I am using a P2200 GPU with the following software on Ubuntu 18.04:
Nvidia-smi Driver Version : 495.29.05
CUDA Version : 11.5
I am attempting to run a workload with specific core and memory clock frequencies, e.g. 139 MHz and 405 MHz respectively. I use ‘nvidia-smi -ac 405,139’ to set the application clocks and this completes successfully. I can see the settings under ‘Application Clocks’ in the output of ‘nvidia-smi -q -d CLOCK’.
However, when I run the workload, the memory clock goes up to 5005 MHz. I am expecting that the core and memory clocks stay at the specified application clock frequencies but that does not happen.
If the application clocks do not stay at the set frequencies, it is not possible to correlate the workload’s performance to the desired (set) frequencies.
Why does the memory clock reach 5005 MHz even if the application clocks have been set to 405 MHz for memory?
While running the workload after setting the application clocks at 278 MHz (core)/405 MHz (memory): the core frequency shoots up to 1746 MHz. In other words, how do we ensure that the core clock stays always at the specific application clock frequency?
a. The clock throttling reason shows ‘Active’ only for ‘Idle’. It is inactive for the reason ‘Application Clock Setting’. Is this related to the unwanted frequency boost issue? If so, how do we fix this?
Also, is it possible to have the core and memory clocks stay at the specified application clock frequencies even when the applications are not running?
Thanks in advance for your inputs.