2

If I'm correct (I may be not) there were no attempts to hardcode any strong chess engine directly into silicon and try to optimize it all the way to the transistor level. Is there any reason for that, other, than costs? For a long time there were many really strong CPUs (like Stockfish). Recently there were also some really strong TPU (Alpha Zero) and GPU (Leela Zero)-based engines, but as far as I'm aware up until now no one has tried to create a powerful ASIC chess engine.

I do know, that FPGAs are a bad choice for creating engines due to their low memory bandwidth as well as highly limited clock speeds.

However except for costs ASICs do look like a holy grail of chess engines. I'm not sure, why it's a bad idea (or why no one is doing that, if it isn't), however I do believe, that Stockfish rewritten from CPU code to an integrated circuit could possibly gain over 200 ELO.

Expense in the range of tens of thousands of dollars may be the reason, but Alpha Zero was significantly more expensive even than that.

3
  • 1
    Alpha Zero was a special project. There is no money in computer chess these days and all top engines are made by volunteers. I know noobpwnftw did donate some server hardware for stockfish.
    – qwr
    May 29 at 3:03
  • 3
    Many decades ago, specialized chess computers were a thing. However, nowadays hardware is so fast that it is not worth the effort. Who is going to buy a computing device that is only useful for chess and nothing else, when you can just buy a general purpose computer that is easily good enough at chess while also just being a normal computer?
    – koedem
    May 29 at 10:44
  • 1
    @koedem I think your comment should be upgraded to an answer, because that's what it really comes down to - the cost.
    – Ray
    Jun 2 at 14:09

2 Answers 2

3

Belle was a chess computer created by Joe Condon and Ken Thompson at bell labs. Move generation and position evaluation was implemented in hardware. It was the first computer to achieve master play with a USCF rating of 2250.

2

Expense in the range of tens of thousands of dollars may be the reason

The thing that you're missing is that this is a dramatic underestimate for anything that will be competitive. If you want to beat a 64 core CPU (which will be a lot cheaper) you will need some serious horsepower meaning you need a well designed chip on a relatively recent process. That's going to cost at least a few million dollars (and you will likely end up with a few thousand chips).

Making matters worse, the cost isn't really the same type as AlphaZero. Alphazero was research in machine learning and is essentially just a way to test out improvements in general algorithms for reinforcement learning. An asic by contrast will be a fully custom effort that won't be applicable to other games or more general research problems.

3
  • I think, that you're overestimating performance of CPUs as well. Usually ASICs im the term of computing power are at least 2 orders of magnitude faster, than CPUs, often as much as 4 orders of magnitude (here mining cryptocurrencies is a great example). It would cost at least a few milion USD to absolutely obliterate almost any resonably sized CPU engine and be able to compete againist resonably sized neural network, however to beat Stockfish on mid-range server CPU wouldn't be that difficult (or expensive). Jun 1 at 16:04
  • Usually ASICs im the term of computing power are at least 2 orders of magnitude faster This is only true because ASICs are only used where they are fast enough to be worth it. I'm pretty sure there isn't an ASIC that can evaluate a NNUE network more than 10x faster than a CPU (and even getting that 10x would be pretty difficult). Also mid range CPU is the wrong target because once you're talking ASIC prices, the competition is an equal priced CPU and dual top of the line 64 core CPUs are going to be way cheaper than an ASIC. Jun 1 at 17:55
  • I wouldn't even try to use ASIC for NNUE computation in the first place. I know that NNUE is basically matrix multiplication, so they're often bottlenecked by memory instead of computing power. Anyway, NNUEs are mostly integer multiplication, so I wouldn't expect speed-up to be any less, than 10 times. Anyway if I was tasked to build as strong a chess computer as I can I would probably decide to use ASICs to manage memory, generate positions and compare NNUE results to each other. For NNUE itself I would probably use reprogrammed SSDs, just like like Mystic AI did for image recognition. Jun 5 at 18:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.