Vibecoding an inference chip
58 points by ankitvgupta 3 hours ago | 19 comments
datadrivenangel 2 hours ago
Does not say what type of model or size they were serving, but one of the traces mentions Maple, which is a 20b model. Still, getting ~70+ tok/s is quite good.
replyrisyachka 2 hours ago
Fpga is software-driven, model can verify its output easily.
replyZero surprise it did it
datadrivenangel 2 hours ago
it only cost them $10,000 in Claude API usage. :)
replyhmate9 2 hours ago
You get that much usage in a monthly $200 subscription.
replydvh 2 hours ago
So who's paying the cost?
replystronglikedan 26 minutes ago
Everyone else that doesn't use the entirety of the plan they pay for.
replynyc_data_geek 22 minutes ago
Retail and society as ANT shares get pulled into index and wealth funds
replycodekansas 2 hours ago
This is cool
replyNext step is tuning your post training pipeline to optimize your model for your chosen FPGA architecture...
lxe 12 minutes ago
Repetitive "make it faster" prompting is essentially what all these "RSI" claims are
replyspwa4 26 minutes ago
I must say I do see one application for this, that you can't really match reasonably otherwise: you can bake weights into an FPGA and then run them for 5W at a speed that will be competitive with a big Nvidia card.
reply(unfortunately for a decent model it will also be price competitive with a big nvidia card)
saejox 2 hours ago
i wonder if we can emulate an FPGA on a rtx 3090? then test it on an AWS EC2 F2 few hours a day.
reply
It seems implementing transformers is the new "hello world" of AI assisted coding.