Hypersort: High-performance Parallel Sorting on HBM-enabled FPGA
Soundarya Jayaraman, Bingyi Zhang, Viktor K. Prasanna · 2022
Accelerating sorting on FPGA has been extensively studied by leveraging the fine-grained data parallelism of FPGAs. However, with the optimized hardware pipelines, the performance of sorting algorithms is bounded by the off-chip memory band-width. The integration of high-bandwidth memory (HBM) on FPGAs offers significantly more off-chip memory bandwidth compared with traditional DDR memory, which enables new opportunities for accelerating sorting. In this paper, we develop Hypersort, a hardware accelerator to accelerate sorting on HBM-enabled FPGA. We use columnsort to merge HBM channels. To support the data communication pat-terns of Columnsort, we propose several optimizations to reduce external memory (HBM) traffic and hide data communication latency to further improve the overall throughput. We implement our accelerator on a state-of-the-art HBM-enabled FPGA. Ex-perimental results show that our implementation achieves overall sorting throughput of 34 GB/s, which is up to 14.8×, 4.73× and 2.18 ×faster than the state-of-the-art implementations on CPU, FPGA with external DDR and HBM-enabled FPGA, respectively. The proposed approach demonstrates higher efficiency for merging sorted arrays in HBM channels compared with the state-of-the-art implementation on HBM-enabled FPGA.