I have a server that runs very light ML inference workloads, and I’m trying to keep power consumption to a minimum. At completely idle, the GPU and memory clock comes down and power sits around 20W. At “almost” idle 0% GPU load, the GPU and memory clock are almost maxed out, and power sits around 43W. Does this sound right, or is there anything I can do to tune it to be more power efficient?
RTX 3060, Ubuntu 22.04, ESXi 7, driver 530.41.03 open kernel.
GIGABYTE RTX 3060 12GB GDDR6 GV-N3060GAMING OC-12GD Rev2.0
See graphs for further numbers. Thank you!
Hi there @marcus34 and welcome to the NVIDIA developer forums!
For P2 the numbers are rather normal in your case, especially since I assume you are running in a VM headless in compute mode? The only thing to do here if you want to avoid switching into P2 is to experiment with the different nvidia-smi settings of influencing display/persistence/compute mode. The list of possible settings is quite extensive, but not all are supported by consumer GeForce GPUs from OEMs like Gigabyte. Their VBIOS might actually cause a switch between modes based on heuristics that the driver cannot influence.
I hope that helps!
