My model in C++ is slower than Python?

[Koruvika]

Model: FPN (Face Parsing)
Checkpoints: FPN + ResNet-50
OS: Windows 10
GPU: NVIDIA RTX 3050Ti 4GB

I converted it to ONNX format. When I timed it on Python using time.time(), it took approximately 17ms.
But when I timed it on C++ using std::chrono::high_resolution_clock, it took approximately 41ms.

I think the issue is from Windows. Does Windows limit ONNX model abilities?