That is probably true with 1060 being the borderline case. However, if you use higher NN cache values and start approaching late middle game IMHO impact of CPU could become noticeable.Laskos wrote: ↑Fri Aug 10, 2018 12:49 pm Price-wise Nvidia 1060 is a pretty fair comparison to Ryzen 7 1700. A bit lower price for GPU, but higher consumption power-wise. I still think that Lc0 on GTX 1060 performs somewhat better against regular engines on 8 core Ryzen 7 1700 than what TCEC shows. I think that two CPU cores allocation for Lc0 is a bit bogus, one can allocate 2 threads of the same core with very little loss, a large part of CPU cycles Lc0 needs being anyway garbage. In fact I might check the GPU load with affinites set to 2 threads of the same CPU core.
But that was not my point anyway. My point was that you need a CPU no matter how weak in order to run Lc0. GPU alone is not enough. And it is a flawed comparison to compare only price and power consumption of GPU vs. some CPU price/power for A/B engine. You have to take the whole package into comparison.
1060 6GB is around 300$ which is identical price to Ryzen 7 1700. You need almost another 100$ for CPU that runs that GPU even if you take the cheapest i3 or Ryzen 3. And it is gonna spend at least additional 30W of power to those 120W of 1060, while Ryzen 7 1700 is at 65W.
So 150W vs 65W and 100$ more whichever way you take it

And in that configuration strength-wise it is 400-500Elo (depending on TC) advantage of SFdev vs Lc0 with newest test net.
I don't think even die-hard NN fans believe 400-500Elo is something that is realistically achievable on the current 20x256 net...