Anatomy of High-Performance GEMM with Online Fault Tolerance on GPUs

General Matrix Multiplication (GEMM) is a crucial algorithm for various applications such as machine learning and scientific computing, and an efficient GEMM implementation is essential for the performance of these systems. While researchers often strive for faster performance by using large compute platforms, the increased scale of these systems can raise concerns about hardware and software reliability. In this paper, we present a design for a high-performance GEMM with algorithm-based fault tolerance for use on GPUs. We describe fault-tolerant designs for GEMM at the thread, warp, and threadblock levels, and also provide a baseline GEMM implementation that is competitive with or faster than the state-of-the-art, proprietary cuBLAS GEMM. We present a kernel fusion strategy to overlap and mitigate the memory latency due to fault tolerance with the original GEMM computation. To support a wide range of input matrix shapes and reduce development costs, we present a template-based approach for automatic code generation for both fault-tolerant and non-fault-tolerant GEMM implementations. We evaluate our work on NVIDIA Tesla T4 and A100 server GPUs. Experimental results demonstrate that our baseline GEMM presents comparable or superior performance compared to the closed-source cuBLAS. The fault-tolerant GEMM incurs only a minimal overhead (8.89\% on average) compared to cuBLAS even with hundreds of errors injected per minute. For irregularly shaped inputs, the code generator-generated kernels show remarkable speedups of \(160\% \sim 183.5\%\) and \(148.55\% \sim 165.12\%\) for fault-tolerant and non-fault-tolerant GEMMs, outperforming cuBLAS by up to \(41.40\%\).