HN.zip

Vibecoding an inference chip

57 points by - 17 comments
[3 hidden]5 mins ago
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.

[3 hidden]5 mins 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.
[3 hidden]5 mins ago
Fpga is software-driven, model can verify its output easily.

Zero surprise it did it

[3 hidden]5 mins ago
it only cost them $10,000 in Claude API usage. :)
[3 hidden]5 mins ago
You get that much usage in a monthly $200 subscription.
[3 hidden]5 mins ago
So who's paying the cost?
[3 hidden]5 mins ago
Everyone else that doesn't use the entirety of the plan they pay for.
[3 hidden]5 mins ago
API pricing vs. individual plan pricing is just price discrimination. They're both positive gross margin.
[3 hidden]5 mins ago
People believe in some wild stuff.
[3 hidden]5 mins ago
Corporate customers
[3 hidden]5 mins ago
And a kitten.
[3 hidden]5 mins ago
This is cool

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

[3 hidden]5 mins ago
Good chance I'm not the only one thinking this: share the code
[3 hidden]5 mins ago
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)

[3 hidden]5 mins 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.
[3 hidden]5 mins ago
Now do silicon. ;)
[3 hidden]5 mins ago