As a side note, using -t3 and after switching from cudnn-9.2-windows7-x64-v7.14 to cudnn-9.2-windows10-x64-v7.14, I'm getting much better speeds, over 18k, as of now, and over 95 % GPU Utilization.
(I'm using Windows 8.1).
Better and better....
lc0-win-20180512-cuda90-cudnn712-00
Moderators: hgm, Rebel, chrisw
-
- Posts: 1339
- Joined: Fri Nov 02, 2012 9:43 am
- Location: New Delhi, India
Re: lc0-win-20180512-cuda90-cudnn712-00
i7 5960X @ 4.1 Ghz, 64 GB G.Skill RipJaws RAM, Twin Asus ROG Strix OC 11 GB Geforce 2080 Tis
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: lc0-win-20180512-cuda90-cudnn712-00
Again, as I said it's an indication of speed not necessarily strength. Try running your set of parameters and default set of parameters with fixed number of nodes and you'll see that higher PUCT values are not so good at all.Albert Silver wrote: ↑Mon May 21, 2018 4:39 amIt's a good theory, except that I tested all my PUCT values at 3+0 and 5+0, and then proposed them to GCP. He in turn tested them at very fast TCs, but stopped the test early due to disastrous results.The lower PUCT value was stronger at very short TCs, while the higher PUCT values only shined at longer TCs. Mind you, these CLOP optimized values are not the ones I had hit on, just similar. Here are the two for comparison, though i have not tested the optimized ones to compare.
CLOP
Slowmover: 2.75
cPUCT: 2.8
FPU: -0.08
Mine
Slowmover: 2.0
cPUCT: 3.0
FPU: 0.0
Problem is no one has a good clue about what is the best MCTS for chess. It depends on too many things.
AB search has been developed for decades. On the other hand MCTS is still in its infancy and it's current implementation is naive compared to AB. Currently ppl are trying to find the best solution with a single fixed parameter. Ofc it's not gonna work.
-
- Posts: 3019
- Joined: Wed Mar 08, 2006 9:57 pm
- Location: Rio de Janeiro, Brazil
Re: lc0-win-20180512-cuda90-cudnn712-00
That's not actually true, but I admit it is well hidden. If you want to see the real GPU usage when running Leela (any build), in Task Manager, go to the Performance tab, then in one of the empty graphs, over them, change the drop-down menu (it is easy to not realize it is one) to Compute_0.
I attached a screenshot to illustrate.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."
-
- Posts: 3019
- Joined: Wed Mar 08, 2006 9:57 pm
- Location: Rio de Janeiro, Brazil
Re: lc0-win-20180512-cuda90-cudnn712-00
I admit I'm pretty confused by this comment. In your first reply, you said it increases the nodes per second, but not necessarily the strength. Fair enough. Now you are telling my to run them with a fixed nodes count? How is that going to illustrate a lack of strength from more nodes??Milos wrote: ↑Mon May 21, 2018 6:39 pmAgain, as I said it's an indication of speed not necessarily strength. Try running your set of parameters and default set of parameters with fixed number of nodes and you'll see that higher PUCT values are not so good at all.Albert Silver wrote: ↑Mon May 21, 2018 4:39 amIt's a good theory, except that I tested all my PUCT values at 3+0 and 5+0, and then proposed them to GCP. He in turn tested them at very fast TCs, but stopped the test early due to disastrous results.The lower PUCT value was stronger at very short TCs, while the higher PUCT values only shined at longer TCs. Mind you, these CLOP optimized values are not the ones I had hit on, just similar. Here are the two for comparison, though i have not tested the optimized ones to compare.
CLOP
Slowmover: 2.75
cPUCT: 2.8
FPU: -0.08
Mine
Slowmover: 2.0
cPUCT: 3.0
FPU: 0.0
That said, I am running a practical test, which will take some time. I am running LC0 with default settings and then modded ones against Spike 1.4 and will simply look at the results. Each gauntlet is 1m+1s and takes hundreds of games each so a bit of patience will be needed, but it should help shed some light on whether there are clear-cut gains, a loss of ability, or just nothing that stands out.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: lc0-win-20180512-cuda90-cudnn712-00
What you select as Compute_0 is actually a GPU engine that Task Manager is using to get calculation and that is dependent on the GPU you are using and also it's not there in every version of Task Manager. In may case is Engine 1 that is taking measurements from Lc0 not Engine 0.Albert Silver wrote: ↑Mon May 21, 2018 6:45 pmThat's not actually true, but I admit it is well hidden. If you want to see the real GPU usage when running Leela (any build), in Task Manager, go to the Performance tab, then in one of the empty graphs, over them, change the drop-down menu (it is easy to not realize it is one) to Compute_0.
I attached a screenshot to illustrate.
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: lc0-win-20180512-cuda90-cudnn712-00
Pretty simple nps gain is kind of fixed, i.e. 20% more nps ~15 more Elo. However, on different TCs you reach different depth and than loss due to higher PUCT value can be more than nps gain so overall net result could be negative.Albert Silver wrote: ↑Mon May 21, 2018 6:50 pmI admit I'm pretty confused by this comment. In your first reply, you said it increases the nodes per second, but not necessarily the strength. Fair enough. Now you are telling my to run them with a fixed nodes count? How is that going to illustrate a lack of strength from more nodes??Milos wrote: ↑Mon May 21, 2018 6:39 pmAgain, as I said it's an indication of speed not necessarily strength. Try running your set of parameters and default set of parameters with fixed number of nodes and you'll see that higher PUCT values are not so good at all.Albert Silver wrote: ↑Mon May 21, 2018 4:39 amIt's a good theory, except that I tested all my PUCT values at 3+0 and 5+0, and then proposed them to GCP. He in turn tested them at very fast TCs, but stopped the test early due to disastrous results.The lower PUCT value was stronger at very short TCs, while the higher PUCT values only shined at longer TCs. Mind you, these CLOP optimized values are not the ones I had hit on, just similar. Here are the two for comparison, though i have not tested the optimized ones to compare.
CLOP
Slowmover: 2.75
cPUCT: 2.8
FPU: -0.08
Mine
Slowmover: 2.0
cPUCT: 3.0
FPU: 0.0
Last edited by Milos on Mon May 21, 2018 7:09 pm, edited 1 time in total.
-
- Posts: 3019
- Joined: Wed Mar 08, 2006 9:57 pm
- Location: Rio de Janeiro, Brazil
Re: lc0-win-20180512-cuda90-cudnn712-00
Not in my case.Milos wrote: ↑Mon May 21, 2018 7:04 pmPretty simple nps gain is kind of fixed, i.e. 20% more nps ~15 more Elo. However, on different TCs you reach different depth and than loss due to higher PUCT value can be more than nps gain so overall net result is negative.Albert Silver wrote: ↑Mon May 21, 2018 6:50 pmI admit I'm pretty confused by this comment. In your first reply, you said it increases the nodes per second, but not necessarily the strength. Fair enough. Now you are telling my to run them with a fixed nodes count? How is that going to illustrate a lack of strength from more nodes??
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: lc0-win-20180512-cuda90-cudnn712-00
Did you try running at longer TC for example 20s per move fixed Lc0-default parameters vs Lc0-your parameters?Albert Silver wrote: ↑Mon May 21, 2018 7:09 pmNot in my case.Milos wrote: ↑Mon May 21, 2018 7:04 pmPretty simple nps gain is kind of fixed, i.e. 20% more nps ~15 more Elo. However, on different TCs you reach different depth and than loss due to higher PUCT value can be more than nps gain so overall net result is negative.Albert Silver wrote: ↑Mon May 21, 2018 6:50 pm
I admit I'm pretty confused by this comment. In your first reply, you said it increases the nodes per second, but not necessarily the strength. Fair enough. Now you are telling my to run them with a fixed nodes count? How is that going to illustrate a lack of strength from more nodes??
-
- Posts: 3019
- Joined: Wed Mar 08, 2006 9:57 pm
- Location: Rio de Janeiro, Brazil
Re: lc0-win-20180512-cuda90-cudnn712-00
You can choose Compute_1 instead.Milos wrote: ↑Mon May 21, 2018 7:02 pmWhat you select as Compute_0 is actually a GPU engine that Task Manager is using to get calculation and that is dependent on the GPU you are using and also it's not there in every version of Task Manager. In may case is Engine 1 that is taking measurements from Lc0 not Engine 0.Albert Silver wrote: ↑Mon May 21, 2018 6:45 pmThat's not actually true, but I admit it is well hidden. If you want to see the real GPU usage when running Leela (any build), in Task Manager, go to the Performance tab, then in one of the empty graphs, over them, change the drop-down menu (it is easy to not realize it is one) to Compute_0.
I attached a screenshot to illustrate.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: lc0-win-20180512-cuda90-cudnn712-00
I use process explorer so there I selected Engine 1.Albert Silver wrote: ↑Mon May 21, 2018 7:12 pmYou can choose Compute_1 instead.Milos wrote: ↑Mon May 21, 2018 7:02 pmWhat you select as Compute_0 is actually a GPU engine that Task Manager is using to get calculation and that is dependent on the GPU you are using and also it's not there in every version of Task Manager. In may case is Engine 1 that is taking measurements from Lc0 not Engine 0.Albert Silver wrote: ↑Mon May 21, 2018 6:45 pm
That's not actually true, but I admit it is well hidden. If you want to see the real GPU usage when running Leela (any build), in Task Manager, go to the Performance tab, then in one of the empty graphs, over them, change the drop-down menu (it is easy to not realize it is one) to Compute_0.
I attached a screenshot to illustrate.