Leela Chess Zero Makes Mind-Blowing Piece Sacrifice!

[jp]

Anyway, my point is that from the get-go, maybe I was born with a brain that wasn't as adept at playing chess as the world's champion. That's unfair!

No, it's not.

A "fair" test is one that truly tests what it is supposed to be testing and is not affected by things it is not supposed to be testing. e.g. A test of engines is supposed to be testing software, not hardware. A chess game or match between humans is supposed to be testing their ability at chess. etc. etc.

[Dann Corbit]

However, NN engines stink on normal hardware so they require hardware roughly 1000 times faster than normal chess hardware. If you run them on a CPU (which is possible) they will stink up the joint.

So toss out the concept of equal hardware. Or don't bother entering the NN engines in the contests.

The equal dollars thing won't work either, unless you buy one of those special boxes that have a raft of GPUs in them. The machine used for TCEC is really a big ticket item, and the GPUs are just commodity items of about $1000 each.

[jhellis3]

You just set equal power budget if you want "fairness." And I would say consumer level GPUs are normal hardware. Maybe not familiar to non-gamers, but still quite normal.

[Ovyron]

A test of engines is supposed to be testing software, not hardware.

I can wave my hand and say with 100% certainty that if the hardware is CPU then Stockfish wins and if the hardware is GPU then Leela wins.

And I didn't even need to do any testing.

This is the most possible fairness, so this was solved.

*Washes hands*

[Ovyron]

You just set equal power budget if you want "fairness." And I would say consumer level GPUs are normal hardware. Maybe not familiar to non-gamers, but still quite normal.

So now the power of software depends entirely of the supply and demand of the hardware those entities run on, which dictates their price for the budget.

Now I can make Stockfish or Leela be much stronger by filling the market with the hardware they run on so their prices plummet and now whoever I want gets to be the victor in a "fair" match. Wait, this looks like result fixing, see the problem now?

[jhellis3]

Nope. My post had nothing to do with price btw.... Not sure what that strawman is all about, but it doesn't concern me so whatever...

[Milos]

The equal dollars thing won't work either, unless you buy one of those special boxes that have a raft of GPUs in them. The machine used for TCEC is really a big ticket item, and the GPUs are just commodity items of about $1000 each.

You keep repeating this BS ad nauseam. Are you aware that Ryzen 3990X is stronger than TCEC CPU setup (almost same nps but with less threads)? Probably not. And 3990X you can buy for already 3450$. And RTX2880Ti you can't get anywhere for less than 1200$ plus that machine also needs a CPU of at least 400$, plus V100 are noticeably more powerful than 2880Ti. So stop BSing about 4x1000$ total cost.
A really fair comparison is what I already wrote to you (but you are just ignoring it because you know Lc0 would loose badly) 3990X machine vs 2x2080Ti+3950X.

[Milos]

A test of engines is supposed to be testing software, not hardware.

I can wave my hand and say with 100% certainty that if the hardware is CPU then Stockfish wins and if the hardware is GPU then Leela wins.

And I didn't even need to do any testing.

This is the most possible fairness, so this was solved.

*Washes hands*

Do you ever do anything but trolling? Gee.

[jp]

The equal dollars thing won't work either, unless you buy one of those special boxes that have a raft of GPUs in them. The machine used for TCEC is really a big ticket item, and the GPUs are just commodity items of about $1000 each.

You keep repeating this BS ad nauseam. Are you aware that Ryzen 3990X is stronger than TCEC CPU setup (almost same nps but with less threads)? Probably not. And 3990X you can buy for already 3450$. And RTX2880Ti you can't get anywhere for less than 1200$ plus that machine also needs a CPU of at least 400$, plus V100 are noticeably more powerful than 2880Ti. So stop BSing about 4x1000$ total cost.
A really fair comparison is what I already wrote to you (but you are just ignoring it because you know Lc0 would loose badly) 3990X machine vs 2x2080Ti+3950X.

And 3990X PBO is clearly better than TCEC's hardware.

[Milos]

The equal dollars thing won't work either, unless you buy one of those special boxes that have a raft of GPUs in them. The machine used for TCEC is really a big ticket item, and the GPUs are just commodity items of about $1000 each.

You keep repeating this BS ad nauseam. Are you aware that Ryzen 3990X is stronger than TCEC CPU setup (almost same nps but with less threads)? Probably not. And 3990X you can buy for already 3450$. And RTX2880Ti you can't get anywhere for less than 1200$ plus that machine also needs a CPU of at least 400$, plus V100 are noticeably more powerful than 2880Ti. So stop BSing about 4x1000$ total cost.
A really fair comparison is what I already wrote to you (but you are just ignoring it because you know Lc0 would loose badly) 3990X machine vs 2x2080Ti+3950X.

And 3990X PBO is clearly better than TCEC's hardware.

That is probably true even in terms of nps, 3990X is around 3.45GHz all cores frequency and E5-4669v4 is only 2.4GHz. 64/88*3.45/2.4=1.045 so roughly 4.5% more nps. But even the same nps with 30% less threads is a noticeable improvement.