NN-testrun of Lc0 0.27.0 68695 finished - impressive progress!
https://www.sp-cc.de
(Perhaps you have to clear your browsercache or reload the website)
SPCC: Testrun of Lc0 68695 finished
Moderators: hgm, Rebel, chrisw
-
- Posts: 2444
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
-
- Posts: 671
- Joined: Sun Jan 26, 2020 10:38 pm
- Location: Turkey
- Full name: Mehmet Karaman
Re: SPCC: Testrun of Lc0 68695 finished
It is very surprising that even after 1 year, progress is still going on. If Deep Mind didn't end the project at the end of 44 million games, we could see a much strong AlphaZero chess engine.
-
- Posts: 3553
- Joined: Thu Jun 07, 2012 11:02 pm
Re: SPCC: Testrun of Lc0 68695 finished
Impressive indeed.
I'm curious, have you done any duplicate test runs at any point using 12 threads instead of 20 threads on the Ryzen, to see if the results are the same with SMT, and any analysis on the games re average depths etc ?
-
- Posts: 2444
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
Re: SPCC: Testrun of Lc0 68695 finished
If you would read the testing-conditions on my website, you would know, that the NN-Testruns are not running on a Ryzen CPU...Modern Times wrote: ↑Sun Apr 25, 2021 11:48 amImpressive indeed.
I'm curious, have you done any duplicate test runs at any point using 12 threads instead of 20 threads on the Ryzen, to see if the results are the same with SMT, and any analysis on the games re average depths etc ?
Using less threads? Duplicate tests? Definitly not. Waste of time and resources.
Analyzing games? Feel free to do it. All games can be downloaded on my website. All played moves with eval, depth and thinkingtime.
-
- Posts: 3553
- Joined: Thu Jun 07, 2012 11:02 pm
Re: SPCC: Testrun of Lc0 68695 finished
Yes I know the NN testing is done on your Intel machine. It was more of a general question on the other testing that you do on the Ryzen.pohl4711 wrote: ↑Sun Apr 25, 2021 3:21 pm
If you would read the testing-conditions on my website, you would know, that the NN-Testruns are not running on a Ryzen CPU...
Using less threads? Duplicate tests? Definitly not. Waste of time and resources.
Analyzing games? Feel free to do it. All games can be downloaded on my website. All played moves with eval, depth and thinkingtime.
-
- Posts: 26
- Joined: Mon Oct 26, 2020 3:05 pm
- Full name: Pedro
Re: SPCC: Testrun of Lc0 68695 finished
I remember that Alphazero trained only 24 hours and turned that monster (today it has already been overcome). But imagine if Deep Mind left A0 training longer with the resources they have? I think it would reach a level beyond the limits of reality, don't you think?
-
- Posts: 2444
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
Re: SPCC: Testrun of Lc0 68695 finished
OK.Modern Times wrote: ↑Sun Apr 25, 2021 4:29 pmYes I know the NN testing is done on your Intel machine. It was more of a general question on the other testing that you do on the Ryzen.pohl4711 wrote: ↑Sun Apr 25, 2021 3:21 pm
If you would read the testing-conditions on my website, you would know, that the NN-Testruns are not running on a Ryzen CPU...
Using less threads? Duplicate tests? Definitly not. Waste of time and resources.
Analyzing games? Feel free to do it. All games can be downloaded on my website. All played moves with eval, depth and thinkingtime.
The point is, on 99% of all notebooks, it is not possible to switch off hyperthreading. And so, on a 12 core machine, you have 24 virtual cores and there is no way, to use the 12 hardware-cores directly. And the OS distributes the computing power in very small timesnippets to the virtual cores. So, when playing engine-matches with an time-increment, the move- and calculation time is much, much longer and this means, each thinking engine gets "the same peace of cake" from the CPU.
No need, to worry about that.
And using only 12 cores would be just a waste of CPU-Power.And the engines still would not have direct access to a hardware-core, but to virtual cores. So, this makes no sense at all.
-
- Posts: 3293
- Joined: Wed Mar 08, 2006 8:15 pm
-
- Posts: 3293
- Joined: Wed Mar 08, 2006 8:15 pm
Re: SPCC: Testrun of Lc0 68695 finished
I played match against J94-130 in GTX card 60+0.6:
68695 has 45% higher NPS.
Code: Select all
1 68695 +16 +17/=123/-10 52.33% 78.5/150
2 J94-130 -16 +10/=123/-17 47.67% 71.5/150
Jouni