Reverse-Engineering the RK3588 NPU: Building an Open Compiler to Run GPT-2 at 36 tok/s

Last year I posted about hacking the RK3588 NPU to run one vision encoder (previous post). This year I opened the whole thing up: reverse-engineered the register format, built an open compiler + runtime, and now GPT-2 and SigLIP run from PyTorch, ONNX, and JAX, no vendor SDK. Story: https://amohan.d