Reproduce “Anatomy of High-Performance GEMM with Online Fault Tolerance on 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
Shixun Wu, Yujia Zhai, Jinyang Liu, Jiajun Huang, Zizhe Jian, Bryan M. Wong, Zizhong Chen
A General Matrix Multiplication (GEMM) implementation for GPUs integrates online Algorithm-Based Fault Tolerance (ABFT) through architecture-aware kernel fusion and automatic code generation, achieving on-the-fly error correction with an average overhead of 8.89% compared to cuBLAS while outperforming prior fault-tolerant methods by up to 39% on Tesla T4 GPUs.
Key insights
- Fusing ABFT operations at the threadblock level within the GEMM computation pipeline minimizes overhead by leveraging data locality and maximizing overlap between computation and fault tolerance tasks.
- A deep understanding of GPU architecture and meticulous, step-wise optimization of the baseline GEMM kernel is crucial for integrating fault tolerance with minimal performance impact.
- Automatic code generation is essential for making fault-tolerant, high-performance GEMM practical and broadly applicable across diverse and irregular matrix input shapes, addressing a key challenge in specialized kernel optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.