🔧 Mainline Linux is getting NPU support for the Rockchip RK3576, thanks to serious reverse-engineering work by Ga Hing Woo (the chip's TRM doesn't even document the compute registers involved)!
-> Byte-exact int8 convolution achieved
-> Chained-layer stalls documented as an honest negative result, not hidden
-> Tested on Radxa ROCK 4D: Llama-3.2-1B ~13 tok/s, MobileNet ~169 fps
👇 (1/2)
