Abstract:This work presents a hybrid CMOS-RRAM integration of spiking nonvolatile computing-in-memory (nvCIM) processing engine (PE) that includes a 64Kb RRAM macro and a novel in situ nonlinear activation (ISNA) module. We integrate the computing controller and nonlinear activation function on-chip to compute convolutional or fully-connected neural network. ISNA merges A/D conversion and activation computation by leveraging its nonlinear working region. This eliminates the need for additional circuits to realize nonlinearity and reduces area by 43.7 x w.r.t. the ADC scheme. The activation precision of ISNA can be configured from 1 to 8 bits to balance throughput, accuracy and power efficiency. The measurement of 4-layer LeNet shows such optimization improves 23.1% of computing speed via compromising a 2.5% relative accuracy drop. The proposed nvCIM PE achieves 16.9 TOPS/W power efficiency and a maximum spike frequency of 99.24 MHz.

RRAM-based Spiking Nonvolatile Computing-In-Memory Processing Engine with Precision-Configurable in Situ Nonlinear Activation