Is a GPU useful for Xiangqi engines?
It depends on the engine's architecture. For engines like Pikafish, which use alpha-beta pruning + NNUE, only the CPU takes part in the calculation and the GPU is of no use. For engines like Px0 and Lc0, the GPU is very important.
A GPU cannot make Pikafish stronger. Even if your computer has a high-end graphics card such as an RTX 5090, it takes no part in the position search while Pikafish is running.
Pikafish uses a neural network, so why doesn't it use the GPU?
Because "neural network" does not mean "must use a GPU".
The NNUE used by Pikafish stands for "Efficiently Updatable Neural Network". Its network is small, and it exploits the fact that only a few pieces change between neighboring positions by incrementally updating intermediate results, instead of recomputing the whole network every time. Within an alpha-beta search, it is hard to gain any efficiency advantage from a GPU.