Why the Go Engine on SynerWorlds Isn't as Strong as the Xiangqi and Chess Engines

If you have played the level-12 engine at Xiangqi or Chess on SynerWorlds, you know the feeling: there is almost no way to win. Switch to Go, though, and an amateur around 3-dan can already beat the top level. We did not deliberately hold Go back — the difference lies in engine technology, and in the hardware the engine is allowed to use.

Two completely different engine schools

The Chess and Xiangqi engines, Stockfish and Fairy-Stockfish, search with alpha-beta pruning guided by a small neural network called NNUE. NNUE is designed to update incrementally from one move to the next, which makes it extremely cheap: an ordinary CPU evaluates millions of positions per second. Even compiled to WebAssembly inside a browser, these engines stay fast enough to be superhuman.

The Go engine, KataGo, belongs to a different school. It searches with Monte Carlo tree search (MCTS), and its "intuition" is a large neural network that must run for every node of the search: the network says which moves are worth trying and who is ahead. That intuition is powerful, but asking the network costs far more than one NNUE evaluation.

The GPU is KataGo's home turf

On a desktop install, KataGo draws its power from the graphics card (GPU) through its CUDA, TensorRT or OpenCL backends. A GPU evaluates thousands of positions in parallel batches, letting KataGo reach tens of thousands of visits per second with its largest networks. It is the GPU that makes full-strength KataGo stronger than the world's best professionals.

In a web browser: CPU only

A web page is not allowed to use the GPU that way. The KataGo build SynerWorlds ships runs entirely on the CPU (the Eigen backend, compiled to WebAssembly), using at most 8 CPU threads, with a compact network called b10 — and it steps down to an even smaller b6 network on weak devices.

The newer WebGPU technology does bring GPU computing to browsers, but KataGo has no official WebGPU backend yet, so that power cannot be brought to the web today. The result: the web build achieves only a small fraction of the visits per second of a machine with a discrete GPU.

Two more limits

First, a 19x19 board offers around 250 legal moves per turn, against roughly 35 in chess. The same search budget is spread far thinner, and Go games also run much longer.

Second, we cap thinking time per level so games stay responsive: at level 12 the engine gets at most 15 seconds per move. When time is up, it must move — no hardware, however strong, may think longer.

So how strong is level-12 Go?

Around amateur 3-dan. That is still more than enough to teach most beginners and casual players, but it is not the monster its two siblings are. We label our levels by their real strength, and this article exists so you can see exactly why the gap is there.

When WebGPU matures, or a GPU server becomes cost-effective, that gap can narrow. Until then, SynerWorlds' priorities stay the same: free, opens and plays instantly on any device — and honest about its limits.