01PaperDraft · 52 min read
GPU Mechanism
Why the same training program produces different bytes on different GPUs — and how to make those bytes agree by controlling the arithmetic and kernel choices.
Read the paper03 / PublicationsFrom the research notebook
Questions. Observations. Possibilities.
Publications
A place for ideas to find their words.
And for words to lead somewhere new.
02PaperRevision 3 · 69 min read
The step oracle
A CPU predicts a GPU training step before the card runs, then checks the loss, every parameter, and every optimizer tensor — byte for byte.
Read the paper