On Ubuntu 20.04.2 LTS
Profiling a CUDA application with ncu
using the CUDA driver 510.47.03 shows that the SM frequency has been capped at 765MHz. This has caused a performance regression for many CUDA applications due to upgrading the driver. Trying to fix this with:
sudo nvidia-smi --application-clocks=1215,1410
Does not make a difference in the SM frequency reported by ncu. However it does fix the performance regression that is otherwise observed when running latest driver. (Running without profiling)
The output of nvidia-smi -q -i 0
(after the above command) shows:
Applications Clocks
Graphics : 1410 MHz
Memory : 1215 MHz
Default Applications Clocks
Graphics : 765 MHz
Memory : 1215 MHz
If there is some rationale for the capped frequency value please let me know. Hopefully this behavior would be fixed and the default applications clocks for graphics should be 1410MHz.