Vibecoding an Inference Chip

(ankitg.me)

26 points | by ankitvgupta 1 hour ago

2 comments

  • codekansas 11 minutes ago
    This is cool

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

  • datadrivenangel 33 minutes 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 17 minutes ago
      Fpga is software-driven, model can verify its output easily.

      Zero surprise it did it

    • datadrivenangel 33 minutes ago
      it only cost them $10,000 in Claude API usage. :)
      • hmate9 28 minutes ago
        You get that much usage in a monthly $200 subscription.
        • dofm 7 minutes ago
          And a kitten.