Blackwell Microbenchmarks A collection of microbenchmarks for NVIDIA Blackwell (SM 100) GPUs, covering memory throughput, latency, tensor core (UMMA) performance, and HBM-resident elementwise throughput. https://newsletter.semianalysis.com/p/dissecting-nvidia-blackwell-tensor Benchmarks Path Purpose ldgsts_throughput/ LDGSTS HBM throughput tma2d_throughput/ TMA 2D HBM throughput ldgsts_latency/ LDGSTS latency tma2d_latency/ TMA 2D latency umma_throughput/ UMMA tensor-core throughput umma_latency/ UMMA tensor-core latency elementwise_throughput/ fp32 HBM-resident activation/elementwise throughput Acknowledgements Compute for this project is generously sponsored by Nebius and Verda.
