What about NVIDIA sleep mode and power consumption !

Hello everyone,

I need some ressources that illustrates the sleep mode that the GPU enter when no work is given, because I can see this with nvidia-smi but no documentation about this.

For example, for my tesla M2075, it passes from 75W to 28W (diff=47W). This GPU (fermi) is a bit old but in the fermi whitepaper I don’t find any information about this .

I can estimate that there is the impact of launching the nvidia-smi! Even though, the idle of this GPU is ~= 56W so (diff=18W).

Thank you,
Dorra

This may be of interest:

https://stackoverflow.com/questions/45360006/what-does-persistence-mode-actually-do-which-reduces-cuda-startup-time/45377054#45377054

It links to various official documentation items.

Thank you forthe link. It was very helpful to understand th device intialization and CUDA context cost.

Still, I have some questions:

Q1: When we do the cudaFree, why we don’t see a decrease in the power consumption ? Should we release the primary cuda context to see this drop in power consumption ?

Q2: As it was mentionned in the post, the GPU manages pstates: so it changes dynamically. Is this true for all GPU tesla (from fermi to volta) ??

Q3: Does nvidia provide an API to keep track of Perf/power states ?? or there is only nvidia-smi ??

Thank you in advance,
Dorra

Why would you expect to see a reduction in power consumption when you do a cudaFree?

If you do a free() operation in CPU code, does that mean the CPU is consuming less power?

Q2: Yes

Q3: Power states are not directly visible using the CUDA runtime API. You can use the NVML library (what nvidia-smi uses) and it may be possible to get some visibility into GPU power consumption using the profilers, and/or CUPTI.