ML Agents Camera Sensor Lag

[lime vigil]

I have a RTX 2070 ssuper and an AMD Ryzen 5 3600x. I want to train my Agent to run towards a red block. I Use a camera sensor to make my Agent see. However i have around 30 fps as soon as i start the learning process with environments at the same time. The Sensor resolution is 160 * 90. Can i do anything to decrease lag? I already use Cuda but my GPU has like 10% usage max and my CPU has 40% usage max. I dont get why it is lagging. Someone knows a solution?

[warped forum]

Convoliutions are really GPU heavy, they do better on TPUs. I suggest training using Colab's TPU instead.