I am running experiments where I want to evaluate the performance of applications with reduced memory clock rates. I am trying to use nvdia-smi to do this. Although I am able to reduce the SM clock to the desired level, the memory clock always appears to max out at 5005.
My CUDA driver is 9.0
(ada)% sudo nvidia-smi -ac 810,607
Applications clocks set to "(MEM 810, SM 607)" for GPU 00000000:84:00.0
All done.
(ada)% nvprof -u ms --system-profiling on ~/BabelStream/cuda-stream
BabelStream
Version: 3.3
Implementation: CUDA
Running kernels 100 times
Precision: double
Array size: 268.4 MB (=0.3 GB)
Total size: 805.3 MB (=0.8 GB)
==16946== NVPROF is profiling process 16946, command: BabelStream/cuda-stream
=16946== System profiling result:
Device "TITAN X (Pascal) (0)"
Count Avg Min Max
SM Clock (MHz) 21 562.43 139.00 607.00
Memory Clock (MHz) 21 4145.19 405.00 5005.00
Temperature (C) 42 27.29 26.00 28.00
Power (mW) 42 70258.55 15929.00 83478.00
Fan (%)
21 23.00 23.00 23.00