Decoupling Compute and Memory for Async GPUs

4 comments

• bobbyzhu2008 1 day 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 1 day ago
  This seems like a better version of CUDA, for Hopper GPUs?
• preetham_rangu 1 day ago
  [dead]
• jackofficial643 13 hours ago
  [flagged]