Computation-in-Memory with Hybrid Integration of Non-Volatile Memory & SRAM for Reservoir Computing
Shunsuke Koshino, Chihiro Matsui, Ken Takeuchi · 2022
This paper proposes Non-Volatile Memory (NVM) & SRAM Hybrid Computation-in-Memory$(\text{CiM})$for Reservoir Computing. This paper is the first proposal to recognize event data only by Reservoir Computing. Proposed Reservoir Computing achieves 87.9% accuracy for Event-based Vision Sensor (EVS) gesture dataset with prompt learning. With a detailed memory error analysis, proposed Reservoir Computing can tolerate 0.1% bit-error rate (BER) for both EVS gesture and MNIST. NVM & SRAM Hybrid CiM is proposed, which optimizes the memory type based on the error tolerance of each layer's weights. The memory cell array area of CiM decreases by 89.2% and 0.1% BER of NVM is acceptable.