@lovego
There is no katago release yet that supports the new 18b architecture and the code is still in branches, so I can't release anything yet.

But I have made a test version that supports the new 18b network, available at https://github.com/aki65/aki65.github.io/releases/tag/v1.18.1. I have added two "optimized" versions of the UEC-network (unfortunately it turned out that HW support on android for the structure and operations of the new network architecture is far worse than for the old networks, so they are much slower):

1) b18c384nbt-uec-20221121b-quant
This is an 8-bit quantized network. On Snapdragon 888 and 8+ gen 1, it behaves as follows:
If you choose UseSpecificNpuAccelerator = 5 in settings (and numSearchThreads = 3 and tick "parallelEvaluation"), it will run on CPU and be about 50% faster than the original network. If you choose UseSpecificNpuAccelerator = 1, it should run on NPU, but:
On SD 888 it stays on CPU, so no speed change
ON SD 8+ gen 1 it runs on NPU (pretty fast), but produces wrong results.
It would be interesting how it behaves on Dimensity 9000, so it would be very nice, if your friend could check that.

2) b18c384nbt-uec-20221121b
This is a 32-bit network. On Snapdragon 888 and 8+ gen 1, it runs best if you choose UseSpecificNpuAccelerator = 3 in settings (and numSearchThreads = 3 and tick "parallelEvaluation"). Then it's about 3.5 times faster than the original network. Here also it would be interesting if your friend could check its behaviour on Dimensity 9000.

Thank you for your new version. I tested it on my mobile phone. It is really slow and weaker than the previous 20b weight

Hello, I found a new problem. I tested a new mobile phone using Qualcomm's latest CPU (8Gen2) and found that the speed was very slow, about 15n/s. What may be the problem? I have switched and tried all the npu options

BadukAI uses the android NNAPI interface for neural networks. Here benchmark scores for SD8gen2 are much lower than for SD8(+)gen1. Compare

SD8gen2: https://browser.geekbench.com/search?k=ml_inference&q=kalama

SD8(+)gen1: https://browser.geekbench.com/search?k=ml_inference&q=taro

This is very strange since the new chipset should be more powerful. I can only speculate on the reason: Maybe the NNAPI driver for the new chipset isn't working correctly yet (after all the new phones were rushed out only a few weeks after presentation of SD8gen2).

KataGo 1.12.2 has been released! Will badukAI update it?

I will certainly update at some point, but not right now, since I'm still investigating if the new architecture networks can somehow be used effectively on android. Since the result may affect the code of my android backend, I will only update when that task is finished (otherwise I might have to update twice).

I made a new test release at https://github.com/aki65/aki65.github.io/releases/tag/v1.18.2 which provides an improved version of the quantized 18b network. If you are interested, check it out.

will baduk_ai__armeabi-v7a come out?
on my devices, 18b will be faster and stronger anyway

I'm also waiting for version armeabi-v7a

The merge with KataGo 1.12.2 is still in progress, but I made a preliminary version for both arm64-v8a and armeabi-v7a here: https://github.com/aki65/aki65.github.io/releases/tag/v1.18.3

@lovego
If you still have access to Snapdragon 8 Gen 2 device, you could try the latest test release: https://github.com/aki65/aki65.github.io/releases/tag/v1.18.4

It's crazy, 400+n/s, amazing speed! Now it's faster than my computer

After the Kata engine is updated to v1.12.4, I think it's time to release an official version. This version is so great!

I tested on my huawei phone, this engine crashes!

@lovego
Which chipset does the huawei phone have ?
Did you load a .dlc file or a .tflite file ?

Kirin 990

It should be that dlc weights are not supported. Other weights are OK

I know this is a longshot and offtopic, but akigo, since you are such a tensorflow wizard and a charitable person I thought I would try . Would you know if there is any way force GPU support of the sorting of abstracts in this hobby project I am using to learn python:

https://github.com/GithubUser017/Entrez ... on_Windows

The code is here:

https://github.com/GithubUser017/Entrez ... rch_gui.py

It is very slow because it sorts and organizes the abstracts using only one CPU thread. Whether you have time to look at it or not, sincere thanks for your wonderful project!

To speed up sorting of strings in python, I found this:

https://stackoverflow.com/questions/48039359/what-is-the-fastest-way-to-sort-strings-in-python-if-locale-is-a-non-concern

Are you sure (i. e. have you profiled) that the sorting step is your bottleneck, instead of, say, using a scripting language or that sleep(1) in the middle?

_________________
A good system naturally covers all corner cases without further effort.

Thanks so much for the replies, will look into it!