upvote
Vibecoding an inference chip

(ankitg.me)

I did something like this with Opus 4.5: https://github.com/cpldcpu/smollm.c

It seems implementing transformers is the new "hello world" of AI assisted coding.

reply
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.
reply
Fpga is software-driven, model can verify its output easily.

Zero surprise it did it

reply
it only cost them $10,000 in Claude API usage. :)
reply
You get that much usage in a monthly $200 subscription.
reply
So who's paying the cost?
reply
Retail and society as ANT shares get pulled into index and wealth funds
reply
Everyone else that doesn't use the entirety of the plan they pay for.
reply
API pricing vs. individual plan pricing is just price discrimination. They're both positive gross margin.
reply
People believe in some wild stuff.
reply
Corporate customers
reply
And a kitten.
reply
This is cool

Next step is tuning your post training pipeline to optimize your model for your chosen FPGA architecture...

reply
Good chance I'm not the only one thinking this: share the code
reply
I must say I do see one application for this, that you can't really match reasonably: 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.

(unfortunately for a decent model it will also be price competitive with a big nvidia card)

reply
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
Now do silicon. ;)
reply