Anyone got some more converted models to share

[StuartIanNaylor]

Apols guys as a dumb englishman quite impressed by the rk3588 npu and would like to bench some models.

I have been running the basic resnet-18 model but found its pretty rare compared to resnet-50 or MobilenetV2.

This isn't a issue but QQ looks a bit of a mare for an Englishman :) I haven't got round to converting any models yet and just curious to why the conversion toolkit is on Ubuntu and runtime on Debian, no bother but just curious as will run up docker with 18.04

The NPU on the rk3588 is it really a x3 2 TOP core NPU? That as well as running a model on all cores with scaling loss you can run x3 models either in auto or chose a core? The governor is userspace is that via rknn-toolkit2 as was trying to get an approx fps with resnet-18 and was confused by how much variance. Changing to performance seems to keep variance minimal. Also on the Rock5b the sram has not been implemented and is this needed to get updates to /sys/devices/platform/fdab0000.npu/devfreq/fdab0000.npu/load as it seems static at 100@1000000000Hz

PS it would be fantastic to have a model zoo somewhere for this NPU :)