Local AI & on-device inference

AMD IRON

Close-to-metal Python for AMD Ryzen AI NPUs — the layer under friendlier local-inference stacks.

  • Local AI & on-device inference
  • Open source
  • Python
  • Contributor
Python C++ NPU PyTorch mlir-aie

What it is

IRON is AMD’s open-source, close-to-metal Python API for running work efficiently on AMD Ryzen AI NPUs. If Lemonade and GAIA are the products people launch, IRON is one of the layers you reach when you care about what the NPU is actually doing.

Why it matters

NPU performance is not a compiler flag you flip once. It is kernels, data movement, and CI that still passes on real hardware. I wanted that skill in my hands, not only in someone else’s runtime.

How it works

The public tree is mostly Python with C++ underneath. Career-database attributes a large set of my commits on amd/IRON to kernel and production-path work: conv2d, AIE2P bf16 harness updates for real NPU runs, L3 staging, operator CI, and callable/spec wiring after an ABC refactor. I describe that as contribution, not ownership of the project.

Status

Upstream lives at amd/IRON. Apache-2.0. I am a contributor.

Related