Low GPU utilization and slow inference on Win11 compared to Linux

The GPU utilization is about only 50% when running inference on WIn11 (including WSL), and the inference time is 2X longer; whereas GPU can be fully utilized running the same model on Linux (more than 90%)

The environment:

  • OS: WIN11 and WSL
  • RTX 4060ti 16g as an eGPU
  • Cuda version:12.3

GPU utilization when running inferencing on WIN11 or WSL:

+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 545.33.01              Driver Version: 546.29       CUDA Version: 12.3     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  NVIDIA GeForce RTX 3050 ...    On  | 00000000:01:00.0  On |                  N/A |
| N/A   37C    P8               4W /  60W |   1413MiB /  4096MiB |     53%      Default |
|                                         |                      |                  N/A |
+-----------------------------------------+----------------------+----------------------+
|   1  NVIDIA GeForce RTX 4060 Ti     On  | 00000000:06:00.0 Off |                  N/A |
|  0%   47C    P2              57W / 165W |   5329MiB / 16380MiB |     48%      Default |
|                                         |                      |                  N/A |
+-----------------------------------------+----------------------+----------------------+

I tried this with a text-to-audio model bark and another LLaMA model. Any ideas about what might cause this?