Polaris 23: a high throughput neuromorphic processing element by RISC-V customized instruction extension for spiking neural network (RV-SNN 2.0) and SIMD-style implementation of LIF model with backpropagation STDP

With the rapid evolution of neuromorphic computing, particularly in the realm of spike neural networks, the need for high-performance neuromorphic chips has escalated significantly. These chips must exhibit exceptional data throughput, necessitating both robust computing capabilities and neuronal transmission bandwidth. Addressing this imperative, our research presents a neuromorphic processing unit (NPU) that boasts both high data throughput and a customized spike neural network instruction set with backpropagation acceleration functionality. The cornerstone of this NPU is the Polaris 23 Processing Element (PE), which leverages a multi-issue super-scalar architecture to enhance instruction parallelism and mitigate the average latency of high-delay instructions. Furthermore, to ensure high-bandwidth neuronal and synaptic state transmission, Polaris 23 incorporates multi-bank caches utilizing SRAM arrays and facilitates efficient data access. Rigorous hardware and software testing have been conducted on Polaris 23. The results are compelling, demonstrating that, when compared to the PE of SpiNNaker 2, a leading neuromorphic chip, Polaris 23 doubles the neuronal transmission throughput, achieving a remarkable 16GBps/GHz. Additionally, it surpasses SpiNNaker 2 in neuron precision, maintaining the same neuronal computing efficiency. Notably, the MNIST model implemented on the Polaris 23 platform achieves an impressive accuracy of 91%.