Skip to content

Comment on Groqchatparent

Comments

They are putting the whole LLM into SRAM across multiple computing chips, IIRC. That is a very expensive way to go about serving a model, but should give pretty great speed at low batch size.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.