Design and implementation of high performance SIMT processor storage system for machine learning

Aiming at the problem of fast access of parallel data in multi-thread operation in self-developed high performance SIMT processor, a storage system suitable for SIMT architecture is designed, which mainly includes storage controller, data cache and instruction cache. This design uses Verilog HDL language which can be integrated to realize its hardware circuit, and meanwhile builds a verification platform based on FPGA to perform functional verification of the storage system. The integrated maximum clock frequency can reach 285MHz on Xilinx's FPGA chip xcvu440-flga-2892-2-e. The verification of all aspects shows that the designed storage system meets the system requirements.