The future of computer chess

[Uri Blass]

To fully satisfy Thorsten's question (since he demands a 6502 5Mhz comparison doing (in my case) about 500-600 nodes per second) is a match based on nodes.

That's what Thorsten wants. What I, and many others want, is...

Nodes = Legal Moves (1 ply)

The engine evaluates each legal move in the position, but is forbidden from generating a game tree, or otherwise generating any forward moves from the current position. It's then all about what the engine knows about chess, and nothing else whatsoever.

Software engineers are always going to love recursion more than they love a calculator - but a calculator is the true demonstration of subject mastery: you ask, you get the correct answer. No messing.

I do not know about other but I see no point in 1 ply search competition when people practically can use search inside their evaluation function(if you evaluate if a position is mate or not mate it is also search because you do it by searching pseudo legal moves of the opponent and finding that they are illegal).

Knowldge in chess is also knowledge to search.

[towforce]

To fully satisfy Thorsten's question (since he demands a 6502 5Mhz comparison doing (in my case) about 500-600 nodes per second) is a match based on nodes.

That's what Thorsten wants. What I, and many others want, is...

Nodes = Legal Moves (1 ply)

The engine evaluates each legal move in the position, but is forbidden from generating a game tree, or otherwise generating any forward moves from the current position. It's then all about what the engine knows about chess, and nothing else whatsoever.

Software engineers are always going to love recursion more than they love a calculator - but a calculator is the true demonstration of subject mastery: you ask, you get the correct answer. No messing.

I do not know about other but I see no point in 1 ply search competition when people practically can use search inside their evaluation function(if you evaluate if a position is mate or not mate it is also search because you do it by searching pseudo legal moves of the opponent and finding that they are illegal).

Knowldge in chess is also knowledge to search.

1. The stronger a human player is, the less likely that he will need to check all the opponent's moves to know that his move is checkmate

2. A good evaluation would give the highest score to a move that gives checkmate

3. After making the move, other parts of the program would then be allowed to check for checkmate/draw by repetition etc for presentation (announcing the outcome and stopping the game)

[mclane]

To be precise, my posting included all big 8bit CPUs of that age:

6502 or z80a or Motorola M6809

IMO Ed Schroeders 8 Bit engines had the highest elo performance on 5 mhz 6502.

6502 was also in Apple II, C64, Acorn BBC,
Z80a was in ZX Spectrum, Sam Coupe, Amstrad CPC464
M6809 was in Dragon32 and TRS 80 i think.

[towforce]

To be precise, my posting included all big 8bit CPUs of that age:

6502 or z80a or Motorola M6809

IMO Ed Schroeders 8 Bit engines had the highest elo performance on 5 mhz 6502.

6502 was also in Apple II, C64, Acorn BBC,
Z80a was in ZX Spectrum, Sam Coupe, Amstrad CPC464
M6809 was in Dragon32 and TRS 80 i think.

Back in the day, these were revolutionary CPUs that made home computing devices practical - but by today's standards, they are very poor CPUs: running stockfish on them would be a non-starter, so comparison that way is not possible.

In your situation, I would do comparison as follows:

8-bit single thread 5 MHz CPU with 6 registers and 56 instructions + 16kb RAM = 1 compute unit

I can see that the arguments will start already: the 6502 had 151 unique opcodes - so should you pick opcodes or instructions? My answer: doesn't matter! The comparison is going to be hopelessly inaccurate - but it will be nice to see the order of magnitude difference!

Compare that with a 96 core Threadripper CPU with an NVIDIA H100 graphics card and come up with a comparable "compute units" number for that setup. Then you can draw a chart of Elo against compute units.

I have no doubt that you'll see dramatic diminishing returns. You might even need to make the compute units a logarithmic scale!

What this won't do, though, is separate hardware gains from software gains: this is why I suggested reverse compiling engines for the 6502 CPU and recompiling them for a kick-ass modern rig. Per Ed's post, they'll probably pick up an extra 200 Elo points if you're lucky.

We can still say, of course, that the 8-bit pioneers pulled out an impressive standard of chess for the level of knowledge and the hardware constraints of the time.

[smatovic]

[...]
What this won't do, though, is separate hardware gains from software gains: this is why I suggested reverse compiling engines for the 6502 CPU and recompiling them for a kick-ass modern rig. Per Ed's post, they'll probably pick up an extra 200 Elo points if you're lucky.
[...]

Ed and Uri estimated +200 Elo for backporting new techniques onto 6502, and Bill Foster did his Sargon 1 port onto x86, 3 ply extra.

[...]
I know, the Sargon port by Bill Foster:

Re: History of Memory Wall in Computer Chess?
viewtopic.php?p=856414#p856414
[...]

--
Srdja

[towforce]

A nice exercise for mclane: the programs for the dedicated chess computers from the golden age were only about 15kb or so, so quite limited in their complexity: as he is strongly attracted to them and their style of play, how about he:

1. Reverse compiles them to a high level language

2. If that language isn't his favourite, then transpile it into a different language

3. Create a document briefly outlining the principles used in chess in the 8-bit era

4. Feed the program and the description to a good chatbot and ask it to explain what's happening

He will then have perfect insight into the great minds of the golden age of the dedicated chess computer: more so than those pioneers themselves, who will have forgotten what they did by now.

[towforce]

Ed and Uri estimated +200 Elo for backporting new techniques onto 6502, and Bill Foster did his Sargon 1 port onto x86, 3 ply extra.

[...]
I know, the Sargon port by Bill Foster:

Re: History of Memory Wall in Computer Chess?
viewtopic.php?p=856414#p856414
[...]

That's not what Bill Forster did: according to the post you linked, he took the Sargon program, unchanged, and ran it on hardware which was 6000x faster.

[syzygy]

To be precise, my posting included all big 8bit CPUs of that age:

6502 or z80a or Motorola M6809

IMO Ed Schroeders 8 Bit engines had the highest elo performance on 5 mhz 6502.

6502 was also in Apple II, C64, Acorn BBC,
Z80a was in ZX Spectrum, Sam Coupe, Amstrad CPC464
M6809 was in Dragon32 and TRS 80 i think.

I thought you were mistaken on the TRS-80, which was based on the Z80, but apparently there was also a "TRS-80 Color Computer" based on the 6809:
https://en.wikipedia.org/wiki/TRS-80_Color_Computer

[mclane]

Everything in those years was running on this hardware. As home computers or dedicated units.

Either software bought for Homecomputers or dedicated chess computers bought from Novag, Mephisto, Scisys, Fidelity, CXG, Conchess, Mephisto, Saitek.

[mclane]

To be precise, my posting included all big 8bit CPUs of that age:

6502 or z80a or Motorola M6809

IMO Ed Schroeders 8 Bit engines had the highest elo performance on 5 mhz 6502.

6502 was also in Apple II, C64, Acorn BBC,
Z80a was in ZX Spectrum, Sam Coupe, Amstrad CPC464
M6809 was in Dragon32 and TRS 80 i think.

I thought you were mistaken on the TRS-80, which was based on the Z80, but apparently there was also a "TRS-80 Color Computer" based on the 6809:
https://en.wikipedia.org/wiki/TRS-80_Color_Computer

Yes. You are absolutely correct.
But these 3 cpus rocked the world.