I do PyTorch training with two independent processes on GPU 0 and GPU 1 on a Windows 11 machine with 2x3090. The scripts are limited to the corresponding GPUs using CUDA_VISIBLE_DEVICES env var. The monitor is attached to GPU 1. I am connected via RDP. When I disconnect from the machine, GPU 1 that has a monitor attached halves its performance in the training script.
In nvidia-smi -q I see “Active” next to the “Clocks Throttle Reasons” “Idle” subsection for GPU 1. It shows 100% GPU utilization, but seems limited to 150W.
In contrast the GPU 0 continues chugging at 350W, 99% utilization and “Clocks Throttle Reasons” “SW Power Cap”.
Attached the full output of nvidia-smi -q for GPU 1. WSL is not involved. Training script is Windows-native.
nvidia-smi.txt (6.6 KB)