The future of computer chess

[Uri Blass]

Which version of stockfish was the last version using only CPU.
And which version of Stockfish began suddenly using GPU as well ?!
Would you count the usage of gpu as a software or a hardware jump ?

Without gpu usage neural net development would not make such a big thing.
Its the enormous hardware power of the gpu that allows for using neural nets to the limit.

So the new hardware possibilities allowed a new approach. Instead of tuning evaluations by humans, neural nets replaced the hand made fine tuned evaluations.

And this Clou is with almost any hardware change. Hardware revolutions do not only give speed increase, they also allow different software ideas because suddenly the address space is bigger, the ram and rom is bigger,
Suddenly you can access Floppys instead of Tape, that allows you to make a bigger book that even learns or you can put endgame databases on floppy/cd/dvd or hdd or ssd.

Early dedicated chess computers were not only slower. They had limited ram/rom not allowing different techniques.
With the same speed but more ram/rom you could have used different ideas how to implement things.

Hash tables, bitboards and all those features came with the 16 bit cpus.
On the 8 bit cpus these new ideas couldn’t be realised.

Bigger opening books and/or growing/ learning opening books came with better media storage hardware.
On tapes this would not have been possible.

It needed floppy, hdd, cd, dvd.

Endgame tablebases came with cd/dvd hdd and later ssd.

Its not all about speed.
Its about ways to do it different.

I think that stockfish does not use gpu.

Leela with NN use GPU but stockfish found a cheap alternative of CPU because it use NNUE and Not NN like Leela.

[mclane]

Yes true. It uses nnue.
But these developments are side effects of the lc0 GPU usage. Before neural nets were not considered in computer chess.
It was all hand made.

[jkominek]

Which version of stockfish was the last version using only CPU.
And which version of Stockfish began suddenly using GPU as well ?!
Would you count the usage of gpu as a software or a hardware jump ?

Without gpu usage neural net development would not make such a big thing.
Its the enormous hardware power of the gpu that allows for using neural nets to the limit.

So the new hardware possibilities allowed a new approach. Instead of tuning evaluations by humans, neural nets replaced the hand made fine tuned evaluations.

And this Clou is with almost any hardware change. Hardware revolutions do not only give speed increase, they also allow different software ideas because suddenly the address space is bigger, the ram and rom is bigger,
Suddenly you can access Floppys instead of Tape, that allows you to make a bigger book that even learns or you can put endgame databases on floppy/cd/dvd or hdd or ssd.

Early dedicated chess computers were not only slower. They had limited ram/rom not allowing different techniques.
With the same speed but more ram/rom you could have used different ideas how to implement things.

Hash tables, bitboards and all those features came with the 16 bit cpus.
On the 8 bit cpus these new ideas couldn’t be realised.

Bigger opening books and/or growing/ learning opening books came with better media storage hardware.
On tapes this would not have been possible.

It needed floppy, hdd, cd, dvd.

Endgame tablebases came with cd/dvd hdd and later ssd.

Its not all about speed.
Its about ways to do it different.

Hi Thorsten,

You make many true points (except for the bit about Stockfish using GPUs). Are you raising these as a way of saying that my interpretation of the data I've presented a couple posts above is simplistic? In other words are you offering an invitation for me to weigh in on the main debate of this thread? My post was already long enough with an embedded graph and table to boot. I felt I had consumed my word budget for a single visit. But I could take the time to share my thoughts on the topic. I see merit to both your point of view as well of that of the opposite side, advocated most strongly by Ronald de Man (syzygy).

john

[mclane]

I don't see your point or my point in opposition.
You said speed. I said it's often not only speed but resources like the jump from tape to floppy, from floppy to CD/dvd and HDD to SSD.
Or from 8 bit to 16 bit giving bigger address space for different ideas.

[Hai]

The future of computer chess:

8-Piece Syzygy Endgame Tablebases
New Endgame Tablebases which need 1/10 space on ssd.

Engines and
Stockfish which runs on GPU only.
Stockfish which runs on NPU only.
Stockfish which runs on CPU + GPU.
Stockfish which runs on CPU + NPU.
Stockfish which runs on CPU + GPU + NPU.

Who wants to be famous should create an experimental engine.

[jkominek]

I don't see your point or my point in opposition.

It is nice that you think so. I was mostly interested in sharing measurements on the evolution of scaling curves and the Elo upper bound, rather than weighing in on the off-and-on running debate.

You said speed. I said it's often not only speed but resources like the jump from tape to floppy, from floppy to CD/dvd and HDD to SSD.
Or from 8 bit to 16 bit giving bigger address space for different ideas.

Yes, when considering the type and degree of computing resources that have enabled software advances, these contributing factors are well worth remembering. It is a causal loop. Unwinding the causal loop of what enabled or caused what and assigning attribution is not fully possible.

I'll add a current example. The Stockfish team makes use of what is effectively a supercomputer cluster in building their software, which is primarily targeted to desktops, laptops, and smartphones. The stockfish we know and enjoy today would not exist if the authors were constrained to develop on the resources of iphones alone. Improved strength is had by down-porting the work from big machines to small.

A similar point applies when back-porting from recent machines to older ones, to a certain extent, highlighting dramatic improvements in software regardless of how they came to be.

[mclane]

Exactly.
Mephisto III was one of the first machines that realised opening book stored as Positions. You could setup a position and it would recognise it as theory. It was an 8 bit computer.

Colossus chess was able to let the opening book grow with learned lines.

Scisys Maestro/Analyst was able to play openings AND find transpositions in the book. This heavily increased book range. It was an 8 bit machine.

Scisys MK V and MK VI were B-Strategy dedicated chess computers by David Broughton.

Mephisto III by Hegener + Glaser was the one and only dedicated chess computer to win a championship title, and come out with a software on 8 or 16 bit that only computer 1-3 or 3-10 NPS (depending on 8 bit or 16 bit hardware)