DNN performance optimization based on Gemmini neural network hardware accelerator

Gemmini is an open-source, full-stack DNN accelerator generator. For an DNN algorithm running on Gemmini, traditional optimizations focus on changing Gemmini's hardware configuration (e.g., changing the size of internal storage and the size of the systolic array). This work proposes an an optimization-test iteration route combining Spike simulation and FPGA simulation under the constraints of unacceptable running speed of verilator simulator and limited FPGA resources. According to this route, this work starts from two aspects of hardware optimization and algorithm optimization. The internal storage size and calculation scale of Gemmini are changed in hardware, and the convolutional layer of Resnet50 neural network algorithm is optimized in algorithm. The experimental results show that for Resnet50 DNN algorithm, Gemmini with basic configuration provides 2000x speedup compared with running on CPU. On the platform of Gemmini accelerator, the optimization of hardware configuration provides 1.59x speedup for Resnet50. On the basis of the hardware optimization, the algorithm optimization provides 1.012x speedup.