2）LeelaZero：您可以在15个方块的网络（针对40b自打训练）和40个方块的网络（LZ 273）之间进行选择。该实现针对android进行了优化，可以在最新的Snapdragon芯片组上使用DSP，因此在此类设备上将非常快。例如，在Snapdragon 855+上，它的速度约为200 n / s（对于15个块）和50 n / s（对于40个块）。
I have created a new android app "BadukAI" which allows you to play and analyse with two AIs:
1) Katago: To my knowledge the first native implementation on android. It uses the 20 blocks network from version 1.5.0 and runs on the GPU via OpenCL, so your device needs OpenCL capability (which most have nowadays). On Snapdragon 855+ processor it does about 10 nodes per second (on top OpenCL-performers like the latest Exynos 990 it should be twice as fast).
2) LeelaZero: You can choose between a 15 blocks network (trained against 40b self-play) and a 40 blocks network (LZ 273). The implementation is optimized for android and can use the DSP on recent Snapdragon chipsets so will be quite fast on such devices. On Snapdragon 855+ for example it does about 200 n/s (for 15 blocks) and 50 n/s (for 40 blocks).
Unfortunately I can't distribute the app via google play store since google forces new apps on play store to target at least SDK 29 and this SDK does not permit an app to start another executable anymore (such as leelazero or katago).
Therefore I made it available via my github page:
https://aki65.github.io/
If you are interested, you can read a complete description and download the app from there.
发表于 2020-12-7 19:37
提升卡
喧嚣卡
变色卡
发表于 2020-12-8 01:29
楼主
