I use this test run in my PC 2700x + 2060, it cost 10s rather than 3s, and I run this test in only cpu without cuda, it's perfomance only change a little. The cuda version test and cpu version test both use about 10s.

python3 test/discrete/test_pg.py --seed 0 --render 0.03

I am new to cuda and machine learning, but the system linux mint is fresh install, and the torch.cuda.is_available() is True