[mantap]

GPU have high latency. For a chess engine like stockfish which is designed to search as many positions as possible, the latency of a GPU is a big problem.

Engines like LC0 that do use the GPU work by searching fewer positions but with a heavier eval function. This makes the latency less relevant because it is a smaller percentage of the GPU time.

[nl]

3D computer games are much more latency sensitive than chess, and work on GPUs very successfully.

This seems like a solvable problem.

[johbjo]

Board game AI works by searching through the state space, and evaluating each state with the neural network, then picking the move with best expectation.

So it needs to load the comparatively tiny game state (chess board) into the GPU for each evaluation. The more game states it can evaluate per move, the better it is. It can be in the order of millions.