Commercial Release of Ethereal 13.00 (NNUE) for AVX2 Systems

[AndrewGrant]

...SSDF is a little obsolete

Was my reliable source 30 years ago to have the exact ranking of Mephisto, Kasparov, Fidelity and Novag... (only one hardware and 1 opening book for each) Now Stockfish-dev and test NNUEs are updated everyday, but they are stuck at Stockfish 12.

Regards, AlexChess

You can anchor a list to any rating. Elo numbers have no meaning on their own. Only the difference between two elo values has any meaning mathematically. 3200, 3000, 10000. Its all the same, just tell me the difference of A and B.

[AlexChess]

Andrew, I have found a workaround to make both Ethereal 13.00 ssse3 github and Pedone 3.1 working on my emulated system!!!

Setting only 1 CPU they do not crash! 1 mN/s... Slowly but they play.

The strange thing is that all other 50+ engines tested (including 12.75 NNUE) work perfectly with 4 CPUs.

And on infinite level they work fine also with 4 CPUs.

I refer that also to Fabio Gobbato (Pedone 3.1), maybe there is a workaround for it.

Regards, AlexChess

[Cornfed]

I just ignore the absolute Elo values and just look at the differences between engines. The absolute values are different on most of the lists depending on how the scale the lists and what ratings tool they use.

Yes, this is the ONLY thing that makes sense. There will always be 'new and better' hardware...so actual 'strength' should (my opinion) only ever be in comparison with other engines...and frankly, fairly minimal - or at least 'average' hardware should be used. Let the engine speak...not the hardware!

[AndrewGrant]

I just ignore the absolute Elo values and just look at the differences between engines. The absolute values are different on most of the lists depending on how the scale the lists and what ratings tool they use.

Yes, this is the ONLY thing that makes sense. There will always be 'new and better' hardware...so actual 'strength' should (my opinion) only ever be in comparison with other engines...and frankly, fairly minimal - or at least 'average' hardware should be used. Let the engine speak...not the hardware!

"Let the engine speak, not the hardware."

For the most part, two engines will have equal elo differences on CPU A and on CPU B. But these NNUE versions break that, and now it really does matter. NNUE without AVX2 is underwhelming. NNUE with AVX2 using only 128-bit lanes is a sad marketing joke.

Its interesting to me that everyone went into a frenzy to quickly buy GPUs for Rating Lists when Leela was coming out in full force, but I've not seen any such push to have AVX2 or AVX512 machines. I was going through rating lists and of the ones that state specific CPUs, only 1 in ~8 had AVX2 support. Many lists are old (even if recently bought) Xeons with AVX only. Some lists even have machines with faux AVX2 support, using only 128-bit lanes. My original desire to release only AVX2 was because of exactly this reason.

[AlexChess]

Right, I have also completely disabled opening books. Eg if you use strongly tuned Goi 6.3 or others up to 30 moves, also if the same for all engines, the ranking results are completely fake

Also the idea to use only 1 CPU isn't so bad

[toliveanddie76]

Its interesting to me that everyone went into a frenzy to quickly buy GPUs for Rating Lists when Leela was coming out in full force,

I would also like to add that the 2000 series raytracing cards from nvidia were the first time in quite a few years where the bump in performance didn't come with much more cost.
this same thing can be said about the AMD CPUs in the last 3 years as well.
which is only now starting to bleed over into AMD's line of graphics cards.
if the leela project were 10 years older, one would have to assume that the same trend in hardware upgrades would be inevitable.

chalk it up to coincidence that the catchy word "raytracing" would directly affect the leela project in a very positive way.

[Modern Times]

Its interesting to me that everyone went into a frenzy to quickly buy GPUs for Rating Lists when Leela was coming out in full force,

Not me ! I bought a lowly GTX1050 to replace an ageing and failing power hungry old AMD GPU and used it for Leela for a time, but that was it. One of my machines does not even have a separate GPU, relies on Intel graphics on the CPU. I was never on that bandwagon.

[Chessqueen]

Right, I have also completely disabled opening books. Eg if you use strongly tuned Goi 6.3 or others up to 30 moves, also if the same for all engines, the ranking results are completely fake

Also the idea to use only 1 CPU isn't so bad

That would be a close result, between these two gladiators, can you match them at least 100 games

[Cornfed]

Its interesting to me that everyone went into a frenzy to quickly buy GPUs for Rating Lists when Leela was coming out in full force,

Not me ! I bought a lowly GTX1050 to replace an ageing and failing power hungry old AMD GPU and used it for Leela for a time, but that was it. One of my machines does not even have a separate GPU, relies on Intel graphics on the CPU. I was never on that bandwagon.

Me, I only have a stock GTX750Ti that came with my Costco desktop from 2 yrs ago! Serves me well though, I don't play games...only chess.

[AlexChess]

Right, I have also completely disabled opening books. Eg if you use strongly tuned Goi 6.3 or others up to 30 moves, also if the same for all engines, the ranking results are completely fake

Also the idea to use only 1 CPU isn't so bad

That would be a close result, between these two gladiators, can you match them at least 100 games

OK, but I must wait to have finished the other tournament, because I can only use 1 CPU with it and I have only the github version without NNUE.

Meanwhile the evergreen Ethereal 12.75 (Etherlito) using 4CPUs with MikeB's NNUE https://tests.stockfishchess.org/api/nn ... d60d95.nnu is doing really well! Next 37 games gauntlet will be with Igel 3.0.5 ARM64 Neon. I'll update the crosstable