Life In 19x19

Dear y_ich,

When I updated the app, it repeatedly opened for one second, then crashed. I deleted and reinstalled the app, then restored katago, and that fixed everything.

On the new iPhone 13 pro, I went from 20 visits per second on the last version to nearly 200 visits per second after the fix. That’s approaching decent desktop GPU performance!

Great job, and thanks!

y_ich wrote:The version 5.15.1 has been released.

1. fixed a problem that KataGo ran extremely slowly on iPadOS/iOS 15 depending on devices. As a bonus, you may get performance boost on iPadOS/iOS 15.

Enjoy!

On further experimentation, I think there is a problem with batch size. Reinstalling reset batch size to one, and it worked great.

When I reset batch size to four or more it instantly crashes when I start study. Batch size of two is inconsistent, sometimes crashing.

I’ll happily settle for batch size one now.

Thanks and good luck y_ich!

wineandgolover san,

Thank you for your helpful information!
I will investigate it soon.

The version 5.15.2 was released.

It fixed a bug that the app crashed on A14/A15 devices with iOS/iPadOS 15 when you use KataGo with batchSize more than (or equal to) 2.

As far as I can tell, the 13 Pro is not taking advantage of its extra gpu core versus the regular 13 for a Master of Go. I tested both side by side, and they alternated being faster by at most 3%.

So unless y_ich can modify the code to leverage the pro's advantage, I don’t recommend buying the pro just for this app.

wineandgolover san,

I have no way to know inside A15 and Core ML, but your experiment is quite reasonable.

A15's Neural Engine runs at 15.8FLOPS (though Apple says TOPS instead of FLOPS).
A15's GPU with 5 cores seems run at about 2.5FLOPS with FP16.
It means that 1 core runs at about 0.5FLOPS.

If Core ML uses both Neural Engine and GPU effectively, A15 with 5 cores has total 18.3FLOPS, while A15 with 4 cores has 17.8FLOPS.
The difference is 0.5FLOPS as described above, 0.5/18.3 = 0.027 ~ 3%.

I think that the match with your test is by chance.
Thermal variation may larger than it.

FYI

Thanks y_ich,

I had no idea how much more powerful the neural engine was compared to the gpu. I am glad your app either effectively uses both, or that it’s not a big deal if it isn’t using the gpu. I’m a little surprised the development tools don’t let you dig in and find out more.

Regardless I am delighted with 200 visits per second on the 40 block katago engine. So thank you again!

And I’ll repeat the caveat for others, don’t upgrade to the Pro version of iPhone 13 for this app, the regular is good enough. (The zoom camera was enough reason for me.)

y_ich wrote:wineandgolover san,

I have no way to know inside A15 and Core ML, but your experiment is quite reasonable.

A15's Neural Engine runs at 15.8FLOPS (though Apple says TOPS instead of FLOPS).
A15's GPU with 5 cores seems run at about 2.5FLOPS with FP16.
It means that 1 core runs at about 0.5FLOPS.

If Core ML uses both Neural Engine and GPU effectively, A15 with 5 cores has total 18.3FLOPS, while A15 with 4 cores has 17.8FLOPS.
The difference is 0.5FLOPS as described above, 0.5/18.3 = 0.027 ~ 3%.

I think that the match with your test is by chance.
Thermal variation may larger than it.

FYI

Hi, there seems to be a problem with the new ios 15.4. Can‘t put stones on the board.

Pippen san,

Thank you for letting me know it.
I cannot reproduce your issue on my iOS/iPadOS 15.4.
My devices are iPhone SE 2 and iPad Air 4.

Could you send e-mail to me via the app?
The template includes your device/os information.

Thanks.

I just bought a new M1-equipped ipad pro. A Master of Go is amazing. How can I see the nps on the ipad?

ez4u wrote:A Master of Go is amazing. How can I see the nps on the ipad?

If you rotate the iPad in portrait mode, you should see them in the lower right.

joachim wrote:

ez4u wrote:A Master of Go is amazing. How can I see the nps on the ipad?

If you rotate the iPad in portrait mode, you should see them in the lower right.

Thanks joachim!!

I have found a bug tested on both iPad Pro M1 and M2 versions. Start a game using KataGo 40b, then after playing a few moves, pull down the control centre from the top right, and swipe the the control centre back up. Then continue playing the next move, this next stone can be placed anywhere even within the squares and after that the app will freeze and must be restarted. Other AI engines also exhibit a similar issue of freezing.

CWE1010 wrote:I have found a bug tested on both iPad Pro M1 and M2 versions. Start a game using KataGo 40b, then after playing a few moves, pull down the control centre from the top right, and swipe the the control centre back up. Then continue playing the next move, this next stone can be placed anywhere even within the squares and after that the app will freeze and must be restarted. Other AI engines also exhibit a similar issue of freezing.

Update: I have emailed the developer of the app and he has fixed the issue and it will be solved in the next update. Thank you.

The app now shows Katago with three choices: KataGo_Light, KataGo_Medium, and KataGo_Heavy. I could not find any info that shows what these different choices represent. Does anyone know or can anyone point me in the direction of that info?