Decoupling Compute and Memory for Async GPUs

Posted by yiyingzhang on Thursday June 25, 2026 @01:08PM
from the self-post dept.

Cool open-source project that introduces a new programming model for decoupling compute and memory for NVIDIA GPUs that supports asynchronous memory operations (e.g., Hopper). 12% perf improvement over SOTA and 67% less kernel code.

Paper: "VDCores: Resource Decoupled Programming and Execution for Asynchronous GPU" arXiv:2605.03190

( 2 comments | 5 points | on HN )

Re:Decoupling Compute and Memory for Async GPUs by bobbyzhu2008 on Thursday June 25, 2026 @01:10PM (#48676402)

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.

[ Parent | on HN ]

Re:Decoupling Compute and Memory for Async GPUs by jhap on Thursday June 25, 2026 @03:30PM (#48678136)

This seems like a better version of CUDA, for Hopper GPUs?

[ Parent | on HN ]