ThunderKittens Runs on NVIDIA Vera Rubin: NVFP4 GEMM Hits 22 PFLOPS
NVIDIAAI · x · 2026-09-15
Together Compute announced ThunderKittens is now running on NVIDIA Vera Rubin. Their kernel team got early access to NVL72, dug through the new ISA, and after reworking kernels for Rubin pushed NVFP4 and FP8 GEMMs to 22 and 12 PFLOPS respectively—competitive with cuBLAS + CUTLASS DSL. NVIDIA AI's official account replied with approval.
More from Infra
- Expert-lookahead delivers 10%+ speedup for MoE inference on low-memory Macs — carloslfu · 2026-09-15
- Local voice assistant keeps its memory across model swaps, thanks to open-source Mnemosyne — Enough_Leopard3524 · 2026-09-15
- Hugging Face reportedly acquired by NVIDIA — Turn_Trout · 2026-09-15
- NextDC accused of using AI to draft lobbying letters, with AI errors, for 225MW datacenter expansion — nordicinst · 2026-09-15
- vLLM-Omni: serving voice, video, and diffusion models explained by Red Hat AI — vllm_project · 2026-09-15
- Running Z-Image Turbo locally on an RX 6800: full ROCm setup, 43s per image — AstroFieldsGlowing · 2026-09-15