Beating cuBLAS by 4.7%: NVFP4 Kernels Hand-Built on GB300, 100% Claude-Generated
pranjalssh · x · 2026-08-16
Two years ago, the author entered the GPU world with a blog on outperforming cuBLAS on the H100 — a post that unexpectedly sparked a mini wave of hand-written Blackwell kernels: others reached 99% of cuBLAS with MXFP8, beat it with BF16, and the Modular team did it too. Now he's back with a sequel: an NVFP4 GEMM built from scratch on a GB300 node (CUDA 13.1) in plain CUDA plus inline PTX, with no framework dependency.
- Beats cuBLAS by 4.7% on the classic 8192×8192×8192 shape, with results on several other shapes; all code is open-sourced on GitHub.
- NVFP4 was chosen because it's the hardest dtype: as tensor-core instructions get faster, the bottleneck shifts to feeding the pipeline, which is where the most extreme GPU performance tricks shine.
- This time the author has something better than Hilbert curves, an approach that exactly fits NVIDIA's hardware layout.
- The entire kernel is 100% generated by Claude (details to follow).
He hopes this motivates people to own their kernels as ordinary code rather than treating them as black boxes.
More from Infra
- CoreWeave's debt costs falling: DDTL spreads tightened from S+9.62% to S+2.25% — BenBajarin · 2026-10-16
- SoftBank lines up US IPOs for SB Energy and AI robotics spinoff Roze, valuations top $50B — nordicinst · 2026-10-16
- TRL reproduces Modal's mixed-precision async RL: plain GRPO collapses, SC+MIS holds 0.72 — QGallouedec · 2026-10-16
- Hugging Face scales Indic pretraining with 2.8T tokens and an open data pipeline — Hugging Face Blog · 2026-10-16
- NVIDIA Vera Rubin NVL72 hits up to 5.7x GB300 throughput per MW on agent benchmark — ArtificialAnlys · 2026-10-16
- 14 open-source LLM observability tools to know in 2026, from Langfuse to Laminar — TheTuringPost · 2026-10-16