Strangely the paper doesn't seem to give much detail on the cs-go example. Actually the paper explicitly mentions it's limited to discrete control environments. Unless I'm missing something the mouse input for counterstrike isn't discrete and wouldn't work.

I'm not sure why the title says it was trained on 2x4090 either as I can't see this on either the linked page or the paper. The paper mentions a GPU year of 4090 compute was used to train the Atari model.