NVIDIA Watch posted an update
Jason Goodison, co-founder and CTO of General Compute, says Nvidia GPUs handle the prefill stage well but struggle with decode on very large AI models. His company uses Cerebras chips for that token-by-token generation, according to Yahoo Finance’s account of his comments.
Why it mattersGoodison also argues that customers may pay three to five times more for faster tokens. That is his commercial assessment, not a published benchmark or proof that Nvidia is losing sales: he says inference demand remains supply-constrained.
Discuss: If faster responses command a premium, should AI providers optimise for speed even when it costs more, or focus on making ordinary inference cheaper?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.