The Efficiency of Convolution on Gemmini Deep Learning Hardware Accelerator

The successful use of deep learning (DL) algorithms in a variety of applications is conceptually based on convolutions. Though convolution is a simple operation, it suffers from severe performance degradation when implemented in software. Recently, with the advancement of CMOS technology, the convolution operation in DL algorithms has been accelerated by being delegated to specialized hardware platforms such as Field Programmable Gate Array (FPGA) devices. On hardware platforms, the convolution operation can be implemented on a synthesizable processor core or custom hardware accelerators based on systolic arrays (SA). Choosing an optimal hardware implementation should not be done analytically but instead employ the use of tools for fast and accurate estimation of metrics such as execution cycles, hardware resource utilization, and power consumption. This work evaluates the efficiency of implementing the convolution operation on various SA dimensions (8 × 8, 16 × 16, and 32 × 32) on the open-source Gemmini DL hardware accelerator with a comparison to the synthesizable RISC-V Rocket processor core. In terms of execution cycles the 8 × 8, 16 × 16, and 32 × 32 Gemmini configurations offer speedups of 323×, 249×, and 204× relative to the Rocket core. This work shows that, unlike the General Matrix to Matrix Multiplication (GEMM), the performance of the convolution operation degrades by an average factor of 2 when the Gemmini SA is doubled. In terms of hardware resource utilization on the Zynq Ultrascale+ ZCU104 FPGA evaluation board, the area and power consumption increased by 3.1× and 2.7× when the Gemmini SA dimension is doubled. Overall, the 8 × 8 Gemmini SA dimension recorded the highest performance-per-area metric making it the most efficient for a popular convolution configuration.