Vibecoding an inference chip

(ankitg.me)

41 points | by ankitvgupta 1 hour ago

5 comments

  • cpldcpu 23 minutes 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.

  • chalkycrimp 32 minutes ago
    Good chance I'm not the only one thinking this: share the code
  • codekansas 44 minutes ago
    This is cool

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

  • datadrivenangel 1 hour 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.
    • risyachka 51 minutes ago
      Fpga is software-driven, model can verify its output easily.

      Zero surprise it did it

    • datadrivenangel 1 hour ago
      it only cost them $10,000 in Claude API usage. :)
      • hmate9 1 hour ago
        You get that much usage in a monthly $200 subscription.
        • dvh 23 minutes ago
          So who's paying the cost?
          • exe34 22 minutes ago
            Corporate customers
        • dofm 40 minutes ago
          And a kitten.
  • saejox 23 minutes 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.