US 12,487,626 B1
High precision analog optical computing method and system based on bit slice
Weiwen Zou, Shanghai (CN); Sicheng Yi, Shanghai (CN); Yuting Chen, Shanghai (CN); Shaoyang Zhang, Shanghai (CN); and Shaofu Xu, Shanghai (CN)
Assigned to Shanghai Jiao Tong University, Shanghai (CN)
Filed by Shanghai Jiao Tong University, Shanghai (CN)
Filed on May 7, 2025, as Appl. No. 19/201,747.
Claims priority of application No. 202510142467.8 (CN), filed on Feb. 10, 2025.
Int. Cl. G06E 3/00 (2006.01); G06E 1/04 (2006.01)
CPC G06E 3/005 (2013.01) [G06E 1/045 (2013.01)] 7 Claims
OG exemplary drawing
 
1. A high-precision simulated light computing method based on bit slices, comprising the following steps:
(1) respectively performing bit slice decomposition on an input signal and a weight signal according to predetermined precision requirements to generate a plurality of signal combinations with low bit precision, wherein each signal combination with low bit precision comprises a part of bits of the input signal or the weight signal,
wherein the bit slice decomposition is to respectively decompose the input signal and the weight signal into a plurality of binary bit fragments, and the number of bits of each binary bit fragment is determined according to required computing precision;
a high-precision input data stream with a word length of K bit is split into M low-precision input bit slices with word lengths of K/M bit by an input bit slice reconstruction module, wherein K is a positive integer multiple of M, and M is a positive integer greater than or equal to 2, wherein the input bit slice reconstruction module comprises one electrical input port, M optical output ports, and M electro-optical converters, wherein optical signals are converted from electrical signals of the M low-precision input bit slices by the electro-optical converters and are then outputted;
a high-precision weight with a word length of L bit is split into N low-precision weight bit slices with word lengths of L/N bit by a weight bit slice reconstruction module, and M low-precision weight bit slices are copied and output, L is a positive integer multiple of N, and N is a positive integer greater than or equal to 2;
(2) respectively performing independent computation on each signal combination with low bit precision generated in step (1) and the corresponding weight signal with low bit precision by using a simulated light computing technology to obtain a plurality of computation results with low precision; and
(3) accumulating and combining, according to a predetermined weight, the plurality of computation results with low precision obtained in step (2) to generate a matrix multiplication or convolutional computation result with high precision.