SPCC: New longtime test SF nnue vs Lc0

Discussion of computer chess matches and engine tournaments.

Moderators: hgm, Rebel, chrisw

User avatar
pohl4711
Posts: 2434
Joined: Sat Sep 03, 2011 7:25 am
Location: Berlin, Germany
Full name: Stefan Pohl

SPCC: New longtime test SF nnue vs Lc0

Post by pohl4711 »

I started a new kind of testruns: SF (with nnue) vs. Lc0 with longer thinking-time (average game-duration 20 minutes) and J.Noomen's lowdraw-openings. Learn more in the NN vs SF testing section (scroll down below the NBSC-Armageddon testings of Lc0 nets) of my website. Short wins can be seen in the "View SF vs Lc0 games"- section.
Each testrun 300 games with 150 Noomen lowdraw-openings (selected openings from TCEC superfinals) and 5'+3'' thinking-time (Lc0) / 7.5'+4.5'' (Stockfish). This thinking-time gives a perfect Leela-Ratio of 1.0 on the used PC hardware: i7-8750H 2.6GHz (Hexacore, TurboBoost mode off) Notebook, RTX 2060 GPU. Stockfish running on 11 hyperthreading threads (1 thread for Windows). Average game-duration: 20 minutes

First testrun:
Stockfish 200810 112bb1c8cdb5 (3662 SPCC-Elo) vs Lc0 0.26.1 LS 15: 300 (+ 78,=196,- 26), 58.7 % (+62 Elo)

As you can see, the first testrun is a clear win for Stockfish with nnue vs Lc0 with Leelenstein 15 net.


https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website)
Albert Silver
Posts: 3019
Joined: Wed Mar 08, 2006 9:57 pm
Location: Rio de Janeiro, Brazil

Re: SPCC: New longtime test SF nnue vs Lc0

Post by Albert Silver »

In my experience, high-end laptop GPUs tend to throttle after 30mins, even in near ideal conditions, when running something as GPU-intensive as Lc0 or NN training (where the GPU is regularly pushed to 90% or more.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."
User avatar
pohl4711
Posts: 2434
Joined: Sat Sep 03, 2011 7:25 am
Location: Berlin, Germany
Full name: Stefan Pohl

Re: SPCC: New longtime test SF nnue vs Lc0

Post by pohl4711 »

Albert Silver wrote: Mon Aug 17, 2020 1:53 pm In my experience, high-end laptop GPUs tend to throttle after 30mins, even in near ideal conditions, when running something as GPU-intensive as Lc0 or NN training (where the GPU is regularly pushed to 90% or more.
Then you buy the wrong laptops.

In my XMG notebook the RTX never is hotter than 65 Celsius. Even though it is very hot in my appartement in Berlin right now...
And lc0 plays vs Stockfish. When Stockfish thinks, the GPU is not in use. So the GPU runs, then stops and cools down, then runs again...with each move.
Absolutely no problem. By the way, it is the same machine, where I do the faster testruns of Lc0 vs Stockfish 200418, 500 games each run (2.5 days) And here, the best lc0 nets (Leelenstein 14.3 and 15) score around 60% vs Stockfish. https://www.sp-cc.de/nn-vs-sf-testing.htm
Does anybody believe, this is possible, on an overheated, throttled GPU???

Just buy quality, Made In Germany.
https://bestware.com/en/
ernest
Posts: 2041
Joined: Wed Mar 08, 2006 8:30 pm

Re: SPCC: New longtime test SF nnue vs Lc0

Post by ernest »

pohl4711 wrote: Mon Aug 17, 2020 2:46 pm
In my XMG notebook...
https://bestware.com/en/
Hi Stefan,
Can you detail the config of your XMG Notebook ?
TIA.
User avatar
pohl4711
Posts: 2434
Joined: Sat Sep 03, 2011 7:25 am
Location: Berlin, Germany
Full name: Stefan Pohl

Re: SPCC: New longtime test SF nnue vs Lc0

Post by pohl4711 »

ernest wrote: Mon Aug 17, 2020 5:48 pm
pohl4711 wrote: Mon Aug 17, 2020 2:46 pm
In my XMG notebook...
https://bestware.com/en/
Hi Stefan,
Can you detail the config of your XMG Notebook ?
TIA.
I have 2 of them.
On the older one, I do my Lc0-testruns: RTX 2060 (mobile,but NOT MaxQ-version (would be slower)), 16 GB and Intel i7-8750H Hexacore CPU.
The new one, for my Stockfish-ratinglist tests - a true monster...XMG is the first company on the planet, to build an AMD Desktop Ryzen 9 CPU into a notebook:
RTX 2060, 32 GB, and AMD Ryzen 3900 12-core CPU (!!!). Paid around 1850€.
Here the direct link to the configuration-website - all these notebooke are built in Germany with the individually chosen components, the customer wants to buy:
https://bestware.com/de/xmg-apex-15.html
The cooling fans are awesome - the machines never get hot. But, I have to say it clear: If you use them for computerchess, the fans are loud. Dont expect a silent machine!

PS: You can have a 16-core Ryzen CPU, too, but this one is much more expensive (+400€) and these CPU is running slower, when all cores are used, otherwise its getting too hot, so you will not gain much Elo out of this CPU. So, I believe the 12core-CPU is the best choice for computerchess in a notebook.
Sam Watson
Posts: 8
Joined: Thu Aug 17, 2017 2:31 pm

Re: SPCC: New longtime test SF nnue vs Lc0

Post by Sam Watson »

pohl4711 wrote: Mon Aug 17, 2020 2:46 pm In my XMG notebook the RTX never is hotter than 65 Celsius.
Do you have some software to know the temperture?
User avatar
pohl4711
Posts: 2434
Joined: Sat Sep 03, 2011 7:25 am
Location: Berlin, Germany
Full name: Stefan Pohl

Re: SPCC: New longtime test SF nnue vs Lc0

Post by pohl4711 »

Of Course. On all XMG machines, a program is installed by XMG, for manual controlling the fans of the GPU and CPU. And this program displays the temperature of CPU and GPU.