Post #2558214
2026-05-01 12:09 UTC
@kroekermartin.bsky.social ~20. But I have the same negative effect when I limit the number of threads, e.g. to 8.
That is: much of the time OpenBLAS threading appears to be at best useless, at worst actually negative.
Replies (2)
-
@itamarst@hachyderm.io 2026-05-01 12:09
@kroekermartin.bsky.social I have on my TODO list doing some profiling to see if I can figure out what is going on, and maybe file a bug if I can figure it out.
-
@kroekermartin.bsky.social@bsky.brid.gy 2026-05-01 20:56
Hmm that's weird, normally I'd expect a sweet spot somewhere, but not single-threading categorically better (unless your code is actually performance-limited by a call to a single-threaded, unoptimized LAPACK function - in particular anything that relies on ?LASR , like the ?SYEV functions)