Conversation

Notices

1. Embed this notice
   :blobcathug: (jain@blob.cat)'s status on Thursday, 06-Mar-2025 00:55:35 JST :blobcathug:
   Using DirectML to benchmark ML Productivity is a fail...
   Why does DirectML has performance issues on AMD Cards and why ROCm can match CUDA?
   So why not ROCm vs CUDA Benchmarks in the first place?
   • Embed this notice
     :blobcathug: (jain@blob.cat)'s status on Thursday, 06-Mar-2025 01:12:45 JST :blobcathug:

     in reply to

     • Iska
     @iska yes, ML stuff works best on linux anyway and ROCm (or AMD HIP SDK how they call it) for windows is slightly behind in Versions
   • Embed this notice
     Iska (iska@catposter.club)'s status on Thursday, 06-Mar-2025 01:12:46 JST Iska

     in reply to

     @Jain@blob.cat not native on windows or something, needs WSL

   • Embed this notice
     Swurl (swurl@swurl.land)'s status on Thursday, 06-Mar-2025 05:13:31 JST Swurl

     in reply to

     • Iska

     @Jain@blob.cat @iska@catposter.club I hate how ROCm requires PCIE Atomics and CUDA doesn't

     :blobcathug: likes this.