I am running LeelaZero on my small and slow pc, Win10, without modern GPU, it is running on the CPU only (something like a cherry trail processor).
Could you give me some hints which settings might perform best under such conditions?
I presume the NN used should be as small as possible?! Any experience you want to share is welcome, thanks in advance.
(I dont expect miracles, of course, with such hardware limitations.)
Leela on a weak pc, question
Moderators: hgm, Rebel, chrisw
-
- Posts: 58
- Joined: Mon Jan 28, 2013 5:27 pm
- Location: Germany
-
- Posts: 41454
- Joined: Sun Feb 26, 2006 10:52 am
- Location: Auckland, NZ
Re: Leela on a weak pc, question
I used the 11248 network in my 72nd Amateur Series Division 4 tournament.chessico wrote: ↑Wed Jan 09, 2019 8:39 pm I am running LeelaZero on my small and slow pc, Win10, without modern GPU, it is running on the CPU only (something like a cherry trail processor).
Could you give me some hints which settings might perform best under such conditions?
I presume the NN used should be as small as possible?! Any experience you want to share is welcome, thanks in advance.
(I dont expect miracles, of course, with such hardware limitations.)
--threads=1 ramlimit-mb=256 --no-ponder
gbanksnz at gmail.com
-
- Posts: 708
- Joined: Mon Jan 16, 2012 6:34 am
Re: Leela on a weak pc, question
128x10 networks or smaller networks would be the best your cpu.chessico wrote: ↑Wed Jan 09, 2019 8:39 pm I am running LeelaZero on my small and slow pc, Win10, without modern GPU, it is running on the CPU only (something like a cherry trail processor).
Could you give me some hints which settings might perform best under such conditions?
I presume the NN used should be as small as possible?! Any experience you want to share is welcome, thanks in advance.
(I dont expect miracles, of course, with such hardware limitations.)
Try these networks or 35xx series network(128x10) .This rating list shows 35xx network are as strong as 32xx network in blitz.
https://docs.google.com/spreadsheets/d/ ... =868347223
https://github.com/dkappe/leela-chess-w ... d-Networks
-
- Posts: 708
- Joined: Mon Jan 16, 2012 6:34 am
Re: Leela on a weak pc, question
As a general rule, chess engines requires both search and evaluation.Graham Banks wrote: ↑Wed Jan 09, 2019 10:08 pmI used the 11248 network in my 72nd Amateur Series Division 4 tournament.chessico wrote: ↑Wed Jan 09, 2019 8:39 pm I am running LeelaZero on my small and slow pc, Win10, without modern GPU, it is running on the CPU only (something like a cherry trail processor).
Could you give me some hints which settings might perform best under such conditions?
I presume the NN used should be as small as possible?! Any experience you want to share is welcome, thanks in advance.
(I dont expect miracles, of course, with such hardware limitations.)
--threads=1 ramlimit-mb=256 --no-ponder
20x256 networks can give you the best evaluation only when you have adequate power to search, GPU with cuda cores.
I am pretty sure that 35xx networks or distilled networks that I mentioned in above post will be much stronger than your 20x256 in your cpu setting.
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: Leela on a weak pc, question
I’m in the process of finding the optimal dimensions for a cpu-only network.
The approach I am taking is to distill a 20 block network into a smaller size using the larger network to train the smaller one. My network of choice is 11258. I am using se–film networks, Which are somewhat stronger than the vanilla resent networks of the same size.
So far I have distilled 16x2, 96x8 and am currently distilling 112x9. They are available from my distilled networks page, along with some others. https://github.com/dkappe/leela-chess-w ... d-Networks
To give you a sense of performance, I have an ongoing tournament (including some t35 128x10-se nets).
The approach I am taking is to distill a 20 block network into a smaller size using the larger network to train the smaller one. My network of choice is 11258. I am using se–film networks, Which are somewhat stronger than the vanilla resent networks of the same size.
So far I have distilled 16x2, 96x8 and am currently distilling 112x9. They are available from my distilled networks page, along with some others. https://github.com/dkappe/leela-chess-w ... d-Networks
To give you a sense of performance, I have an ongoing tournament (including some t35 128x10-se nets).
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%)
1 ethereal : 3373 58 237.5 259 91.7%
2 crafty25.2 : 3057 36 196.5 305 64.4%
3 ID36092 : 3038 73 46.5 73 63.7%
4 ID11248-128x10-se : 3034 36 192.5 318 60.5%
5 ID35975 : 3025 46 109.0 188 58.0%
6 ID35689 : 2994 54 78.5 131 59.9%
7 ID11258-96x8-se-5 : 2949 86 22.5 49 45.9%
8 ID11248-256x12-se : 2803 76 24.5 75 32.7%
9 crafty19.18 : 2693 38 68.0 305 22.3%
10 ID11258 : 2681 63 26.0 124 21.0%
11 ID11258-16x2-se-3 : 2595 67 21.5 126 17.1%
12 ID11258-16x2-se-4 : 2588 66 23.0 139 16.5%
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
-
- Posts: 58
- Joined: Mon Jan 28, 2013 5:27 pm
- Location: Germany
Re: Leela on a weak pc, question
I thank you all very much, your answers will certainly help me to get Leela working a little better on my mini pc. And special thanks to Dietrich for his distillations, I will try them and compare how they work on my machine.
-
- Posts: 79
- Joined: Sat Aug 01, 2015 6:16 pm
- Location: France
Re: Leela on a weak pc, question
Hi
for CPU use I would suggest, worth a try:
lc0 >= 0.17 to allow ponder on
scs 64x8 network run1 550000 steps (link). Not too large for speed sake, but skilled enough. So well balanced...
Cpuct MCTS 1.700000
Minibatch size 16
Max prefetch nodes 0
Scale thinking time 1.200000
NN cache size 400000 (to be tuned according to the RAM use you grant)
Other settings=default ones
Indeed 35xx series networks (128x10) are getting interesting, latest ones seem to reach even level as the above arrangement (older ones do not). But I played very few games opposing these, so this needs further testing.
16x2 distilled 11258 is far too weak, despite its speed. Cannot compete with the scs 64x8, and loses as well against the non-distilled, very slow on CPU, same 11258 NN. I tried several higher MCTS settings to wider the search itself, as speed provides depth anyway, but no way. Big thank anyway to dkappe for his work and for sharing it, much useful (and fun!) & appreciated. I am looking forward to test the promising 96x8 one.
All my tests were using a single core on CPU (no GPU). So my conclusions may of course not apply to more powerful hardware
for CPU use I would suggest, worth a try:
lc0 >= 0.17 to allow ponder on
scs 64x8 network run1 550000 steps (link). Not too large for speed sake, but skilled enough. So well balanced...
Cpuct MCTS 1.700000
Minibatch size 16
Max prefetch nodes 0
Scale thinking time 1.200000
NN cache size 400000 (to be tuned according to the RAM use you grant)
Other settings=default ones
Indeed 35xx series networks (128x10) are getting interesting, latest ones seem to reach even level as the above arrangement (older ones do not). But I played very few games opposing these, so this needs further testing.
16x2 distilled 11258 is far too weak, despite its speed. Cannot compete with the scs 64x8, and loses as well against the non-distilled, very slow on CPU, same 11258 NN. I tried several higher MCTS settings to wider the search itself, as speed provides depth anyway, but no way. Big thank anyway to dkappe for his work and for sharing it, much useful (and fun!) & appreciated. I am looking forward to test the promising 96x8 one.
All my tests were using a single core on CPU (no GPU). So my conclusions may of course not apply to more powerful hardware
-
- Posts: 1439
- Joined: Sat Oct 27, 2018 12:58 am
- Location: Germany
- Full name: N.N.
Re: Leela on a weak pc, question
I have a question. Can someone tell me if the following installation is correct?
Currently I have only a laptop core duo 2x 2.4 Ghz 64 Bit.
Also I have LC0 v20.1 installed as Windows only.
NN = 11258-16x2-se-4.pb
client.exe
Lc0.exe
libopenblas.dll
(all in the same folder)
I clicked on the Lc0.exe and gave the command
"go nodes 100".
Then I installed Lc0 as UCI engine under ChessBase.
Lc0 runs well and reaches 1 - 3 kN /s. Search Depth 13 in the middlegame: after 150 seconds.
Is that good with my hardware? Or have I forgotten something, perhaps an important command?
Currently I have only a laptop core duo 2x 2.4 Ghz 64 Bit.
Also I have LC0 v20.1 installed as Windows only.
NN = 11258-16x2-se-4.pb
client.exe
Lc0.exe
libopenblas.dll
(all in the same folder)
I clicked on the Lc0.exe and gave the command
"go nodes 100".
Then I installed Lc0 as UCI engine under ChessBase.
Lc0 runs well and reaches 1 - 3 kN /s. Search Depth 13 in the middlegame: after 150 seconds.
Is that good with my hardware? Or have I forgotten something, perhaps an important command?
-
- Posts: 41454
- Joined: Sun Feb 26, 2006 10:52 am
- Location: Auckland, NZ
Re: Leela on a weak pc, question
Better depth than I was getting.Eduard wrote: ↑Fri Jan 11, 2019 10:42 pm I have a question. Can someone tell me if the following installation is correct?
Currently I have only a laptop core duo 2x 2.4 Ghz 64 Bit.
Also I have LC0 v20.1 installed as Windows only.
NN = 11258-16x2-se-4.pb
client.exe
Lc0.exe
libopenblas.dll
(all in the same folder)
I clicked on the Lc0.exe and gave the command
"go nodes 100".
Then I installed Lc0 as UCI engine under ChessBase.
Lc0 runs well and reaches 1 - 3 kN /s. Search Depth 13 in the middlegame: after 150 seconds.
Is that good with my hardware? Or have I forgotten something, perhaps an important command?
gbanksnz at gmail.com
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: Leela on a weak pc, question
64x6 or 64x8? The 96x8-se is already released. 112x9-se is at almost 140k of 200k steps.
Next is 144x11-se.
Next is 144x11-se.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".