I am running some tests regarding various iterative solvers’ performance using cuSPARSE and AMGX. For some reason, despite of (1) having persistent mode on and (2) preferable maximum perfomance on NVIDIA X Server Settings, nvidia-smi shows GPU1 (GPU0 is used to render Xorg enviroment) dropping to power state P2 while running the benchmarks (while GPU1 load is nearly 100%), thus skewing my benchmark results. It gives variable solving times between reboot (more than 50% difference).
Before I did (1) and (2) I used to see GPU1 falling back to P8 while running the benchmarks. What am I missing here? Is there a way through nvidia-smi to lock GPU1 to P0?