New android app "BadukAI"

[lovego]

Dimensity 9000 is about 300n/s (KG 20b) when useSpecificNpuAccelerator is 3
The fastest chip at present!

[lovego]

Oh, I get 400+n/s when numSearchThread is 3 and paralletEvalution is on

[akigo]

It's great to see that. Thank you very much for your testing. I would like to add an entry for Dimensity 9000 in the performance table on BadukAI's homepage. Since all the visit counts in this table were measured in this position (black to play)



it would be nice if you could check the speed of Dimensity 9000 in this position.

It's important to compare speeds only in the same position, since the type of position affects the visit count: In a highly symmetrical position (for example an empty board), the search tree will include a lot of positions which are mirrors of already visited positions. KataGo will handle a visit of such a position by cache lookup instead of computation, so the visit count will be higher.

By the way, correct move is E3 (not too difficult, I am told, but already beyond my level), so also a good way to check if computations are correct.

[lovego]

OK, I'll let my friend try.
Another thing, in version 1.16, the LZ engine crashes when the analysis is paused, and the speed is very slow on Dimensity 9000, can LZ choose the NPU accelerator to use?

[lovego]

Regarding board recognition, I found a good solution!
Use yolov3 to identify black stone, white stone, corners (no stone), sidelines (no stone) and empty position.

https://blog.csdn.net/wyzegg/article/details/108029789
https://github.com/zhuoyiyao97/YOLO-GO

[lovego]

Dimensity 9000 in that position, my friend test it has 280n/s
Snapdragon 888 in that position, I test it only has 220n/s

[akigo]

Regarding board recognition, I found a good solution!
Use yolov3 to identify black stone, white stone, corners (no stone), sidelines (no stone) and empty position.

Very interesting, I'll have a closer look when I am less busy.

OK, I'll let my friend try.
Another thing, in version 1.16, the LZ engine crashes when the analysis is paused, and the speed is very slow on Dimensity 9000, can LZ choose the NPU accelerator to use?

This is now included in a new test version https://github.com/aki65/aki65.github.i ... ag/v1.15.2

Dimensity 9000 in that position, my friend test it has 280n/s
Snapdragon 888 in that position, I test it only has 220n/s

I checked again and still get 260 n/s for SD 888, but this may be caused by different measuring methods:

I set "seconds/move" to 30 in "Manage time settings", wait a few seconds after loading the sgf (to make sure it's completely loaded), then tap "analyse all" and watch how high the visit count goes.

In engine config I use "parallelEvaluation" and 2 search threads.

Moreover, as usual on android, all performance measurement results vary with device temperature, battery state, background activity and so on ...

Anyway, Dimensity 9000 is clearly faster, great work by Mediatek and TSMC.

[lovego]

OK, look forward to the official release of the new version

[And]

viewtopic.php?p=271968#p271968
Interestingly, in the analysis mode, Zenith 7 quickly (less than a minute) found the correct move, and after a few minutes of analysis, CS ZERO did not even consider it (from 20 different moves). in game mode 9 dan, both programs chose a different move

[And]

@akigo

when do you intend to make a new optimized network 60b?

[akigo]

When 60b is stronger than 40b at time parity. Before that, it doesn't make much sense considering all the disadvantages of 60b (twice the creation time, twice the disk space, twice the loading time in the app).

[And]

on my tablet with 1 GB of RAM, the optimized 60b network works well! it's incredible!

[And]

When 60b is stronger than 40b at time parity. Before that, it doesn't make much sense considering all the disadvantages of 60b (twice the creation time, twice the disk space, twice the loading time in the app).

network 60b not only plays differently, but also "reads" better even with 5 playouts.
in the match (24 games) BadukAI (opt 40b 10 pl) - Sabaki (40b 10 pl) there were 8 games lasting less than 100 moves: 99,76,84,41,96,92,99,40
in a match (24 games) Sabaki (40b 10 pl) - Sabaki (60b 5 pl) in those games where the network 60b lost, only one game lasted less than 100 moves - 94
EDIT60b - kata1-b60c320-s583

[And]

the optimized 60b network appeared almost 8 months ago. is it possible to make a new network at least once every half a year? on some devices, such as mine, an unoptimized network does not work even 40b

[Vargo]

... when do you intend to make a new optimized network 60b?
... When 60b is stronger than 40b at time parity.

This result is probably not directly useable, but at time parity (0.5 s/move, laptop with GTX1660Ti)
b40s111d271 wins 80-20 against b60c320s583 (twogtp, 100 games, average length ~180 moves, average time/game for 40b : ~46s, and 48s for 60b)