320x24 Jh net is doing well in cccc bonus. Good news for upcoming T60 series!!

[Nay Lin Tun]

As it is early era of NN, no one knows the optimum size of network for current consumer hardwares.

Although 20x256 are much stronger than 10x128 in 20xx RTX hardwares, bigger networks did not do well previously.

It is the first time to see this bigger net 320x24 doing well in cccc.

It is a very good news for upcoming T60 series with 320 x24 net.

Downlink for Jh 370 000 net. https://github.com/jhorthos/lczero-trai ... a-Training

[Eduard]

Thanks for the networks! The smaller 192x16 even run well on the smartphone. Nice thing.

[Hai]

As it is early era of NN, no one knows the optimum size of network for current consumer hardwares.

Although 20x256 are much stronger than 10x128 in 20xx RTX hardwares, bigger networks did not do well previously.

It is the first time to see this bigger net 320x24 doing well in cccc.

It is a very good news for upcoming T60 series with 320 x24 net.

Downlink for Jh 370 000 net. https://github.com/jhorthos/lczero-trai ... a-Training

Current consumer hardware: All available gpus, weakest to strongest.

Optimum is: stronger gpu = bigger network size, weaker gpu = smaller network size.

It was obvious that 24 x 320 size will be better when using strong gpus.

Even the 40 x 256 network was at least equal or even better than the 20 x 256, when using stronger gpus and more time.
Don't forget that 20 x 256 got 10 to 20 times more training games / new networks per day compared to the 40 x 256.
With this development speedup 20 x 256 should have been some hundreads of elos stronger than 40 x 256.
But it wasn't so it's a terrible result and a waste of ressources.