HyperDMA: Enhancing High-Performance Computing and AI Workflows with Advanced Data Transfer Capabilities
Minghao Peng, Haiyan Chen, Yang Zhang, Sheng Liu · 2024
In the field of high-performance computing (HPC) and artificial intelligence (AI), the performance and flexibility of direct memory access (DMA) controllers are crucial. This paper proposes a novel DMA architecture: HyperDMA, which achieves efficient data transfer and processing through a hierar-chical structure and a multi-core cooperative mechanism. Within HyperDMA, we design a data processing module that supports N -dimensional tensor transfer, gather transfer of non-contiguous source data, as well as broadcast and segmented transfer in a multi-core context. Finally, HyperDMA is integrated into the V-DSP computational core of the multi-core accelerator Fusion, and RTL simulation is performed using a simulation platform to evaluate its performance under different transfer modes and data granularities. The results show that HyperDMA significantly outperforms traditional DMA in data transfer efficiency and bandwidth utilization. Compared with not using HyperDMA, we achieve at least 39% performance improvement of GEMM algorithm. Compared with not using HyperDMA, we achieve at least 39% performance improvement of GEMM algorithm. This research provides an efficient and reliable solution for data transfer in HPC and AI applications.