Lc0 in CCC matches

[Jouni]

Why they still use outdated v0.27.0-rc1 and J94-100??

[Pedro]

And what are the strongest nets or version of Lc0 today?

[Jouni]

Obviously same situation as with SF: no progress with nets in 6 months! Both have reached their limits?

[pohl4711]

And what are the strongest nets or version of Lc0 today?

In my testings, T60 net 68002 is the strongest and clearly stronger than J94-100.

https://www.sp-cc.de/nn-vs-sf-testing.htm

[Dann Corbit]

I think they should go to double sized networks, like FF2.
On high end hardware at decent time controls, 384x30 are clearly the strongest amongst those tested by the LC0 group.
I have found on my hardware the big networks are considerably stronger tactically.
Since tactical strength appears to be a function of network size, it makes sense to try 2x nets and maybe 4x nets.

[Cornfed]

I think they should go to double sized networks, like FF2.
On high end hardware at decent time controls, 384x30 are clearly the strongest amongst those tested by the LC0 group.
I have found on my hardware the big networks are considerably stronger tactically.
Since tactical strength appears to be a function of network size, it makes sense to try 2x nets and maybe 4x nets.

Do I smell an area where Team Dragon are going?

[Nay Lin Tun]

I think they should go to double sized networks, like FF2.
On high end hardware at decent time controls, 384x30 are clearly the strongest amongst those tested by the LC0 group.
I have found on my hardware the big networks are considerably stronger tactically.
Since tactical strength appears to be a function of network size, it makes sense to try 2x nets and maybe 4x nets.

The point is resources!
Bigger nets need longer time to train.And nobody knows which training methods/parameters is best to train. They are training on smaller nets like T70 to get those parameters before they can run on bigger nets.

Also, nobody knows plateau in current 30x384 nets are due to fully trained or just improperly trained.

[Dann Corbit]

GPUs are also getting more powerful. Every 6 months or so, there will be a new crop, much more powerful than the previous.
I think the real problem with the LC0 approach right now is the miners who are hoarding the GPU cards.
It ramps up the price to sky-high and makes it hard for people to get the new technology.
And what is the produce of these bitcoin type miners?
It is the currency of criminals.

[MikeB]

GPUs are also getting more powerful. Every 6 months or so, there will be a new crop, much more powerful than the previous.
I think the real problem with the LC0 approach right now is the miners who are hoarding the GPU cards.
It ramps up the price to sky-high and makes it hard for people to get the new technology.
And what is the produce of these bitcoin type miners?
It is the currency of criminals.

So true Dann,
Bitcoin Energy Consumption Debate

They say it crashes every 4 years, so wait for next Bitcoin Winter to get the latest GPU.

[Milos]

GPUs are also getting more powerful. Every 6 months or so, there will be a new crop, much more powerful than the previous.
I think the real problem with the LC0 approach right now is the miners who are hoarding the GPU cards.
It ramps up the price to sky-high and makes it hard for people to get the new technology.
And what is the produce of these bitcoin type miners?
It is the currency of criminals.

Lol, what do computer chess enthusiasts produce? What do you produce except burning electricity? What is the purpose of pushing a few Elo in a practically solved and dead game of chess???
Miners at least produce money and enable crypto transactions .