Research On Optoelectronic Heterogeneous Computing Method For Large-Scale Matrix-Vector Multiplication

With the advent of the post-Moore era,on the one hand,Artificial Intelligence applications such as deep learning,autonomous driving are booming,which leads to the demand for computing power and memory.On the other hand,Moore’s Law is gradually facing failure,and traditional Von Neumann architecture computers have been unable to cope with the increasing demand for high-performance computing.There is an urgent need for new computing architectures to solve computing power needs.Since matrix computing occupies most of the computational overhead of AI algorithm.Optical computing,due to its high parallelism,low energy consumption,and strong reconfigurability,can accelerate certain types of computations in the optical domain,especially matrix multiplication.However,pure photonic computing systems lack of effective optical buffering,difficulty in large-scale integration,and lack of flexibility to face the diverse performance requirements of different applications.In this regard,this thesis proposed an optoelectronic heterogeneous computing architecture.The main research contents are as follows:Firstly,this thesis designed and implemented an optoelectronic heterogeneous computing architecture based on an array of optical modulators.The optical modulators are cascaded to realize multiplication,and the outputs are combined to realize addition to construct an optical computing unit.And based on the FPGA design of the electrical control unit,the control methods of two performance indicators are discussed in detail:the block algorithm is designed on the FPGA to realize the dimensional control,and the error of the computing system is analyzed,and a scheme with adjustable precision is proposed.Finally,an optoelectronic heterogeneous computing platform is built,and it is verified that the platform supports matrix-vector multiplication of multiple dimensions and multiple precisions.Secondly,this thesis designed and implemented an optoelectronic heterogeneous computing architecture based on scattering medium.Firstly,the principle of optical computing is analyzed and explained,and the scheme of matrix-vector multiplication is determined.Then in order to realize photoelectric cooperative computing,FPGA is used to develop USB interface,design storage unit,and control DMD.Finally,a method of dividing macro-pixel blocks in the electrical domain to achieve precision control is proposed,and an experimental platform is built in combination with the optical computing unit.Due to the limited area of the DMD pixel array,when computing large-scale matrix-vector multiplication with large macro-pixel blocks to represent elements,the required computing time is longer.The experiment shows that the computing precision and speed can be controlled.