Page 1 of 2
LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Fri Aug 10, 2018 1:35 pm
by AdminX
Hardware:
i7-4790 CPU @ 3.60GHz
GTX 1060
32 GB Ram
Software:
Windows 10 Pro
Shredder 13 GUI
Book:
Tournament Setting: FCP-Live-Book_V4.49
Engines:
Deep Shredder 13 x64 (4 Cores) (2048 Hash) (Permanent Brain On) (Complete Syzygybases)
LC0 ver 0.16.0 (Default Settings)
Games: https://www59.zippyshare.com/v/AnZHZHGL/file.html
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Fri Aug 10, 2018 4:33 pm
by oreopoulos
GPU hardware is not enough for such a limited time.
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Fri Aug 10, 2018 4:57 pm
by jhellis3
(Permanent Brain On)
Well....
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Fri Aug 10, 2018 6:24 pm
by Javier Ros
The Permanent Brain in Shredder is lowering the performance of LC0 because LC0 uses at least one core of CPU.
It is easy to detect it with a CPU monitor. Depending on the CPU and GPU I have seen it uses until 60 % the CPU.
Fair conditions would involve turning Permanent Brain off.
On the other hand, I also think that an average of 15 or 30 seconds per movement helps a lot to improve the play of LC0.
Do you use CUDA libraries? They multiply by 6 the performance of LC0.
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Fri Aug 10, 2018 6:39 pm
by AdminX
oreopoulos wrote: ↑Fri Aug 10, 2018 4:33 pm
GPU hardware is not enough for such a limited time.
Not so sure this statement is true. I am currently re-running the same test with these modified LC0 settings and LC0 is currently leading this time.
Code: Select all
-t 4 --minibatch-size=512 --nncache=2000000
PS: Using CUDNN
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Fri Aug 10, 2018 9:34 pm
by AdminX
AdminX wrote: ↑Fri Aug 10, 2018 6:39 pm
oreopoulos wrote: ↑Fri Aug 10, 2018 4:33 pm
GPU hardware is not enough for such a limited time.
Not so sure this statement is true. I am currently re-running the same test with these modified LC0 settings and LC0 is currently leading this time.
Code: Select all
-t 4 --minibatch-size=512 --nncache=2000000
PS: Using CUDNN
While match is still on going, here are the current standings.
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Sat Aug 11, 2018 2:11 am
by AdminX
Here is the final score of the rematch using the modified settings.
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Sat Aug 11, 2018 9:24 am
by Ron Langeveld
AdminX wrote: ↑Sat Aug 11, 2018 2:11 am
Here is the final score of the rematch using the modified settings.
With "Permanent brain" still active in Shredder the rerun is still irrelevant as pointed out earlier by Javier. Since you did not mention changing it I guess you decided to keep it enabled, or did you just forget it?
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Sat Aug 11, 2018 10:26 am
by AdminX
Ron Langeveld wrote: ↑Sat Aug 11, 2018 9:24 am
AdminX wrote: ↑Sat Aug 11, 2018 2:11 am
Here is the final score of the rematch using the modified settings.
With "Permanent brain" still active in Shredder the rerun is still irrelevant as pointed out earlier by Javier. Since you did not mention changing it I guess you decided to keep it enabled, or did you just forget it?
Correct, I decided to keep it enabled.
Re: LC0 ver 0.16.0 ID 10520 vs Deep Shredder 13 x64 40/4
Posted: Sat Aug 11, 2018 7:49 pm
by Ron Langeveld
What is your reasoning for keeping it enabled?
Common test protocol and practice is with PONDER OFF for good reason. I am sure a man with your experience knows this all too well.
In this case only Shredder is allowed to ponder during the opponents time.
Furthermore, while pondering Shredder will compete with Leela for CPU thread utilization.
To me this is like a boxing match where Leela had one arm tied to his back.
So what do you think the outcome of your test is telling us?
There are plenty of pointless tests to be found on the internet and I couldn't care less.
It's the fact that you actually decided not to address the obvious issue with your test that triggered me to post.