To Infinity and Beyond: ThunderKittens Now on NVIDIA Vera Rubin NVL72!
- What changed
- ThunderKittens was ported to the NVIDIA Vera Rubin NVL72, rebuilding the NVFP4 GEMM to reach over 22 PFLOPS.
- Why you should care
- New hardware generations require custom instruction set architecture adjustments to fully saturate hardware rooflines.
- Your move
- Watch. Monitor hardware specific kernel adoption rates.
- Evidence
- ConfirmedTogether AI — Blog ↗
- What to watch next
- Independent benchmarks verifying ThunderKittens performance on production workloads.
- Event
- improvement
- Event date
- Sep 10, 2026
- Relevant to
- General AI readers