LC0 network 31740 vs Stockfish 10

[shrapnel]

Except your joke doesn't work, because to even run an OS and chess engines you need a CPU. A GPU alone can't even boot the computer.

I can also beat Leela and Stockfish at Tetris and Pacman easily.

I'll repeat what I already said to Milos and you quoted:

Not my joke, and even if it were, if all jokes relied on every element being 100% scientifically accurate we'd not have many jokes. If you wish to get riled up over such things, that's your preference, of course.

As I've already said, it seemed to me that the post from mwyoung was intended to be a joke but was not interpreted as such, so I was trying to be helpful. I didn't realize that would be interpreted as an invitation to provide condescending explanations that someone else's joke wouldn't make sense as a factual report (well, ok, I expect such nonsense from Milos).

My apologies for trying to help. Won't happen again.

Looking for people with a Sense of Humour in a Chess Forum ?
That's like organizing a Snowball fight in Hell !

[mwyoung]

I was demonstrating the absurdity of his test by being absurd.

Why would anyone run such a test with Lc0 other then to mislead people and disparage Lc0 with such an obvious bias test.

The tester claims to have vast knowledge on testing, but still posted this absurdity.

From Lc0 http://blog.lczero.org/2018/09/guide-se ... s-gui.html

"This means that in order to run Leela in full strength you will need a GPU preferably one that supports CUDA.
Running it on GPU with non CUDA Lc0 binaries means you will get many times worse performance.
Even worse running it on a CPU like normal traditional engines run, means you will get many many times worse performance.
Neural net weights in order to run fast enough they need a GPU. On CPUs they are slow as turtles."

I have been testing chess engines since 1982. And I have seen this reaction before by people wanting to protect their favorite chess programs. I have seen all the chess programs fall to the newer chess programs with better ideas. And Stockfish will be no different, as well as Lc0 if it can take the top spot.

Is Lc0 miles behind Stockfish as shown in his test? The latest TCEC ratings:


Rank Name Elo + - games score oppo. draws
1 LCZero v19-TP-11248 3580(+2) 154 154 14 79% 3404 29%
2 KomodoMCTS 2180.00/2210.00 3538 85 85 59 83% 3272 27%
3 Stockfish 051117/041017 3491 56 56 143 83% 3215 31%
4 Stockfish 010918/270918/18102108 3478 38 38 186 63% 3397 66%
5 Houdini 6.03/6.02 3469 36 36 243 69% 3318 53%
6 Stockfish 100218/260318 3468 38 38 184 65% 3378 66%
7 Komodo 1959.00/1937.00/1970.00 3463 36 36 243 67% 3321 51%
8 Komodo 2121.01/2135.10/2155.00 3448 38 38 182 54% 3416 66%
9 Stockfish 160518/180614 3447 42 42 149 61% 3383 65%
10 Houdini 6.03 3424 28 28 322 56% 3385 7

[AndrewGrant]

I was demonstrating the absurdity of his test by being absurd.

Why would anyone run such a test with Lc0 other then to mislead people and disparage Lc0 with such an obvious bias test.

The tester claims to have vast knowledge on testing, but still posted this absurdity.

From Lc0 http://blog.lczero.org/2018/09/guide-se ... s-gui.html

"This means that in order to run Leela in full strength you will need a GPU preferably one that supports CUDA.
Running it on GPU with non CUDA Lc0 binaries means you will get many times worse performance.
Even worse running it on a CPU like normal traditional engines run, means you will get many many times worse performance.
Neural net weights in order to run fast enough they need a GPU. On CPUs they are slow as turtles."

I have been testing chess engines since 1982. And I have seen this reaction before by people wanting to protect their favorite chess programs. I have seen all the chess programs fall to the newer chess programs with better ideas. And Stockfish will be no different, as well as Lc0 if it can take the top spot.

Is Lc0 miles behind Stockfish as shown in his test? The latest TCEC ratings:


Rank Name Elo + - games score oppo. draws
1 LCZero v19-TP-11248 3580(+2) 154 154 14 79% 3404 29%
2 KomodoMCTS 2180.00/2210.00 3538 85 85 59 83% 3272 27%
3 Stockfish 051117/041017 3491 56 56 143 83% 3215 31%
4 Stockfish 010918/270918/18102108 3478 38 38 186 63% 3397 66%
5 Houdini 6.03/6.02 3469 36 36 243 69% 3318 53%
6 Stockfish 100218/260318 3468 38 38 184 65% 3378 66%
7 Komodo 1959.00/1937.00/1970.00 3463 36 36 243 67% 3321 51%
8 Komodo 2121.01/2135.10/2155.00 3448 38 38 182 54% 3416 66%
9 Stockfish 160518/180614 3447 42 42 149 61% 3383 65%
10 Houdini 6.03 3424 28 28 322 56% 3385 7

Hes clearly mocking a particular poster. The fact that you don't get that is enough to end this thread.

Miles is laughing his ass off reading this thread.

[Chessqueen]

Both using 8 cores Xeon E5-2670 at 3GHz all cores frequency. SF10 running on 16 threads (HT on), Lc0 v0.19 openblas implementation with 8 threads.
Private opening set, 50 openings limited to 6 moves.
TC: 1min+1s
Result:
SF10 99.5 points
Lc0 0.5 points
+99 =1 -0
Winning percentage = 99.5%
Lc0 is 920Elo weaker than SF10.

You should use the strongest Net available for LCzero which is 32930 with 20.2 cpu 3.4 and cpuctbase = 10K for Leela, just release for TCEC. Also you should change the TC to at least 5min+3s .