Why Browser Chess Engines Beat Grandmasters — but Go Needed a Supercomputer

Try a small experiment on your phone. Open a game of Xiangqi in your browser and let the computer play at full strength: the opponent living inside that tiny tab is strong enough to crush any human grandmaster alive. Chess is the same story — Stockfish, running right in your browser on your phone's own chip, still outplays Magnus Carlsen at his all-time peak. But open Go and the story flips: to beat a top human Go player for the first time, in 2016 Google had to mobilise an enormous computing system. Same family of games, same "play against the computer" idea — so why is one side so cheap it fits inside a browser tab, while the other cost like a national project?

Xiangqi and chess: why computers play them "cheaply"

The first reason is size. A xiangqi board has 90 intersections; a chess board has 64 squares. On each turn, a player has on average about 38 legal moves in xiangqi and 35 in chess. The game trees are still astronomical — as far back as 1950, Claude Shannon estimated the number of possible chess games at around 10^120 (the figure now bears his name), and xiangqi, with its longer games, is estimated at roughly 10^150. No supercomputer could ever enumerate trees like those — and no engine needs to.

Because both games have two "weaknesses" that computers exploit beautifully:

• The tree can be pruned. Alpha-beta search lets the engine discard most clearly losing branches and dig deep only where the position is genuinely in doubt.
• Positions reduce to arithmetic. Pieces carry fairly clear relative values — chariots, cannons and queens are worth far more than pawns — so a position can be scored with a number: count the material, adjust for a few positional factors, and you have an evaluation good enough to decide with. Today that scoring is done by a small neural network called NNUE, small enough to run extremely fast on an ordinary CPU.

The historical landmark is 1997: IBM's Deep Blue, with 480 specialised chess chips examining some 200 million positions per second, beat world champion Garry Kasparov 3.5–2.5. It sounds monstrous, but that was 1997 hardware — the computing power that once filled two two-metre cabinets is now comfortably beaten by the chip in the phone in your hand. At full strength, Stockfish tops the CCRL engine rating list at roughly 3,640–3,650 Elo on the engine scale. Even compiled to WebAssembly and run single-threaded inside a browser tab — noticeably slower than a native build on a strong machine — it is still estimated at around 3,200, several hundred points above Carlsen's record 2882. Xiangqi followed exactly the same road: Fairy-Stockfish and Pikafish, the Stockfish-family engines specialised for xiangqi, both sit far beyond every human master, and their browser builds are already more than any human player will ever need.

In short: for xiangqi and chess, the problem was solved by algorithms long ago. The hardware only has to be "fast enough" — and "fast enough" gets cheaper every year.

Go: more expensive in every dimension

A 19×19 Go board has 361 points — four times the xiangqi board. Each move offers on average about 250 choices, and a game typically runs to around 150 moves: the game tree swells to roughly 10^360. The number of legal positions has been counted exactly, digit for digit, by John Tromp and Gunnar Farnebäck: about 2.08×10^170 — against a mere 10^80 or so atoms in the observable universe. But size is not even the biggest barrier.

The biggest barrier is that Go positions cannot be scored by counting material. A Go stone has no "price": whether it lives or dies, whether it is strong or weak, depends entirely on the surrounding shape and the state of the whole board. No formula of additions and subtractions can say who is ahead in a mid-game Go position. That is why, for decades after chess had fallen, the strongest Go programs hovered around high amateur dan level — they were missing the one thing that mattered: intuition about shape.

The breakthrough was to let the machine learn that intuition: a deep neural network trained on millions of self-play games, combined with Monte Carlo tree search (MCTS) to guide it. But intuition has a price in compute: every move needs tens of thousands of simulations, and every simulation means running the neural network again. Strength then scales almost directly with how many simulations you can afford — that is, with how much machine you have.

DeepMind's own numbers make this vivid. In the 2016 Nature paper, the single-machine version of AlphaGo (48 CPUs, 8 GPUs) rated 2,890 on their internal Elo scale; the distributed version, using 1,202 CPUs and 176 GPUs, rated 3,140 — and it was that distributed giant that beat Lee Sedol 4–1. AlphaGo Zero, in 2017, showed how better algorithms slash the bill: it taught itself from scratch in 40 days over 29 million self-play games, and needed only a single machine with 4 TPU chips to actually play. Today's KataGo is more efficient still — running on a strong consumer graphics card, given enough simulations per move, it plays far beyond the best human professionals.

And here is the honest part about ourselves: KataGo inside a browser is a KataGo starved of simulations. It gets only a few seconds per move on your own machine, with no powerful GPU behind it, so its simulation count is cut to a small fraction. That is why the highest level of Go on SynerWorlds plays at roughly amateur 3 dan — exactly what the game's own level label says: more than enough to trouble strong club players, but not the "full" KataGo on a GPU. Same brain; feed it drastically less compute and it plays drastically weaker. That is the nature of Go for computers.

The human side is the same: it demands a stronger brain

The delightful twist is that the computing bill applies not only to machines — it applies to the human brain sitting at the board, too.

Xiangqi asks the brain to read deeply down a corridor. Go asks the brain to hold the entire board at once, on every move — it demands a stronger brain.

A strong xiangqi master reads tactics to a frightening depth, but the decisive battle usually runs along familiar lines: the chariot files, the cannon lines, the palace. Every piece moves on fixed paths, relative values are fairly clear, and there is exactly one final target — checkmate the general. The brain gets to concentrate its firepower down a narrow corridor: narrow, but dug very deep.

Go offers the brain no such foothold. All 361 points are potential battlefields; a stone played in one corner can change the fate of the opposite corner. Every move demands two kinds of work at once: reading a local life-and-death fight exactly, variation by variation, and judging the whole board — where is the biggest point right now, which direction is growing, who is actually ahead when no piece has a fixed price to add up? In other words, Go demands far more calculation than xiangqi — from machines and from humans alike.

Which is why, if anyone should feel challenged by Go, it is precisely the strongest xiangqi brains. They already own the hardest thing to train — the ability to read deeply and calculate far — and Go merely asks them to spread that ability across all 361 points instead of one corridor. Take this as an invitation, not a slight: if you pride yourself on your reading strength at xiangqi, the 19×19 board is where that strength gets tested at a far larger scale.

Strength = algorithms × hardware

Look back over the whole story: chess fell in 1997 to an algorithmic idea from 1950 plus specialised hardware, and xiangqi followed the same path. Go held out for almost twenty more years, and fell only when deep learning met the massive compute of the 2010s. Every game has its own price of calculation — and as we have seen, that price applies to chips and to brains alike.

You can test both sides of that equation yourself on SynerWorlds: face Stockfish and Fairy-Stockfish in Chess and Xiangqi and feel how strong "cheap" can be, then step into Go, learn your way around the 9×9 and 13×13 boards, and challenge your brain on the full 19×19. The machines have paid their computing bill. The rest — taking your seat at the board — is up to you.