Wenquxing 22: A Highly Efficient Neuromorphic Accelerator by RISC-V Customized Instruction Extension for Spiking Neural Network (RV-SNN 1.0), Streamlined LIF Model and Binary Stochastic STDP

This paper proposes Wenquxing 22: a neuromorphic processor to efficiently compute SNN with RISC-V extension instructions. The main idea of Wenquxing 22 is to integrate the SNN computing unit into the pipeline of a generic in-order processor to achieve neuromorphic computing by customized RISC-V SNN instruction extensions 1.0 (RV-SNN 1.0). To integrate the Leaky Integrate-and-Fire (LIF) model into the in-order processor, we prune the complex traditional LIF model called the streamlined LIF model, and apply it to the pipeline; the binary stochastic Spike-timing-dependent-plasticity (STDP) with binary synaptic weights is also proposed to achieve the power efficiency of Wenquxing 22. The experimental results show that, on Xilinx Alveo U250, working in 300MHz with pure 1-bit 2-layer SNN, the effective peak power efficiency of Wenquxing 22 reaches 2.4 TSOPS/W (Tera Synaptic Operations per Second per Watt) and the area efficiency reaches 339.9 SOP/LUT (Synaptic Operations per LUT); the peak classification accuracy on MNIST is 95.75%. We also evaluate the power consumption of physical chip of Wenquxing 22 on Development Board, and the result is 2.48W. Wenquxing 22 outperforms than the existing open-source spiking systems. The Source code of Wenquxing 22 is https://gitee.com/OpenBPU/OpenBPU1