Product
What it is
A kernel lab for AMD AI Engine NPUs. Not a compiler. Not FastFlowLM. Not an ASR runtime.
IRON Lab is a kernel lab. You write IRON/Peano tile programs, compile them to xclbin, occupy the NPU so a second process cannot clobber the array, and test against a CPU reference.
What it is not
- Not a model compiler. We use mlir-aie / IRON as-is. Gaps get upstream bugs, not a fork.
- Not FastFlowLM. FastFlowLM runs LLMs with closed kernels. Different workload, different contract.
- Not an ASR runtime. Transcription for Papyrus / morphist-asr is a later consumer of a compiled bundle.
Who it is for
- End user — check the NPU, load a compiled design, see who occupies the array, unload.
- Kernel author — scaffold an operator, compile with Peano, test vs NumPy, ship an xclbin.
Public name
Repo WorldTreeNetwork/iron-lab. Binary ironlab. License MIT OR Apache-2.0.
See hardware.md, cli.md, occupancy.md.