OpenAI baselines very slow

[MatPoliquin]

When using OpenAI's baselines:
https://github.com/openai/baselines

I get very slow training performance despite the GPU usage being close to 100%
I get around 30 fps of training performance but when using the CPU I get over 200 fps

python -m baselines.run --alg=ppo2 --network=cnn --env=PongNoFrameskip-v4 --num_timesteps=2e7

Same problem when I switch to the NVIDIA MX 150 dedicated GPU, I get 50 fps instead of 500 fps when using CUDA drivers

That said when I use the MLP model the FPS is much better (over 300 fps).

System Information

• Windows 11 Pro 21H2
• Python 3.7
• TensorFlow-DirectML 1.15.5 (latest)
• Intel 620 Integrated GPU (latest driver)
• • Intel i7-8550U