AMD Ryzen R9 3950X new build

[jp]

Research in diagrammatic/combinatorial knot theory and related graph theory and linear algebra over Z/2Z. And an occasional chess analysis too.

Ok, I give up. What kind of knots are you working on and what application do they have?

Hopefully only unknot and trefoil. We need his hardware for chess analysis!

[Illusion]

I would like to thank you all for this thread. It has helped me a lot.

I have another question however. It was stated that even with an RTX2600 card, Leela will run out of RAM when doing longer analysis.
Is there any way this could be quantified? It does not have to be precise. For example, lets say I am running infinite analysis on a tactical line in the Najdorf (around move 15 or such) on 55 cores with AB engine, and also running the same position on a pair of 2080-Ti's using the other 10-11 cores for this on LC0....

I leave it run 'overnight'. Will I run out of my 128GB RAM? I hope you get the question I am asking

Can anyone discuss this to provide some/any insight for this?

Life would be much simpler if correspondence chess were outlawed! Still, I would think that allowing 8 hours per position might be enough even for 2800 level players.

[Hai]

I would like to thank you all for this thread. It has helped me a lot.

I have another question however. It was stated that even with an RTX2600 card, Leela will run out of RAM when doing longer analysis.
Is there any way this could be quantified? It does not have to be precise. For example, lets say I am running infinite analysis on a tactical line in the Najdorf (around move 15 or such) on 55 cores with AB engine, and also running the same position on a pair of 2080-Ti's using the other 10-11 cores for this on LC0....

I leave it run 'overnight'. Will I run out of my 128GB RAM? I hope you get the question I am asking

Can anyone discuss this to provide some/any insight for this?

Life would be much simpler if correspondence chess were outlawed! Still, I would think that allowing 8 hours per position might be enough even for 2800 level players.

It will run out of your RAM after 12 to 16 hours, when you are using a pair of 2080-Ti's, when you are using a LC0 Sergio 30x384 net.
If you use a smaller net, it will run out of RAM much faster.
That's also a reason why I'm using LC0 Sergio 30x384 nets.
Let's take a look how long we must wait for 40x512 nets. Then we can do much longer analysis.

[corres]

...
I have another question however. It was stated that even with an RTX2600 card, Leela will run out of RAM when doing longer analysis.
Is there any way this could be quantified? It does not have to be precise. For example, lets say I am running infinite analysis on a tactical line in the Najdorf (around move 15 or such) on 55 cores with AB engine, and also running the same position on a pair of 2080-Ti's using the other 10-11 cores for this on LC0....
I leave it run 'overnight'. Will I run out of my 128GB RAM? I hope you get the question I am asking.
...

You should make some experience getting answer for your question.
In general if you use rather large pagefile.sys (at least 100-200 GB) you would not fear from freezing of your analysis. For this you can create pagefile.sys even on your HDD but it is better if you use a kind of SSD for pagefile.sys.

[Laskos]

I would like to thank you all for this thread. It has helped me a lot.

I have another question however. It was stated that even with an RTX2600 card, Leela will run out of RAM when doing longer analysis.
Is there any way this could be quantified? It does not have to be precise. For example, lets say I am running infinite analysis on a tactical line in the Najdorf (around move 15 or such) on 55 cores with AB engine, and also running the same position on a pair of 2080-Ti's using the other 10-11 cores for this on LC0....

I leave it run 'overnight'. Will I run out of my 128GB RAM? I hope you get the question I am asking

Can anyone discuss this to provide some/any insight for this?

Life would be much simpler if correspondence chess were outlawed! Still, I would think that allowing 8 hours per position might be enough even for 2800 level players.

With a regular T40-like 256x20 net (and various same-size improvements like Sergio's one), take as rule of thumb (very approximate) for 2x 2080 Ti eat about 30GB RAM every hour. For bigger nets, compare their "surface" to 256x20, and divide the 30GB / hour by that factor.

[mwyoung]

...
I have another question however. It was stated that even with an RTX2600 card, Leela will run out of RAM when doing longer analysis.
Is there any way this could be quantified? It does not have to be precise. For example, lets say I am running infinite analysis on a tactical line in the Najdorf (around move 15 or such) on 55 cores with AB engine, and also running the same position on a pair of 2080-Ti's using the other 10-11 cores for this on LC0....
I leave it run 'overnight'. Will I run out of my 128GB RAM? I hope you get the question I am asking.
...

You should make some experience getting answer for your question.
In general if you use rather large pagefile.sys (at least 100-200 GB) you would not fear from freezing of your analysis. For this you can create pagefile.sys even on your HDD but it is better if you use a kind of SSD for pagefile.sys.

I found even if you do nothing. In windows after the memory is used by Lc0. Lc0 started using virtual memory. And the analysis just continued.