Remix.run Logo
▲ Vibecoding an Inference Chip(ankitg.me)
34 points by ankitvgupta 2 hours ago | 11 comments
▲cpldcpu 2 minutes ago | parent | next [-]

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.

▲saejox 2 minutes ago | parent | prev | next [-]

i wonder if we can emulate an FPGA on a rtx 3090? then test it on an AWS EC2 F2 few hours a day.

▲chalkycrimp 11 minutes ago | parent | prev | next [-]

Good chance I'm not the only one thinking this: share the code

▲codekansas 23 minutes ago | parent | prev | next [-]

This is cool

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

▲datadrivenangel an hour ago | parent | prev [-]

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.

▲risyachka 30 minutes ago | parent | next [-]

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

Zero surprise it did it

▲datadrivenangel an hour ago | parent | prev [-]

it only cost them $10,000 in Claude API usage. :)

▲hmate9 41 minutes ago | parent [-]

You get that much usage in a monthly $200 subscription.

▲dvh 2 minutes ago | parent | next [-]

So who's paying the cost?

▲exe34 a few seconds ago | parent [-]

Corporate customers

▲dofm 19 minutes ago | parent | prev [-]

And a kitten.