Reproduce “Efficient Expert-Parallel Communication on PCIe-Connected Consumer GPUs” 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
Jaehwan Lee, Sangmin Lee, Chaewon Kim, Junsik Shin, Jaejin Lee
ThunderEP addresses expert-parallel communication on no-peer-to-peer consumer GPUs by using host memory as a shared exchange point, DMA engines for bulk transfers, and per-sender completion signals. This replaces NCCL’s relayed ring and lets ready work proceed independently. On RTX 5090, Qwen3-30B-A3B achieved 1.42× average prefill speedup over vLLM across 256-to-4K prompts, in the tested six-GPU PCIe setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.