CN-121998011-A - Phase-change super-surface and holographic light field modulation-based storage and calculation integrated photon computing system

CN121998011ACN 121998011 ACN121998011 ACN 121998011ACN-121998011-A

Abstract

The invention discloses a storage and calculation integrated photon computing system and method based on phase change super-surface and holographic light field modulation, and provides a hybrid computing architecture combining a high-speed spatial light modulator and a photoelectric nonlinear activation circuit by using a blue-ray disc or a phase change storage wafer as a high-density holographic weight storage medium aiming at the problems of storage and calculation separation, low dynamic modulation speed and limited precision in the existing photon computing architecture. Through an area array parallel reading mechanism, a large-scale weight matrix is read at one time by using beam expansion coherent light, the multiplication operation of a light velocity matrix is completed, and the interlayer rapid switching of the deep neural network is realized by combining a flow rotation or light beam scanning technology. The invention realizes high-density on-chip storage and zero-power consumption lossless reading of large model weight by utilizing the nonvolatile and multi-stage refractive index characteristics of the phase change material, and effectively solves the problems of video memory wall and energy consumption wall in large model pushing.

Inventors

YU DONG
XU HUINAN
OuYang Shuting

Assignees

杭州华望系统科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (8)

1. A storage and calculation integrated photon computing system based on phase change super surface and holographic light field modulation, comprising: the light source module is configured to generate a reading light beam with high coherence, and expand and collimate the light beam to form an area array parallel light covering a specific calculation area; the input modulation module is positioned at the downstream of the light path of the light source module and is configured to receive a digital input vector, load the vector into the light field distribution of the area array parallel light and generate a structural light field carrying input information; The holographic storage module is positioned at the downstream of the optical path of the input modulation module and comprises at least one holographic super-surface storage medium constructed based on a phase change material, wherein Fourier holograms or calculation holograms corresponding to a neural network weight matrix are recorded on the medium in advance; the photoelectric activation module is positioned at the downstream of the optical path of the holographic storage module, and is configured to receive the diffraction light field after operation, convert the diffraction light field into an electric signal, and execute activation function operation by utilizing the nonlinear volt-ampere characteristic of the analog circuit; and the control and synchronization module is configured to coordinate the data loading time sequence of the input modulation module with the physical position or the light beam scanning position of the holographic storage module so as to realize multilayer continuous reasoning of the neural network.
2. The integrated photon computing system based on phase change super surface and holographic optical field modulation as claimed in claim 1, wherein the storage medium in the holographic storage module is a blue-ray disc structure or a wafer structure based on chalcogenide phase change alloy; The medium modulates the amplitude or phase of incident light by utilizing the complex refractive index difference of the phase change material in crystalline and amorphous states; the hologram on the medium is generated by adopting a computer holographic algorithm, and is pre-solidified in the medium recording layer by a laser direct writing or photoetching process to form a static weight database.
3. The integrated photon computing system based on phase-change super-surface and holographic light field modulation as claimed in claim 1, wherein said system adopts an area array parallel reading mechanism; the spot diameter of the area array parallel light generated by the light source module covers a complete weight matrix area or a sub-block area thereof on the medium; when the holographic calculation module calculates, the medium is kept relatively static or in a streaming motion state, and the multiplication and addition operation in the whole area is finished at the moment of light flight by utilizing the parallel diffraction characteristic of light without point-by-point scanning and reading.
4. The integrated photonics computing system of claim 1, wherein the control and synchronization module comprises a photoelectric encoder or a position sensor: When the medium is in a rotary optical disc structure, the encoder monitors the rotation angle of the optical disc in real time; when the optical disk rotates to the hologram area corresponding to the specific layer of the neural network, the control and synchronization module triggers the input modulation module to refresh the corresponding input vector and synchronously triggers the photoelectric activation module to acquire signals, so that the flow pipeline operation of interlayer calculation is realized.
5. The storage and calculation integrated photon calculation method based on phase change super surface and holographic light field modulation is characterized by comprising the following steps: the method comprises the following steps of preprocessing, namely partitioning and holographic encoding a neural network weight matrix to be deployed to generate corresponding hologram data, and physically burning the hologram data into a preset track or area of a phase change super-surface storage medium; the input loading step is that an input vector of the current reasoning task is obtained, and the vector is mapped into the spatial light intensity distribution or the phase distribution of the incident light beam by using a spatial light modulator; Controlling the incident light beam to irradiate the corresponding weight area of the storage medium, and realizing optical convolution or matrix multiplication of an input vector and a weight matrix in the light propagation process by utilizing the diffraction characteristic of the medium; a nonlinear activation step, namely receiving the calculated light field by using a photoelectric detector array, converting a light intensity signal into a photocurrent, and applying nonlinear transformation to the photocurrent through a diode or transistor circuit to obtain an activated output vector; and an interlayer iteration step, namely taking the output vector as an input vector of a next-layer network, driving a next-stage light source or modulator, and controlling a storage medium to be switched to a hologram area of the next-layer weight until the reasoning of the whole network is completed.
6. The integrated photon computing system based on phase-change super-surface and holographic light field modulation according to claim 5, wherein in the preprocessing step, a pure phase hologram or a complex amplitude hologram of a weight matrix is computed by using a Gerchberg-Saxton algorithm or a point source hologram algorithm; discretizing the calculated hologram into a binary or multi-value bitmap, and mapping the binary or multi-value bitmap into the crystallization degree distribution of the phase change material.
7. The integrated photon computing system based on phase change super surface and holographic optical field modulation as claimed in claim 5, wherein in the interlayer iteration step, the incidence angle or position of the readout beam is changed rapidly by using an acousto-optic deflector or a micro-electromechanical system galvanometer, microsecond switching of different weight areas is realized on a static storage medium wafer, and mechanical rotation seeking is replaced.
8. The integrated photon computing system based on phase change super surface and holographic optical field modulation as claimed in claim 5, wherein in the interlayer iteration step, the incidence angle or position of the readout beam is changed rapidly by using an acousto-optic deflector or a micro-electromechanical system galvanometer, microsecond switching of different weight areas is realized on a static storage medium wafer, and mechanical rotation seeking is replaced.

Description

Phase-change super-surface and holographic light field modulation-based storage and calculation integrated photon computing system Technical Field The invention relates to the technical fields of photon calculation, artificial intelligent hardware acceleration and optical storage, in particular to a storage and calculation integrated photon calculation system and method based on phase change super-surface and holographic light field modulation. Background With the explosive development of deep learning, particularly Large Language Models (LLMs), the need for artificial intelligence has grown exponentially. However, traditional electronic computer architectures (von neumann architectures) face severe "memory wall" and "energy consumption wall" challenges. When performing large model reasoning, the processor needs to frequently read massive weight data from the memory, resulting in much higher power consumption for data handling than for the computation itself, and bandwidth limits the reasoning speed. Photon calculation is taken as an emerging calculation model, utilizes the volatility (interference and diffraction) of photons to perform operation, and has the natural advantages of high bandwidth, low delay and low power consumption. However, existing photonic computing schemes (such as photonic chips based on Mach-Zehnder interferometers) generally face the following problems: 1. The memory density is low, the memory density is limited by the size of the device, a single chip can integrate thousands of neurons only, an advanced generation model which needs billions of neurons is difficult to support, an external electric memory is still needed, and the problem of a memory wall cannot be really solved. 2. The modulation mechanism is limited, and the traditional thermo-optical or electro-optical modulator has large volume and high power consumption, and is difficult to realize large-scale integration. 3. The lack of nonlinearity-photons do not possess nonlinear interactions themselves, making it difficult to realize all-optical activation functions (e.g., reLU) in neural networks. Phase change materials such as GST are mature for their wide use in optical disc storage and Phase Change Memory (PCM). It has a significant refractive index difference in crystalline and amorphous states and is nonvolatile (data is not lost after power failure). However, the prior art is mainly used for binary data storage, and there are few reports of using it as a holographic optical element for large-scale matrix calculation. Therefore, there is a need for a storage and computing integrated architecture that combines high-density optical storage and high-speed optical computing to implement real-time reasoning of large models in a low-cost, low-power-consumption manner. Disclosure of Invention The invention aims to provide a storage and calculation integrated photon computing system and method based on phase-change super-surface and holographic light field modulation, and the core concept is that a mature blue-ray disc or a phase-change wafer is used as a holographic weight library, a weight matrix of a neural network is pre-frozen in a physical medium, and calculation is completed by utilizing light diffraction instant, so that weight carrying expenditure in a reasoning process is thoroughly eliminated. According to an embodiment of the invention, a storage and calculation integrated photon computing system based on phase change super-surface and holographic light field modulation comprises: the light source module is configured to generate a reading light beam with high coherence, and expand and collimate the light beam to form an area array parallel light covering a specific calculation area; the input modulation module is positioned at the downstream of the light path of the light source module and is configured to receive a digital input vector, load the vector into the light field distribution of the area array parallel light and generate a structural light field carrying input information; The holographic storage module is positioned at the downstream of the optical path of the input modulation module and comprises at least one holographic super-surface storage medium constructed based on a phase change material, wherein Fourier holograms or calculation holograms corresponding to a neural network weight matrix are recorded on the medium in advance; the photoelectric activation module is positioned at the downstream of the optical path of the holographic storage module, and is configured to receive the diffraction light field after operation, convert the diffraction light field into an electric signal, and execute activation function operation by utilizing the nonlinear volt-ampere characteristic of the analog circuit; and the control and synchronization module is configured to coordinate the data loading time sequence of the input modulation module with the physical position or the light beam scanning position of the holo