–cuda-clock nvidia-smi Support for Titan XP etc

Developer Tools Other Tools System Management and Monitoring (NVML)
, ,
1

5m

For compute, Titan XP and other legacy cards are locked at pstate P2 based on this post: Titan V slower than 1080ti tensorflow:18.08-py3 and 396.54 drivers

Is it possible to get support for --cuda-clocks or related for these GPUs? We are trying to study the very tradeoffs that they discuss in this post - the clock-to-instability relationship for problems like deep learning, etc, so that we can advise on this stability issue.

2

You can alternatively search for NVreg_RegistryDwords for driver (over 700 of driver parameters). I resolved P-state problems with RMForcePstate (and use “-ac” or “-pl” setting for nvidia-smi).

3

This seems to not work for me. The default pstate gets set to P0, but still reverts to P2 during cuda applications. Even worse, the performance/power draw seems even more limited, down to ~100 watts and 1/4th of peak performance.