Jacques Mattheij

Technology, Coding and Business

Another Way Of Looking At Lee Sedol vs AlphaGo

Imagine a contest in which we are going to pit man against machine. But instead of measuring who is best in playing the game of ‘Go’ we are going to measure who is fastest. In the one corner: Human, all of 175 pounds of extremely well trained runner. And in the other corner, a Formula 1 racecar with a remote control running down a straight track.

Nonsense, you’d say, that’s not a fair comparison, the car uses much more instantaneous power, and will use a much larger amount of energy during the race, you should handicap that car somehow to make the race a fair one. And I’d be more than happy to agree with you.

Now, in an absolute sense, yes, that racecar is faster than a human. But in a relative sense, the amount of energy it will take the human to get to his destination for an ‘unaided’ human (or even one on a bicycle, probably the easiest way in which a human aided by some non-powered mechanics can achieve substantial speed) is so much lower that in a way even if the human loses on the absolute scale he or she actually wins that race for most practical purposes because the other approach does not scale beyond the race.

In the world of Chess it took until 1996 before a computer won even a single game against the then reigning world champion, Garry Kasparov in a series of 6 matches, and another year before he was defeated in match play with an updated version of Deep Blue. It took 30 computers to achieve this feat, evaluating approximately 200 million positions per second, mostly due to its use of a large number of specially crafted chips. The 30 P2SC nodes alone consumed about 900 watts of power, you still need to add some power consumption for the 480 ASICs that were especially crafted to play chess (we’re not actually talking about a general purpose computer here, more about a special purposes hardware design that happens to be extremely good at evaluating chess positions).

After that it took until 2009 for something (really) interesting to happen, a program called ‘pocket Fritz’ achieved parity with Deep Blue in terms of strength while running on a mobile phone and evaluating only a small fraction of the number of positions evaluated by Deep Blue. In terms of power efficiency that’s an absolutely enormous leap and it brought - in my view at least - for the first time the game on a level footing. Humans and Chess computers were now competing at approximately the same level of energy consumption.

This short detour into the world of Chess is the lead in to what I would like to say about the Lee Sedol vs AlphaGo games.

Now, not to diminish the achievement of the AlphaGo team, what they have done is nothing short of incredible, something experts predicted a very few short years ago would not happen in a decade is a fact today, AlphaGo beat Lee Sedol 4:1 in a match that will go into history as the one where computers beat humans on the most sophisticated game we have constructed. Go is an extremely complex game and to achieve this result took a large team of people and a very large amount of computer hardware in order to win and as impressive as that all is I don’t think it is any more a ‘fair’ match than that Formula 1 car against the runner (or even against a cyclist).

Lee Sedol used about 20 Watts of power to operate. By contrast, AlphaGo runs on a whopping 1920 CPUs and another 280 GPUs for an estimated power consumption of approximately 1 MW (200 W per CPU and 200 W per GPU). That’s 50,000 times as much power as the amount of power that Lee Sedol’s brain uses and the two are not quite evenly matched but it is close enough to use for comparison.

So now the interesting question (to me at least) is: How long before a computer will beat the human Go world champion using no more power than the human. That will be an achievement that is much more important than the one achieved last week, it will on a Joule-for-Joule basis be a fair match, when this one - to me at least - definitely was not, and it will require a very large improvement in our understanding of how humans play Go. Anybody taking bets on how long it will take to achieve that feat?