@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.