digitally controlled oscillation module, a phase accumulation module, an energy detection module, and an annealing control module. The digital time coding module and the digitally controlled oscillation module are respectively connected to the phase accumulation module, the phase accumulation module is connected to the energy detection module, and the energy detection module is connected to the annealing control module; the digital time coding module is used to encode received spin data into time pulse data and send it to the phase accumulation module; the digitally controlled oscillation module is used to convert received weight data into frequency signal data and send it to the phase accumulation module; the phase accumulation module is used to calculate the received time pulse data and frequency signal data to obtain a calculation result, and send the calculation result to the energy detection module; the energy detection module is used to evaluate the system energy value of the closed-loop regulation computing system according to the calculation result, and update the spin data through the annealing control module. Among them, in the technical solution of the present application, in terms of time-frequency calculation, the spin data is encoded as a time pulse, the coupling strength is encoded as a frequency signal, and the multiplication and addition calculation is realized through phase accumulation, avoiding the complexity and high power consumption of the traditional digital multiplier structure, and effectively reducing the calculation complexity. Moreover, the energy detection module and the annealing control module are introduced in the present application to perform energy feedback, which can adjust and update the spin data in real time based on the detected system energy value, can adjust the direction of spin evolution, effectively avoid local optimal traps, improve system accuracy, and take into account low power consumption, high efficiency and reconfigurability, etc., which is suitable for the rapid solution of large-scale combinatorial optimization problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0028] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0029] Figure 1 This is a schematic diagram of the time-frequency calculation framework corresponding to the closed-loop regulation calculation system in an embodiment of the present application;

[0030] Figure 2 A schematic diagram of a plurality of multiplication-accumulation computing units connected in series in an embodiment of the present application;

[0031] Figure 3 Schematic diagram of the structure of the enabling unit in the pipeline preparation stage in an embodiment of the present application;

[0032] Figure 4 This is a data waveform diagram of the enabling unit in the pipeline preparation stage in an embodiment of the present application;

[0033] Figure 5 This is a waveform diagram of the data flow of the fully connected Ising machine pipeline in an embodiment of the present application;

[0034] Figure 6 This is a diagram of the architecture of a fully connected Ising machine with reconfigurable precision and closed-loop regulation in a time-frequency pipeline provided in an embodiment of the present application;

[0035] Figure 7 A schematic diagram of a coupling weight storage array provided in an embodiment of the present application;

[0036] Figure 8 Schematic diagram of the closed-loop adaptive adjustment principle of the annealing control module in an embodiment of the present application;

[0037] Figure 9 This is a flow chart of a closed-loop annealing control mechanism based on system energy drive provided in an embodiment of the present application.

[0038] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0039] To make the above-mentioned purposes, features, and advantages of the present application more clearly understood, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present application without making any creative work are within the scope of protection of this application.

[0040] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0041] In order to better understand the technical solution of this application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0042] At present, in the solution process of many complex problems such as combinatorial optimization, neural network training and graph calculation, Ising model as a general modeling framework has been widely concerned in recent years. Ising machine as a hardware solver of the model, its goal is to seek the minimum state of system energy through the coupling relationship between spin variables, so as to realize the efficient approximation of optimization problem. Especially in the fully connected Ising model, there is a coupling relationship between each node. Such fully connected characteristics can provide stronger modeling ability and convergence performance in theory, but at the same time it brings great challenges in implementation. On the one hand, when implementing the fully connected structure, the traditional digital circuit needs to process a large number of coupling terms, and the hardware resource overhead increases quadratically with the system size, resulting in a sharp rise in power consumption and area. Such architecture is prone to poor scalability, high power consumption and complex structure when facing large-scale problems, which is difficult to meet the application requirements of low power consumption and high integration. In addition, digital logic also needs to cooperate with a large number of timing control and register resources, further increasing the system burden. Therefore, the existing technology has great limitations in implementing the fully connected Ising machine. On the other hand, the solution of energy function depends on the weighted summation operation of multiple coupling terms, and the computational complexity is O(N 2 ). In the existing system, the summation is usually performed in a serial or partially parallel manner, but such methods are difficult to balance speed and resource overhead at the same time. With the increase of the number of nodes, the time required for summation operation increases significantly, which becomes the bottleneck of system throughput, seriously affecting the overall computing performance. Therefore, a data processing method that can efficiently parallelize and support high throughput is needed to alleviate the problem of limited computing speed.

[0043] Furthermore, the precision requirement of coupling strength varies with specific problems, and using a uniform bit width to represent all coupling terms will result in a significant decrease in hardware resource utilization. Strong coupling terms may affect the modeling accuracy due to insufficient bit width, while weak coupling terms waste a large number of redundant bits. In the existing architecture, fixed precision or fixed point format is usually used to represent the coupling strength, and there is a lack of flexible precision configuration mechanism, resulting in uneven resource usage and low efficiency. Therefore, developing a mechanism that can dynamically allocate precision according to the importance of coupling strength is of great significance to improve the overall system efficiency. In addition, the solution process of Ising model has heuristic search characteristics, which is easily affected by initial conditions and evolution path, and then falls into local optimal solution. The traditional method lacks real-time feedback mechanism for energy convergence trend, which makes it difficult for the system to jump out of the suboptimal point in the complex energy surface, affecting the final solution accuracy and reliability. Therefore, in the technical scheme of the embodiments of the present application, the system energy is introduced as a feedback index, and a closed-loop regulation mechanism is designed to improve the solving ability of Ising machine and enhance its approximation ability to global optimum.

[0044] In summary, the existing full connection Ising machine implementation has obvious limitations in power consumption, speed, hardware resource utilization, and solution accuracy. Therefore, a new hardware architecture with low power consumption, high scalability, high throughput, adjustable accuracy, and intelligent adjustment capability is needed to improve the practical application value of the full connection Ising machine.

[0045] To solve the above problems, the embodiments of the present application provide a closed-loop regulation computing system, referring to Figure 1 , Figure 1 The figure is a schematic diagram of a time-frequency computing framework corresponding to the closed-loop regulation computing system of the present application. The closed-loop regulation computing system includes a digital time encoding module 100, a digital control oscillation module 200, a phase accumulation module 300, an energy detection module 400, and an annealing control module 500. The digital time encoding module 100 and the digital control oscillation module 200 are connected with the phase accumulation module 300, the phase accumulation module 300 is connected with the energy detection module 400, and the energy detection module 400 is connected with the annealing control module 500.

[0046] The digital time encoding module 100 is used to encode the received spin data into time pulse data and send it to the phase accumulation module 300.

[0047] The digital control oscillation module 200 is used to convert the received weight data into frequency signal data and send it to the phase accumulation module 300.

[0048] The phase accumulation module 300 is used to calculate the received time pulse data and frequency signal data to obtain a calculation result and send the calculation result to the energy detection module 400.

[0049] The energy detection module 400 is used to evaluate the system energy value of the closed-loop regulation computing system according to the calculation result and update the spin data through the annealing control module 500.

[0050] The closed-loop regulation computing system provided by the embodiments of the present application is equivalent to an implementation structure of a full connection Ising machine based on a time-frequency computing architecture. The spin data includes spin states, which are encoded into time pulse data by a digital time encoding method. The weight data includes coupling weights, which are converted into corresponding frequency signals by a digital control oscillation module (such as a digital control oscillator). Then, both of them are input into the corresponding phase accumulation calculation unit of the phase accumulation module, and the product operation between time and frequency can be realized through the phase accumulation module.

[0051] In the phase accumulation operation, the duration is programmable as the pulse width, and the frequency is expressed as the number of oscillations per unit time. Therefore, in the phase accumulation calculation, the result of time × frequency can be interpreted as a phase increment, that is, a multiplication operation. The multiplication-accumulation operation is then performed by superimposing the phase increments over multiple clock cycles. All calculation results are then aggregated to the system energy detection module, which provides real-time assessment of the current system energy state. Finally, the annealing control module adjusts the system temperature based on energy trends, achieving closed-loop updates and adjustments to the spin state, helping the system escape local optimal solutions and improving overall solution performance.

[0052] It's important to note that in the Ising model, the system's energy state can be expressed through two core parameters: the coupling constant (J) and the external magnetic field strength (H). The system temperature (T), adjusted by the annealing control module, represents the intensity of the ambient thermal noise, thereby controlling the randomness of the spin state. Furthermore, the annealing process dynamically adjusts the system temperature T, guiding the system's evolution from high randomness to low energy. The temperature T in the Ising model does not refer to actual heat, but rather to parameters such as the ambient noise intensity, the Boltzmann distribution scale factor, and the phase transition control parameter.

[0053] After receiving the input spin state data and coupling weights, the closed-loop control calculation system enters the data conversion phase. The coupling weights are converted into frequency signals proportional to the numerical amplitude via a digitally controlled oscillator (DCO). The spin state data is then processed by a digital time encoding module to generate a time pulse signal of corresponding polarity and duration. This time pulse serves as a unified control signal and is multiplexed throughout the multiplication-accumulation calculation array of the phase accumulation module, significantly reducing control complexity and increasing computational parallelism. In the multiplication-accumulation calculation, the time signal and frequency signal are jointly input to the phase accumulator unit, performing an equivalent multiplication operation. By mapping time × frequency into phase increments and accumulating the phases, a multiplication-accumulation (MAC) operation is performed to obtain the calculation result.

[0054] The embodiment of the present application essentially constructs a multiplication and accumulation calculation circuit based on time-frequency coding, which is used to simulate the multiplication and accumulation operation between the spin state and the coupling weight in the Ising model. In this structure, the spin state is encoded as a time pulse signal, and the coupling weight is converted into a frequency signal by a digitally controlled oscillator, and the two are input into the phase accumulator for calculation. The phase accumulation module can realize the product of time and frequency, that is, obtain the phase quantity representing the coupling energy, and accumulate the phase quantity to complete the entire multiplication and accumulation process. The time-frequency coding multiplication and accumulation calculation circuit designed for the Ising model, combined with the digital characteristics of the Ising machine spin and coupling weight, the system avoids the complexity and high power consumption of the traditional digital multiplier structure, and provides an efficient and scalable hardware multiplication and accumulation implementation method, which is suitable for large-scale integration and asynchronous computing scenarios.

[0055] Furthermore, in a feasible embodiment, the phase accumulation module includes multiple multiplication and accumulation calculation units connected in series. In the process of calculating the time pulse data and the frequency signal data, the phase accumulation module determines the starting accumulation position from each multiplication and accumulation calculation unit according to the bit weight level corresponding to the frequency signal data, so as to perform phase accumulation on the phase increment mapped by the time pulse data and the frequency signal data based on the starting accumulation position.

[0056] like Figure 2 As shown, the phase accumulation module includes multiple end-to-end connected and serially connected multiplication and accumulation calculation units (such as Q0, Q1, Q2, Q3). Each multiplication and accumulation calculation unit includes components such as a D flip-flop and a selector. The D flip-flop includes ports such as D (data input), Clk (clock), Reset (reset), Q (output), and Q- (inverted output). Its functions include: when the rising edge (or falling edge) of the clock signal CLK[0] is triggered, the value of the input D is latched to the output terminal Q, Outputs the inverted signal of Q. The Reset port is used to forcibly reset the output to its initial state (usually 0). In addition, CLK[0] is the clock signal for the D flip-flop, triggering the data latch. Rese is the global reset signal, used to forcibly clear the output of the D flip-flop. IN[0] is the control signal for the left selector (i.e., HOLD LOAD), which determines whether to hold or load data. UP / DOWN is the control signal for the right selector, which determines the counting direction (increment / decrement).

[0057] In order to improve the flexibility of the system, the technical solution of the embodiment of the present application also supports a bit-weighted precision expansion mechanism, which can select the starting accumulation position of the phase accumulator according to the bit weight level of the weight, and realize reconfigurable calculations with multiple precisions such as 5bit and 9bit, taking into account energy efficiency and precision requirements.

[0058] For example, the bit weight level corresponding to the frequency signal data can be directly obtained by determining the coupling weight of the frequency signal data. The bit weight level can be represented by a natural number. A higher bit weight level corresponds to a larger natural number, and the starting accumulation position in the phase accumulation module is positioned further to the right. For example, the bit weight levels can include 1, 2, 3, and 4, and their corresponding starting accumulation positions are Q1, Q2, Q3, and Q4, respectively.

[0059] The present embodiment introduces a precision reconfiguration mechanism in the phase module. Specifically, by designing different accumulation starting positions in the phase accumulator, coupling weights are input according to bit weights, thereby supporting multiple precision forms (such as 5-bit and 9-bit). This mechanism can flexibly control the starting point of the phase accumulation process based on the bit weight of the current input coupling weight, completing the bit-by-bit expansion of high-precision multiplication. It also significantly improves the utilization of hardware resources, allowing the system to effectively control circuit area and power consumption while maintaining high computational accuracy, enhancing the versatility and flexibility of the architecture.

[0060] In a feasible embodiment, each multiplication-accumulation calculation unit in the phase accumulation module is activated in sequence based on a preset clock cycle, and during the calculation process of the phase accumulation module, a multiplication-accumulation calculation unit outputs a calculation result in each clock cycle.

[0061] To further improve throughput, the phase accumulation module of the present embodiment adopts a pipeline execution mechanism. During the startup phase of the computing units, the computing units are activated sequentially from left to right, realizing the spatial and temporal expansion of the computing task. After the computing units enter the stable operation phase, only one computing unit completes the output in each clock cycle, and the corresponding calculation results are continuously produced in sequence according to the pipeline order. This allows N spin states to complete a round of parallel updates within N cycles, effectively reducing the computational complexity from the traditional O(N2) to O(N).

[0062] For example, in a sequential pipeline structure, the spin state update operation is expanded in time sequence, thereby constructing a pipeline array composed of multiple multiplication and accumulation units in series. In this architecture, the calculation of each spin state is completed by an independent computing unit in different cycles. The entire system advances from left to right in cycle by cycle, and the start signal is also transmitted in cycle by cycle. The structure of the enabling unit in the pipeline preparation stage is as follows: Figure 3As shown in the figure, it includes a D flip-flop, ENI represents the enable input, which controls whether the element responds to the clock signal (high level is valid); CLK represents the clock input, which receives the external clock signal and triggers the data latch operation (rising edge or falling edge trigger); D represents the data input port, which is the data input port to be stored or processed; Q represents the data output), which is the direct output port of the stored data; ENO represents the enable output, which is the cascade output signal (usually high level) used to control the enable of the next level element; PULSE represents the output start signal. In addition, the waveforms of various data in the enable unit in the pipeline preparation stage are shown as follows Figure 4 As shown, the final output start signal is in the form of a pulse signal.

[0063] For example, based on the above structure, a fully connected Ising machine pipeline data flow waveform is as follows: Figure 5 As shown in the figure, the figure includes multiple waveforms such as PULSE, PE1, PE2, PE3, PE4, PULSE_DE, HSEL[0], H REG, FA State, Add out, and Sum Out, among which PULSE is the waveform of a fixed pulse signal, PE1, PE2, PE3, and PE4 refer to the serial numbers of the processing units (Q1-Q4) respectively, PULSE_DE is the delayed signal of the PULSE signal, HSEL is the Hamiltonian selection signal, H REG is the register signal for storing the Hamiltonian, FA State is the state of the full adder, Add ou is the output accumulation result signal, and Sum Out is the final accumulation result signal, among which h is used to represent energy, H1-H4 are used to represent the calculation results corresponding to each processing unit, j represents the weight, and h1+j 12 Expressions such as σ2 are calculation expressions of the processing unit during the calculation process, and σ is a coefficient.

[0064] Based on the enable signal output by the enable unit and transmitted in sequence according to the cycle, the time complexity can be reduced from O(N 2 ) is reduced to O(N). This structure greatly improves the system throughput and parallel capability, reduces the complexity of local interconnection, and facilitates modular design and resource reuse.

[0065] In a feasible embodiment, the closed-loop regulation calculation system further includes a spin data storage array and a weight data storage array, the spin data storage array is connected to the digital time encoding module and the annealing control module respectively, and the weight data storage array is connected to the digital control oscillation module;

[0066] The annealing control module is used for updating the spin data in the spin data storage array according to the system energy value after receiving the system energy value sent by the energy detection module.

[0067] In the embodiment of the present application, the spin state and the coupling weight are stored respectively by the spin data storage array and the weight data storage array. Figure 6 As shown, the spin state storage array stores multiple spin states, the coupling weight storage array (i.e., the weight data storage array) stores multiple coupling weights, the digital time encoding module can be a digital time encoding unit, the digitally controlled oscillation module can be a digitally controlled oscillation array composed of multiple digital controller oscillators, and the phase accumulation module is a multiplication-accumulation calculation array composed of multiple phase accumulation calculation units. The multiplication-accumulation calculation array outputs the calculation results to the system energy detection module, which transmits the detected energy state to the annealing control module, so that the annealing control module can update the spin state based on the energy state. In summary, the above modules form a fully connected Ising machine architecture with reconfigurable time-frequency pipeline and closed-loop regulation.

[0068] On the basis of the above structure, first, in the data loading and recycling stage, the system realizes the rapid loading of the initial coupling weights and initial spin states through the serial input parallel output shift register (Serial-In Parallel-Out, SIPO) structure. In the subsequent calculation process, the coupling weight data enters each multiplication and accumulation calculation unit through the parallel input parallel output mode (Parallel-In Parallel-Out, PIPO), supporting the cyclic call of the calculation stage. The specific switching between serial input and parallel input modes is controlled by the selector. The specific structure is shown. At the same time, the spin state data is also loaded in parallel, but it needs to cooperate with the annealing control module for state analysis and update to ensure effective control of the state update path during the system annealing process. For example, the structure of the coupling weight storage array is as follows Figure 7 As shown, it includes multiple D flip-flops, where D represents input, CLK represents clock, Q represents output, DOUT represents serial output, which is the final output of the shift register chain, CLK_P represents parallel clock, which is used in parallel data loading mode (writing external data directly into the flip-flop), CLK_S represents serial clock, which is used in serial shift mode, DIN represents serial input, J[0], J[1], J[2], J[3], etc. are parallel data input / internal nodes, which can be used as parallel loading ports and flip-flop output nodes; in addition, Sel_S2P is a selector with the function of selecting multiple data inputs at the signal end.

[0069] Furthermore, in a feasible embodiment, the annealing control module at least includes a closed-loop temperature control unit, a random disturbance source connected to the closed-loop temperature control unit, and a state update processing unit;

[0070] The closed-loop temperature control unit is used to receive the system energy value sent by the system energy detection unit and determine the system control temperature value according to the system energy value;

[0071] The random disturbance source is used to generate a disturbance random number, and the disturbance random number and the system control temperature value are input into the state update processing unit together;

[0072] The state update processing unit is used to determine whether to update the spin data according to the system control temperature value, and determine the spin data update value according to the system control temperature value and the random disturbance number when determining to update the spin data.

[0073] In the embodiment of the present application, the annealing control module monitors the system energy changes in real time and dynamically adjusts the system temperature to achieve efficient management and control of the state evolution process, thereby achieving effective optimization and precise control of the system state. The system energy detection module continuously monitors the trend of total system energy changes to evaluate the stability of the current system and the degree to which it deviates from the optimal state. Among them, energy changes are important feedback signals for the evolution of the system state and provide a basis for subsequent temperature control. Figure 8 As shown, the system energy value is first delivered to the closed-loop temperature control module. For example, when the system energy is high, the closed-loop temperature control module will increase the system temperature value, allowing the system to conduct a large-scale exploration in the solution space, thereby avoiding falling into a local optimum; and when the energy gradually decreases and the system tends to be stable, the system temperature value is then reduced, prompting the system to gradually converge to the global optimal state. In addition, in order to simulate the thermal fluctuation behavior in the physical system, a random disturbance source is also introduced in the embodiment of the present application. A pseudo-random number generator based on a linear feedback shift register (LFSR) is used to simulate thermal noise to introduce disturbances, thereby ensuring randomness and exploration capabilities during the state update process and improving the global optimization capability of the algorithm. Then, the disturbance random number and the system control temperature value output by the closed-loop temperature control module are input to the state update processing unit. The state update processing unit first determines whether to update the spin value, and then determines the spin data update value according to the system control temperature value and the random disturbance number when determining to update the spin data.

[0074] For example, the energy detection module is used to monitor the total energy changes of a fully connected Ising system in real time. Specifically, the outputs of the multiplication and accumulation units in each cycle of the system are input into the energy detection path, and the total system energy H at the current moment is accumulated. This module serves as a pre-reference for annealing control, providing a closed-loop feedback signal for subsequent temperature control, thereby dynamically reflecting the stability and convergence trends of the system at different stages. This module enables accurate assessment of the system's current solution state, improving the responsiveness and adaptability of the annealing process.

[0075] For example, the closed-loop temperature control module can dynamically adjust the system's "temperature" variable based on the total energy changes output by the system status monitoring module to guide the system's search direction and convergence speed. This control mechanism is designed to imitate the physical annealing process: when the system energy is high, increasing the temperature gives the system greater freedom and allows it to escape the local optimum; when the system energy tends to stabilize, lowering the temperature suppresses large fluctuations and encourages the solution to converge toward the optimal state. By implementing energy-driven temperature regulation, a closed-loop annealing system that collaborates with software and hardware can be constructed, significantly enhancing the Ising solver's global search capabilities for large-scale problems.

[0076] For example, a random perturbation source is used to simulate thermal noise perturbations in a physical thermodynamic system, and a linear feedback shift register is used to implement hardware-based pseudorandom number generation. This pseudorandom sequence, acting as a thermal noise signal, participates in the spin update decision process, providing a random perturbation basis for state transitions. The introduction of thermal perturbations effectively prevents the system from falling into local minima at low temperatures, thereby improving the ability to traverse the overall solution space. This module offers advantages such as simple structure, strong adjustability, and low resource consumption, making it suitable for integration into low-power hardware platforms.

[0077] Furthermore, the state update processing unit is also used to determine whether the energy decreases after the spin data is updated based on the system control temperature value. If so, determine to update the spin data; if not, calculate the update probability based on the energy change amplitude after the spin data is updated and the system control temperature value, and determine whether to update the spin data based on the update probability.

[0078] In this embodiment, the state update processing unit determines whether to update spin data using the Metropolis criterion (a probabilistic acceptance strategy): The system, under the influence of current temperature and thermal noise, determines whether to accept a spin flip based on the energy increase or decrease caused by the state change, thereby implementing a state transition mechanism based on energy, temperature, and random perturbations. This allows the system to effectively achieve adaptive state optimization, approaching the optimal solution within a finite period.

[0079] For example, according to the Metropolis criterion, the temperature T(t) from the temperature control module is combined with the random number R(t) from the disturbance source. The random number is used to determine the updated spin state based on the current spin state. Then, based on the energy change of the spin state (increase or decrease), it is decided whether to accept the current proposed spin state. If the energy decreases, the spin data is unconditionally accepted, that is, the spin data is determined to be updated; if the energy increases, it is accepted with a certain probability. The probability function can be expressed as: Among them, P is the update probability, ΔH is the energy change amplitude, and T is the system temperature value.

[0080] Specifically, each spin state is judged and updated separately based on the current system temperature and random perturbation signal. It is understandable that if the current update operation causes the system energy to decrease, the new state is accepted unconditionally; if it causes the energy to increase, the transition is accepted with a certain probability. This probability function is determined by the current temperature and the energy difference (equivalent to the amplitude of the energy change), thereby realizing a controlled spin jump mechanism. The state update processing unit can be integrated with the annealing control circuit in hardware and executed synchronously through sequential logic, effectively realizing a simulated annealing-like spin control process, improving the overall system's optimization capability and the breadth of the solution space search.

[0081] Furthermore, the closed-loop regulation calculation system is further used to initialize and generate multiple copies, and run each copy at multiple temperatures;

[0082] The energy detection module is used to detect the total energy value corresponding to each replica and send it to the annealing control module;

[0083] The annealing control module is used to generate an updated energy value based on the total energy value and the random perturbation number and determine the energy change amplitude of each replica. When the energy change amplitude is lower than 0, the updated energy value is determined as the new total energy value; when the energy change amplitude is greater than or equal to 0, whether to determine the updated energy value as the new total energy value is determined based on the update probability.

[0084] In order to achieve efficient parallel and dynamic optimization of the system convergence process in the embodiment of the present application, a closed-loop tempering control mechanism based on energy feedback is further constructed. Figure 9 As shown, the system will first initialize and generate multiple copies (such as s1, s2, ..., s n ) and initialize multiple temperatures (e.g. t1, t2, ..., t). Each replica corresponds to a replica state, and different replicas operate at different temperatures. In each replica (replica 1-replica n), the total energy E in the current spin state is detected in real time by the energy detection module and its change trend is evaluated, and closed-loop annealing is performed separately. Based on the system stability and energy changes, the annealing control module adjusts the system temperature to construct a control mechanism that simulates the physical annealing process. Specifically, each replica first obtains the replica s and temperature t, and then generates a new solution s based on random perturbations. new , and compare the difference △E between the current energy and the previous state energy (where △E=E new -E old , E new is the energy value corresponding to the new solution, E old is the energy value corresponding to the solution before the update), when △E is greater than or equal to 0, the new solution is accepted according to the Metropolis criterion probability, otherwise, the new solution s is directly accepted. new .

[0085] Furthermore, in a feasible embodiment, the annealing control module is also used to determine whether the number of iterative updates of the current energy value is greater than or equal to a first preset number after determining a new total energy value or determining not to update the energy value; if not, determine whether the energy change amplitude is greater than a preset change threshold; if greater, adjust the temperature value of the replica based on the first cooling rate; if not, adjust the temperature value of the replica based on the second cooling rate; the first cooling rate is greater than the second cooling rate, and return to execute the step of generating an updated energy value according to the total energy value and the random disturbance number and determining the energy change amplitude of each replica; if not, output the current replica state.

[0086] like Figure 9 As shown, after accepting the new solution probabilistically or directly according to the Metropolis criterion, it is determined whether the current number of iterative updates of the energy value is greater than or equal to m (i.e., the first preset number). If not, the iterative update continues. If so, the current replica state is output. Specifically, when the current number of iterative updates of the energy value is not greater than or equal to m, it is determined whether the current energy change is large based on the energy change amplitude and the preset change threshold. If the energy change amplitude is greater than the preset change threshold, it indicates that the current energy change is large, that is, the current replica state is far from the optimal solution. In this case, the annealing speed is accelerated by rapid cooling (corresponding to the first cooling rate). Otherwise, the optimal solution is refined by slow cooling (corresponding to the second cooling rate).

[0087] Furthermore, in a feasible embodiment, the annealing control module is also used to exchange the states corresponding to each of the replicas according to the update probability after the number of iterative updates of the energy value is greater than a first preset number, and then determine whether the current total number of exchanges is greater than or equal to a second preset number; if not, return to execute the step of generating an updated energy value according to the total energy value and the random perturbation number and determining the energy change amplitude of each replica; if not, output the spin data of the optimal replica among all the replicas.

[0088] like Figure 9As shown, after each replica outputs the current replica state, the annealing control module can also exchange the replica states according to the update probability determined based on the Metropolis-Hastings criterion at a certain period. After each exchange of the replica states, it is determined whether the total number of exchanges is greater than or equal to the second preset number M. If not, the closed-loop annealing process of each replica is continued. If so, the state of the optimal replica among all replicas is output as the updated spin state. Among them, the optimal replica refers to the replica with the lowest energy value. The exchange of replica states is also determined based on the Metropolis-Hastings criterion. That is, when the replica energy value decreases, the replica state is directly exchanged, and when the replica energy value increases, the replica state is exchanged. Whether to exchange the replica state is determined based on the calculated update probability. Therefore, the purpose of exchanging the replica state is to minimize the replica energy value and ultimately find the optimal replica state as the updated spin state to maintain system stability.

[0089] The closed-loop regulation computing system proposed in the present application is equivalent to a time-frequency multiplication-accumulation circuit and its annealing control architecture for fully connected Ising calculations, featuring high parallelism, reconfigurable precision, and closed-loop optimization capabilities. By encoding coupling weights as frequencies, spin states as time pulses, and performing multiplication-accumulation operations in the phase domain, the efficiency and precision of multiplication-accumulation calculations in simulated physical systems are significantly improved, making it particularly suitable for large-scale parallel computing tasks. The introduced bit-weighted precision reconstruction mechanism effectively balances resource utilization and computational accuracy, providing excellent flexibility and scalability. In addition, the annealing control module constructs a complete closed-loop path for energy detection, temperature regulation, and state perturbation. It can dynamically control the temperature based on the system energy, achieving precise updates and jump control of the spin state, thereby significantly improving the overall system convergence speed and the global optimality of the solution. The overall architecture features low power consumption, compact structure, and controllable computational process, promising engineering implementation prospects and application value, making it particularly suitable for efficiently solving combinatorial optimization problems and hardware-software collaborative systems in physical simulation scenarios.

[0090] Furthermore, it should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the closed-loop regulation calculation system of the present application. More forms of simple transformations based on this technical concept are all within the scope of protection of the present application.

[0091] The embodiment of the present application also provides a closed-loop regulation computing device, which includes at least the closed-loop regulation computing system in the above embodiment. The closed-loop regulation computing device can be a device with computing capabilities, such as a computer, server, chip or other computing processing device. The closed-loop regulation computing device includes the closed-loop regulation computing system in the above embodiment, which can solve the technical problems of high complexity and poor accuracy in the implementation of the current fully connected Ising model. Compared with the prior art, the beneficial effects of the closed-loop regulation computing device provided by the embodiment of the present application are the same as the beneficial effects of the closed-loop regulation computing system provided by the above embodiment, and the other technical features in the closed-loop regulation computing device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0092] It should be understood that the various parts of the embodiments of the present application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any appropriate manner in any one or more embodiments or examples.

[0093] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the above claims.

[0094] The above is only an exemplary solution of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the patent processing scope of the present application.

Claims

1. A closed-loop regulation calculation system, characterized in that: The closed-loop regulation calculation system includes a digital time encoding module, a digitally controlled oscillation module, a phase accumulation module, an energy detection module, and an annealing control module. The digital time encoding module and the digitally controlled oscillation module are respectively connected to the phase accumulation module, the phase accumulation module is connected to the energy detection module, and the energy detection module is connected to the annealing control module. The digital time encoding module is used to encode the received spin data into time pulse data and send it to the phase accumulation module; The digitally controlled oscillation module is used to convert the received weight data into frequency signal data and send it to the phase accumulation module; The phase accumulation module is used to calculate the received time pulse data and frequency signal data to obtain a calculation result, and send the calculation result to the energy detection module; The energy detection module is used to evaluate the system energy value of the closed-loop regulation computing system according to the calculation result, and update the spin data through the annealing control module.

2. The closed-loop regulation calculation system according to claim 1, wherein: The phase accumulation module includes multiple multiplication and accumulation calculation units connected in series. During the process of calculating the time pulse data and the frequency signal data, the phase accumulation module determines a starting accumulation position from each of the multiplication and accumulation calculation units according to the bit weight level corresponding to the frequency signal data, so as to perform phase accumulation on the phase increment mapped by the time pulse data and the frequency signal data based on the starting accumulation position.

3. The closed-loop regulation calculation system according to claim 2, wherein: Each of the multiplication-accumulation calculation units in the phase accumulation module is activated in sequence based on a preset clock cycle, and during the calculation process of the phase accumulation module, a multiplication-accumulation calculation unit outputs a calculation result in each clock cycle.

4. The closed-loop regulation calculation system according to claim 1, wherein: The closed-loop regulation calculation system further includes a spin data storage array and a weight data storage array, wherein the spin data storage array is connected to the digital time encoding module and the annealing control module respectively, and the weight data storage array is connected to the digital control oscillation module; The annealing control module is configured to update the spin data in the spin data storage array according to the system energy value after receiving the system energy value sent by the energy detection module.

5. The closed-loop regulation calculation system according to claim 4, characterized in that: The annealing control module at least includes a closed-loop temperature control unit, a random disturbance source connected to the closed-loop temperature control unit, and a state update processing unit; The closed-loop temperature control unit is used to receive the system energy value sent by the system energy detection unit, and determine the system control temperature value according to the system energy value; The random disturbance source is used to generate a disturbance random number, and the disturbance random number and the system control temperature value are input into the state update processing unit together; The state update processing unit is used to determine whether to update the spin data according to the system control temperature value, and determine the spin data update value according to the system control temperature value and the random disturbance number when determining to update the spin data.

6. The closed-loop regulation calculation system according to claim 5, characterized in that: The state update processing unit is also used to determine whether the energy decreases after the spin data is updated based on the system control temperature value. If so, determine to update the spin data; if not, calculate the update probability based on the energy change amplitude after the spin data is updated and the system control temperature value, and determine whether to update the spin data based on the update probability.

7. The closed-loop regulation calculation system according to claim 6, wherein: The closed-loop regulation computing system is further used to initialize and generate multiple copies, and run each of the copies at multiple temperatures respectively; The energy detection module is used to detect the total energy value corresponding to each of the replicas and send it to the annealing control module; The annealing control module is used to generate an updated energy value based on the total energy value and the random perturbation number and determine the energy change amplitude of each replica. When the energy change amplitude is lower than 0, the updated energy value is determined as the new total energy value; when the energy change amplitude is greater than or equal to 0, whether to determine the updated energy value as the new total energy value is determined according to the update probability.

8. The closed-loop regulation calculation system according to claim 7, wherein: The annealing control module is further configured to, after determining a new total energy value or determining not to update the energy value, determine whether the number of iterative updates of the current energy value is greater than or equal to a first preset number; if not, determine whether the energy change amplitude is greater than a preset change threshold; if so, adjust the temperature value of the replica based on a first cooling rate; if not, adjust the temperature value of the replica based on a second cooling rate, the first cooling rate being greater than the second cooling rate, and return to executing the steps of generating an updated energy value according to the total energy value and the random perturbation number and determining the energy change amplitude of each replica; If not, output the current copy status.

9. The closed-loop regulation calculation system according to any one of claims 8, wherein: The annealing control module is further configured to, after the number of iterative updates of the energy value exceeds a first preset number, exchange the states corresponding to the respective replicas according to the update probability, and then determine whether the current total number of exchanges is greater than or equal to a second preset number; if not, return to the step of generating an updated energy value according to the total energy value and the random perturbation number and determining the energy change amplitude of each replica; If not, the spin data of the best copy among the copies is output.

10. A closed-loop regulation computing device, characterized in that: The closed-loop regulation computing device comprises the closed-loop regulation computing system according to any one of claims 1 to 9.

Images