is it possible to increase lc0 calculate time even at 1 sec it still takes lc0 at least 5 sec to display a move using the RTX 2080 hardware.
I’m measuring about exactly that too – somewhere around 4 seconds of overhead for calculating the moves, and I’m not exactly sure what is happening in those four seconds.
What’s interesting is that the overhead drops significantly when running LcZero on CPU hardware. I’m seeing about 200-500ms of overhead when using the 20-core CPUs.
So, my guess is that the bulk of the overhead must be doing something to initialize the GPU. I’m not terribly familiar with LcZero’s implementation though so I don’t know this for sure.
Somebody a while ago I thought had a good suggestion about keeping the engine processes running and recycling them between requests. That might be a way to eliminate the overhead for each request, but I’d need to experiment with it a bit before rolling it out to production.
I’m going to be doing some work on the backend soon, so I’ll be investigating this to see if we can shorten that delay!