Reproduce “HG-PIPE: Vision Transformer Acceleration with Hybrid-Grained Pipeline” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qingyu Guo, Jiayong Wan, Songqiang Xu, Meng Li, Yuan Wang
HG-PIPE is an FPGA design for Vision Transformer inference that combines fine-grained streaming with coarse-grained buffering for attention’s data dependencies. It uses low-bit arithmetic and lookup tables to reduce demands. On VCK190, it reports 7,118 images/s versus 2,529 for V100, while residual-path activation-buffer cost is reduced 83.3% versus PIPO; speed comparison uses A3W3 at 71.05% versus fp32 at 74.5%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.