LCzero network

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Milos
Posts: 3313
Joined: Wed Nov 25, 2009 12:47 am

Re: LCzero network

Post by Milos » Thu Dec 13, 2018 4:27 pm

jp wrote:
Thu Dec 13, 2018 3:52 pm
Laskos wrote:
Thu Dec 13, 2018 3:44 pm
Nets vary in speed, even among test30 nets there is quite a variarion. Strength is measured at same time control, not same number of nodes. Same time control could mean different number of nodes.
Interesting. Do you know why? I guess most assume same-size NN = same speed.
Different net, different policy head, different expansion rate, different NNcache usage, different nps.

chrisw
Posts: 1417
Joined: Tue Apr 03, 2012 2:28 pm

Re: LCzero network

Post by chrisw » Thu Dec 13, 2018 4:42 pm

jp wrote:
Thu Dec 13, 2018 3:52 pm
Laskos wrote:
Thu Dec 13, 2018 3:44 pm
Nets vary in speed, even among test30 nets there is quite a variarion. Strength is measured at same time control, not same number of nodes. Same time control could mean different number of nodes.
Interesting. Do you know why? I guess most assume same-size NN = same speed.
Because, amongst other reasons, the amount of data involved, its variability in amount, the need to parallelise and thread and throw data at GPU in volume, means that all kind of weird things conspire together to speed stuff up or slow stuff down. If your code is serial, it’s also quite logical not just in “logic” but in time flow, but when parallel with all sorts of different stuff happening at the same time, usually with unknown times at which one part of the parallel events are done, well, it’s a bit like herding cats. Stuff happens and so on, little changes in flow can have large effects, or not. Robust process it isn’t. Deterministic, it isn’t.

Post Reply