LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Discussion of computer chess matches and engine tournaments.

Moderator: Ras

Hai
Posts: 713
Joined: Sun Aug 04, 2013 1:19 pm

LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Post by Hai »

LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion)
AMD R9 390 (t2) vs i7 3930k (1 core @ 4GHz)
3000 mb vs 3000 mb
TC: 1 minute game + 15 seconds per move
Aquarium 2018 gui
Ponder on selected for both
No book
The 20 possible opening moves were played = 40 games

Result:
+11 =15 -14
18.5 vs 21.5
26 elo difference
46.25% winning percentage

All games: http://www.mediafire.com/?211jrn11ml1gj

ID 310 is 26 elo weaker than Ethereal 9.65.
But it's very strong :D
AndrewGrant
Posts: 1978
Joined: Tue Apr 19, 2016 6:08 am
Location: U.S.A
Full name: Andrew Grant

Re: LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Post by AndrewGrant »

3000mb is not a valid hash size. Most, if not all engines (aside from Sf a few months ago) round the hash down to a power of 2.

2048MB would have been a proper hash size -- which is also absurd for the time control you posted. All major rating lists test with far less.

Ethereal does not support ponder, so you have inflated Leelas elo by at least 70 for that time control.

Also, am I reading this right, an entire GPU versus 1 core?
No book
The 20 possible opening moves were played = 40 games
Those two statements are a contradiction. Which one is it?
If you played all 20 openings of course the elo is close. Some of those openings are awful.

Don't make Ethereal the target of all these poorly executed Leela hit jobs.
User avatar
CMCanavessi
Posts: 1149
Joined: Thu Dec 28, 2017 4:06 pm
Location: Argentina

Re: LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Post by CMCanavessi »

AndrewGrant wrote: Sun May 20, 2018 8:08 pm 3000mb is not a valid hash size. Most, if not all engines (aside from Sf a few months ago) round the hash down to a power of 2.

2048MB would have been a proper hash size -- which is also absurd for the time control you posted. All major rating lists test with far less.

Ethereal does not support ponder, so you have inflated Leelas elo by at least 70 for that time control.

Also, am I reading this right, an entire GPU versus 1 core?
No book
The 20 possible opening moves were played = 40 games
Those two statements are a contradiction. Which one is it?
If you played all 20 openings of course the elo is close. Some of those openings are awful.

Don't make Ethereal the target of all these poorly executed Leela hit jobs.
Wait, Leela doesn't support ponder either, unless it was added very very recently to lc0, which I believe it wasn't. :D
My chess engine Facón: https://github.com/CMCanavessi/facon
AndrewGrant
Posts: 1978
Joined: Tue Apr 19, 2016 6:08 am
Location: U.S.A
Full name: Andrew Grant

Re: LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Post by AndrewGrant »

A quick look at the source shows a whole bunch of ponder related variables.

Not sure if they are all wired up or not, however.

In which case that elo difference does not exist, and Hai has no idea what he is doing.
yanquis1972
Posts: 1766
Joined: Wed Jun 03, 2009 12:14 am

Re: LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Post by yanquis1972 »

A) it's a completely expected outcome.

B) CCRL has only tested the latest version of ethereal on one core.

C) it looks more to me like your engine got noticed because it won a TCEC title than that someone is persecuting you or it.
AndrewGrant
Posts: 1978
Joined: Tue Apr 19, 2016 6:08 am
Location: U.S.A
Full name: Andrew Grant

Re: LC0 (ID 310) vs Ethereal 9.65 (S12 Div.4 Champion) special

Post by AndrewGrant »

A) Given the opening set, yes.

B) Okay?

C) Sure, but I don't want to be apart of these daily threads about Leela vs [Insert literally any engine hundreds of elo higher than Leela]