Simon Willison’s Weblog

Subscribe

GPUs Go Brrr (via) Fascinating, detailed low-level notes on how to get the most out of NVIDIA's H100 GPUs (currently selling for around $40,000 a piece) from the research team at Stanford who created FlashAttention, among other things.

The swizzled memory layouts are flat-out incorrectly documented, which took considerable time for us to figure out.