CUDA GEMM Journey
- Year
- 2026
- Status
- in progress
- Built with
- CUDA, C++, Nsight Compute, cuBLAS, CUTLASS
- Repository
- Source (opens in a new tab)
Technical summary
A from-scratch CUDA single-precision matrix multiplication (SGEMM), optimised step by step and benchmarked against cuBLAS, PyTorch and CUTLASS. Each kernel generation starts from a profile in Nsight Compute, fixes the one bottleneck it finds, and is backed by measured timings, GFLOPS and hardware counters rather than estimates.