quda icon indicating copy to clipboard operation
quda copied to clipboard

Published 20 hours ago • verified publisher icon lattice

Update `cudaMemPrefetchAsync` to use `cudaMemPrefetchAsync_v2`

Open weinbe2 opened this issue 7 months ago • 1 comments

cudaMemPrefetchAsync currently accepts either a GPU device ID for moving data to devices, or -1 to prefetch data to the host. The newer routine cudaMemPrefetchAsync_v2 gives more granularity over migration to the host by providing the option to specify a host NUMA memory space. This will allow for more granular control on multi-socket nodes such as the 4xGH nodes on Jupiter and ALPS.

weinbe2 avatar Apr 10 '25 19:04 weinbe2