Jan. 11(Advanced Photonics) — with the rapid development of artificial intelligence neural networks, large-scale matrix operation and frequent data iteration bring great pressure to traditional electronic processors. Through the collaborative integration of optical processing and electrical processing, optoelectronic hybrid computing exhibits remarkable computing performance. However, practical applications are limited by problems such as the separation of training and reasoning links, offline weight updating, etc. , this leads to the deterioration of information entropy and the decline of calculation accuracy, resulting in low inference accuracy.
A programmable optical processing unit (OPU) based on phase pixel array is proposed by the Institute of Semiconductor Research, Chinese Academy of Sciences, and the flexible programming of OPU is realized by combining Lyapunov stability theory. On this basis, the team constructed an end-to-end closed-loop optoelectronic hybrid computing architecture (ECA) . Through hardware-algorithm co-design, the whole process closed-loop optimization of training and reasoning is realized, which effectively compensates for the loss of information entropy, it breaks the strong coupling relationship between calculation accuracy and accuracy in optical computing.
The architecture realizes the joint optimization of optical and electrical parameters and the adaptive calculation accuracy compensation through the noise self-learning mechanism. The experimental results show that the inference accuracy of ECA on the MNIST handwritten digit recognition task (a classical task in the field of computer vision) reaches 90.8% when a 4-bit OPU is adopted, and the inference accuracy of ECA on the MNIST handwritten digit recognition task (a classical task in the field of computer vision) reaches 90.8% when a 4-bit OPU is adopted, it is close to the theoretical limit (90.9%) of the 8-bit traditional computing architecture (TCA) , which indicates that the optical computing system can still achieve high-precision reasoning under the condition of low hardware accuracy, and provides a new idea for the design of high-performance computing architectures.
The OPU supports 30.67 gbaud/s computing speed, achieving 981.3 gops of computing power and 3.97 TOPS/mm2 of computing density. Theoretical analysis shows that the structure can be further extended to a scale of 128 × 128, with a computing power of 1005TOPS, a computational density of 4.09TOPS/mm2, and an energy efficiency of 37.81FJ/MAC, it shows potential applications in the fields of microwave photonic signal processing, optical communication and neuromorphic artificial intelligence. (Qing Yun)