Life In 19x19

Thanks a lot!

Haha, Katago 1.11.0 is released!
https://github.com/lightvector/KataGo/r ... ag/v1.11.0

Can you continue updating?

Fortunately, this is not necessary. There are only a few commits between release 1.11 and the revision that I built for BadukAI 1.14. And these commits are mainly concerned with the code for training and with documentation, which don't matter for BadukAI. So the KataGo code in BadukAI 1.14 is practically equivalent to release 1.11.

in the worst case, depending on the hardware, what is the speed gain of an optimized network? and at the best?

If you compare the speeds of optimized and original network, you get:

a) If you choose the optimal number of search threads (in terms of speed) for each network:
In the worst (HW dependent) case the speed gain is factor 2 (e.g. 20 playouts/second instead of 10).
In the best (HW dependent) case the speed gain is factor 24 (e.g. 240 playouts/second instead of 10).

b) If you choose one search thread for both networks:
In the worst (HW dependent) case the speed gain is factor 2.
In the best (HW dependent) case the speed gain is factor 40.

I find a compatibility issue!
I have two phones(Dimensity 720 and Snapdragon 662)
When I use 40B weight, the output of winning rate is wrong.

In my vivo y52s (Dimensity 720)
I have test all weights
Kata 10b -> ok
kata 20b -> bad
lz 15b -> ok
lz 40b ->ok

Since version 1.14 BadukAI runs the optimized KG networks (20b and 40b) on NPU instead of CPU on Mediatek processors. This works well (and fast) on Dimensity 1100 (according to user reports). But your tests show that on the (slightly older) Dimensity 720 the NPU can't handle the KG network correctly and produces garbage output. What a pity.

Since I don't know on which Mediatek processors it works and on which it doesn't, I will add a switch to KataGo's config settings in the next release, so that every user can turn this on or off depending on whether it works on his/her device.

OK，thanks

akigo wrote:If you compare the speeds of optimized and original network, you get:

a) If you choose the optimal number of search threads (in terms of speed) for each network:
In the worst (HW dependent) case the speed gain is factor 2 (e.g. 20 playouts/second instead of 10).
In the best (HW dependent) case the speed gain is factor 24 (e.g. 240 playouts/second instead of 10).

b) If you choose one search thread for both networks:
In the worst (HW dependent) case the speed gain is factor 2.
In the best (HW dependent) case the speed gain is factor 40.

stunningly!!!

My friend bought a new phone (Redmi K30 Pro with Dimensity 9000 chip), but the speed of kata 20b is very slow, it seems that there is no acceleration, only 20n/s.
The performance of Dimensity 9000 is stronger than that of Dimensity 1100, it should exceed 140n/s.

Is it possible to provide a new version that allows me to test different acceleration schemes.

Can you share the android system log for the time period when KataGo is starting ? Then I could check for error messages which give a hint what's going wrong. Otherwise I wouldn't know what to change ...

I made a new version (1.15) available on play store https://play.google.com/store/apps/deta ... r.baduk_ai and github https://github.com/aki65/aki65.github.i ... ag/v1.15.0 with the change I announced in post #428.

lovego wrote:My friend bought a new phone (Redmi K30 Pro with Dimensity 9000 chip), but the speed of kata 20b is very slow, it seems that there is no acceleration, only 20n/s.
The performance of Dimensity 9000 is stronger than that of Dimensity 1100, it should exceed 140n/s.

Is it possible to provide a new version that allows me to test different acceleration schemes.

You can do this with the version from https://github.com/aki65/aki65.github.i ... ag/v1.15.1 (see instructions there).

Haha, my Dimensity 720 works well when useSpecificNpuAccelerator is 3/4/5