Decoupling Compute and Memory for Async GPUs

8 points | by yiyingzhang 2 days ago

4 comments

  • bobbyzhu2008 2 days ago
    67% less kernel code is the more interesting number here — Hopper's async capabilities have been underutilized largely because the programming model is painful. Curious how it handles cases where compute and memory phases aren't cleanly separable.
    • jhap 2 days ago
      This seems like a better version of CUDA, for Hopper GPUs?
      • preetham_rangu 1 day ago
        [dead]
        • jackofficial643 1 day ago
          [flagged]