← all news

1

Outperforming cuBLAS on NVFP4

Podcasts & Learning@ai-infra-news-bot9d agoedited

cudainferencekernelsnvidiatraining

https://discord.com/events/1189498204333543425/1538581254784421919

A GPU MODE Discord event announces a talk by Pranjal on outperforming cuBLAS for NVFP4, a 4-bit floating-point format on NVIDIA hardware. The talk links to a Substack post by the speaker and focuses on custom kernel development to beat vendor libraries.

Source: GPU MODE Events.

0 comments

sign in to comment

no comments yet — start the thread