WebGPU in Chrome Canary now supports float16 for faster local LLM inference

  • I hope they gate webGPU behind a per domain allow/deny notification like they do with location and camera access.

    This whole thing is just begging to be exploited by cryptominers.

  • To test, run Chrome Canary with:

      --enable-dawn-features=allow_unsafe_apis,disable_robustness
    
    Then, open: https://webllm.mlc.ai/#chat-demo