What it is
AMD-XDNA-NPU-AI is my own WIP notebook of a stack for running AI on AMD XDNA NPUs. The README is explicit: this is not a finished product.
The intended pieces are an NPU hardware abstraction, an AMD Quark quantization pipeline, IRON ObjectFifo memory management, ONNX Runtime sessions with the VitisAI execution provider, and real-time monitoring (CPU/GPU/NPU via psutil, GPUtil, and WMI).
Why it matters
I wanted a single place to wire the packages I actually use when I think about Ryzen AI: onnxruntime with VitisAI, mlir-aie (IRON), llvm-aie (PEANO), Quark, and torch for loading weights. It is the messy map underneath friendlier demos.
How it works
The repo already has a sketched package layout — config, hardware (npu manager, thermal monitor, resource allocator), quantization, memory, plus planned inference, monitoring, and coordination modules. Treat the README’s later performance targets as targets, not results I am claiming here. I have not independently restated unpublished numbers from that list.
Status
WIP, public at antmikinka/AMD-XDNA-NPU-AI. For closer-to-metal production code I contributed to, see AMD IRON.