All credits to Artificial Analysis for doing this benchmarking but I am really apprehensive of the timing of this. And I am really not in favor of releasing half-baked results even if it comes with a disclaimer that "we'll be further updating these results as more optimizations go in".
This looks more like the continued attempt to assuage concerns about TPUs eating Nvidia's share - which is just panic fueled by "AI experts", the same experts who will give you a 1000-page cheat-sheet to use AI agents to make a 7-figure ARR business over the Thanksgiving weekend.
Either way - half-baked results with disclaimers are only useful when you know the audience is going to spend time reading and understanding the results. Not when there are AI doomers lurking everywhere.
And more importantly - there's certainly a step difference between the stack Google would internally use for their TPU-runs vs what is available to the community today. That's why the CUDA moat exists in the first place - the maturity of the CUDA software stack is a generation ahead of anything out there.
Bye and Happy Thanksgiving. Time to eat some potatoes.
Good model @xAI | prev. d-matrix, @Google.
I am Speculating. You Decode. Opinions my own
