Was talking about this repo on discord, got this comment from lightvector:
lightvector — Today at 4:37 AM
If you want the strongest strength even on weak hardware, I suspect the latest 40-block net is better overall than any smaller net at this point, so long as you can get, say, around at least 100 or so playouts so as cover basic tactics and avoid low-visits weirdness, which can be doable. (My laptop, for example, gets 20 playouts per second on pure CPU when compiling with Eigen AVX2, which means 5 seconds is enough already)
[4:40 AM]
It will be a different balance of strength due to the much shallower search - could be a little worse at capturing races and certain kinds of highly tactical fights, could be a little more vulnerable to certain blind spots, but far better global judgment and shape intuition, including playing correctly many moves that the smaller net won't judge well even with large amounts of search. Overall, probably stronger.
[4:41 AM]
The smaller net will be a bit more robust though, because of the deeper search, probably will make fewer big blunders in the tail even though weaker overall, probably better for analysis and reporting variations for the user for reviewing games.(edited)
[4:42 AM]
So, depends on what you want.
Was talking about this repo on discord, got this comment from lightvector: