Abstract: To address the current shortcomings of RISC-V-based neural network accelerators in accelerating matrix computations and nonlinear operations within Transformer-based models,a neural network acceleration architecture based on RISC-V instruction set extension,named Taurus,is proposed.This architecture introduces matrix instruction extensions tailored to the characte-ristics of Transformer models and employs a systolic array to perform matrix multiply-accumulate operations.To accelerate nonlinear computations,vector instruction extensions are added,along with the design of specialized vector units to efficiently compute operations such as LayerNorm and Softmax.To ensure balanced data supply,memory access instruction extensions are optimized to provide sufficient data throughput to the matrix and vector computation units.The instruction set extensions adopt a scalar register expansion approach,embedding operand data information directly into the registers.This increases the addressing space and reduces the number of instructions required for large-scale data computations.The Taurus neural network accelerator architecture is cycle-accurately simulated on the Gem5 platform.Compared with the open-source accelerator Gemmini,Taurus achieves an 80% improvement in systolic array utilization during general matrix multiplication.For inference tasks on ResNet50 and BERT models,Taurus delivers 1.3× and 31.3× speedups respectively over Gemmini.Compared with the baseline RISC-V,Taurus achieves 1 467× and 4 513× performance improvements respectively.

Key words: Neural networks, Matrix computation, Nonlinear computation, Instruction set extension