lichess.org

AlphaZero is a hoax.... (READ)

Current SF8 running on multiple processors, (not a standard lap top as used in the match vs. the supercomputer, evaluates SF's performance as 2200/2300 with many mistakes every game.)
The playing field was not level. The exhibition was only to display Google's advances in AI (self learning Monte Carlo logarithms) and contributes very little to chess theory.
What is proven, is a super-computer can self learn far better than human can program board games.

@mdinnerspace According to the paper, SF was running not on a standard laptop but on a 64 thread beast getting 70 mnps.

I have a fairly powerful computer of my own which gets 35 mnps, and I have thus far been able to replicate all of the "suspicious" SF8 moves from the match that people with weaker hardware have questioned.

Also AlphaZero didn't run on a supercomputer in the match. Only the training was done on what qualifies as a supercomputer (5000 first-generation TPUs to generate games, 64 second-generation TPUs to train the neural networks; 4 second-generation TPUs to play against Stockfish).

If anything it was AlphaZero that ran on the (really not at all like an actual laptop) equivalent of a laptop compared to what they could have used! (More like Deep Blue running on a fraction of its hardware)



From the original paper hardware used for initialising and training -

Training proceeded for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters, using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation TPUs to train the neural networks

and hardware used for the games - AlphaZero and the previous AlphaGo Zero used a single machine with 4 TPUs Stockfish and Elmo played at their strongest skill level using 64 threads and a hash size of 1GB.

So, AlphaZero used special hardware developed by Google. It used specialized Tensor Processor Units (TPUs) rather than general Central Processing Units (CPUs) as are available commercially.

This is how Wikipedia describes the second generation TPUs they used -

The second generation TPU was announced in May 2017. Google stated the first generation TPU design was memory bandwidth limited, and using 16 GB of High Bandwidth Memory in the second generation design increased bandwidth to 600 GB/s and performance to 45 TFLOPS. The TPUs are then arranged into 4-chip 180 TFLOPS modules

They used 4 TPUs for the games, so a processing power of 180 TFLOPS. Note TFLOPS = 1000 billion floating point operations per second.

For comparison Intel's latest most powerful chip is the Core i9 Extreme Edition processor which clocks in at 1 TFLOP. A top of the line I7 that you would find in a gaming machine would typically be about 100 GFLOPs (i.e. one tenth of a TFLOP).

I think it's fair to say that AlphaZero was using an 800 pound gorilla of a hardware configuration compared to Stockfishes mouse.

Copied

chess.stackexchange.com/questions/19366/hardware-used-in-alphazero-vs-stockfish-match

I'm not surprised at all that SF at a much lower depth (SF on lichess) thinks that some of the moves of SF at a much higher depth (Google's beast) are mistakes :)

You are right that A0 ran on a more powerful machine than SF did. There is no doubt about that. I'm not trying to say otherwise. What I am trying to do is combat two incorrect rumors that I keep hearing:

1) SF ran on a cheap tablet/laptop/whatever

2) Their SF is playing weaker moves than my SF/lichess's SF/chess.com's SF - they must have set it up wrong for it to keep blundering like that!

These two things are false.

They are both glorified calculators.

Saying that these calculators play chess is the same as saying that a bulldozer is the world Sumo wrestling champion, because under Sumo rules vs a human it wins every time.

But in chess people seriously compare a machine to a human and even say that the machine is better than the human in a human sport.

It's very funny that the general opinion is that chess is a game for the smart people when the chess players have a huge portion of people who seriously would crown the bulldozer as the Sumo-champion.

Maybe this silly idea started from the fact that we remember a time when humans defeated the calculators under chess rules. Well, the Sumo wrestlers also defeated the shovel under Sumo rules.

Get over it.

Whatever happened we are missing the point a little bit, it doesn't really matter if AlphaZero destroys every computer in chess, chess itself is not an end for these guys in google, the important thing is the way these guys basically are crushing computer science.

@butterfli

"The games were played at a fixed time of 1 minute/move, which means that Stockfish has no use of its time management heuristics (lot of effort has been put into making Stockfish identify critical points in the game and decide when to spend some extra time on a move; at a fixed time per move, the strength will suffer significantly)."

Post #10

2nd point: opening/end game table bases were disabled for SF.

Monte Carlo System - repeat randomness trillians of times - survival of the fittest - beat stockfish - 42

Reconnecting