New lc0 v0.23-rc1 VERY FAST

[shrapnel]

Hi all lc0 users
Do upgrade to the new lc0 v0.23-rc1.
Its giving greatly improved speed and performance.
My speed using the sergio NNs almost doubled !
It should be good for the other NNs too.
Regards

[smatovic]

Hi all lc0 users
Do upgrade to the new lc0 v0.23-rc1.
Its giving greatly improved speed and performance.
My speed using the sergio NNs almost doubled !
It should be good for the other NNs too.
Regards

Common, numbers please

--
Srdja

[crem]

It should be neither faster nor slower, but there are reports for both.
Being slower is more plausible (e.g. due to bugs, not optimized code etc), but most likely it's just either different weights file or different settings (e.g. you have lc0.config near v0.22).

[corres]

It should be neither faster nor slower, but there are reports for both.
Being slower is more plausible (e.g. due to bugs, not optimized code etc), but most likely it's just either different weights file or different settings (e.g. you have lc0.config near v0.22).

As I saw the main difference is in time management.

[corres]

I made some speed test with this dev version:
Params:
setoption name backend value cudnn-fp16
setoption name nncachesize value 5000000
go nodes 10000000
Others are Default
GPU 1 x RTX2060 OC
Net 42872

Results:
lc0 v0.22.0 max nps = 28.7 kN/s (Depth 19, Nodes 9.8 MN)
lc0 v0.23-rc1 max nps = 23.7 kN/s (Depth 19, Nodes 9.7 MN)
So rc1 is rather slower than lc0 v0.22.0
Note
The source of rc1 contains some bugs.

[crem]

I made some speed test with this dev version:
Params:
setoption name backend value cudnn-fp16
setoption name nncachesize value 5000000
go nodes 10000000
Others are Default
GPU 1 x RTX2060 OC
Net 42872

Results:
lc0 v0.22.0 max nps = 28.7 kN/s (Depth 19, Nodes 9.8 MN)
lc0 v0.23-rc1 max nps = 23.7 kN/s (Depth 19, Nodes 9.7 MN)
So rc1 is rather slower than lc0 v0.22.0
Note
The source of rc1 contains some bugs.

It's a known bug (for few hours already). When there is no ucinewgame or position befor go command, v0.23 starts counting time from the engine start, while v0.22 starts timer at go command. So it's not a real nps drop, just timer started too early.
That will be fixed in RC2.

[corres]

I made some speed test with this dev version:
Params:
setoption name backend value cudnn-fp16
setoption name nncachesize value 5000000
go nodes 10000000
Others are Default
GPU 1 x RTX2060 OC
Net 42872

Results:
lc0 v0.22.0 max nps = 28.7 kN/s (Depth 19, Nodes 9.8 MN)
lc0 v0.23-rc1 max nps = 23.7 kN/s (Depth 19, Nodes 9.7 MN)
So rc1 is rather slower than lc0 v0.22.0
Note
The source of rc1 contains some bugs.

It's a known bug (for few hours already). When there is no ucinewgame or position befor go command, v0.23 starts counting time from the engine start, while v0.22 starts timer at go command. So it's not a real nps drop, just timer started too early.
That will be fixed in RC2.

Thanks.
I am waiting for.

[Ovyron]

Does "RC1" stand for "Release Candidate"? Because with my experience as Beta tester, I'd advice people to avoid RCs like the plague, and wait until the actual release, because the most bugs tend to hide in them. Unless you're wanting to beta test it yourself.

[mwyoung]

Hi all lc0 users
Do upgrade to the new lc0 v0.23-rc1.
Its giving greatly improved speed and performance.
My speed using the sergio NNs almost doubled !
It should be good for the other NNs too.
Regards

"New option --logit-q (UCI: LogitQ). Changes subtree selection algorithm a bit, possibly making it stronger (experimental, default off)."

Any more information on this new option? I will test this to see if works in any meaningful way.

[shrapnel]

"New option --logit-q (UCI: LogitQ). Changes subtree selection algorithm a bit, possibly making it stronger (experimental, default off)."

Any more information on this new option? I will test this to see if works in any meaningful way.

Thanks.
I switched back to v0.22 after all the uproar, but I wasn't having any real problems with v0.23 RC-1, except for a few crashes which I ascribed to some Hardware tweaks I was doing at the same time.
Anyway, do let me know the Results of your Test.