IRON
Close-to-metal Python API for AMD Ryzen™ AI NPUs, built on MLIR-AIE: declare operators, compile them into graphs, autotune them, and run models entirely on the NPU.
- 159
- Python
- Apache-2.0
- 3 d ago
Close-to-metal Python API for AMD Ryzen™ AI NPUs, built on MLIR-AIE: declare operators, compile them into graphs, autotune them, and run models entirely on the NPU.
A close-to-metal Python API for programming AMD Ryzen™ AI NPUs (AI Engines), built on an open-source MLIR-based compiler toolchain.
AI Engine RunTime, lite: language models and image generation on the NPU of AMD Ryzen AI processors
Open-source bring-up + verified matmul on the first-gen AMD XDNA1 (Phoenix/Hawk Point) NPU on Linux — the gen FastFlowLM/Lemonade skip. RyzenAI-npu1, mlir-aie/IRON, XRT.
Transport-neutral, vendor-agnostic virtual accelerator devices. First class cross-platform NPU backend support for Apple, Intel, AMD, Qualcomm devices, as well as Vulkan for cross-platform GPU support.
LLM and speech-to-text on the AMD Ryzen AI NPU with custom AI Engine kernels: LFM2.5-230M at 58 tok/s, Parakeet TDT 0.6B encoder at 185 ms per 10 s audio with FP32 accuracy. IRON/MLIR-AIE, XRT.
Detect, validate, enable and assess the AMD XDNA NPU on Linux. Dependency-free. Gives an honest, machine-specific LLM-feasibility verdict for XDNA 1 (Phoenix) vs XDNA 2 (Strix).
Unlocking the power of the Halo NPU: NPU + iGPU live verification, compression, and routing research on AMD Strix Halo
Open, reproducible AMD XDNA1 + Strix Point XDNA2 compute on Linux: pinned builds, CPU-reference checks, persistent C/Python runners, and an LLM roadmap.
Whisper encoder on the AMD XDNA1 NPU (Ryzen AI, Phoenix) under Linux — open toolchain only: XRT, MLIR-AIE/IRON, peano. Includes an HTTP service speaking whisper.cpp's contract.
Text embeddings on AMD XDNA NPU (Strix Halo) — OpenAI-compatible /v1/embeddings endpoint powered by all-MiniLM-L6-v2 running on the NPU tile array
ROCm/HIP on Ryzen APUs: Radeon 780M gfx1103 setup, MIOpen CK fixes, and XDNA1 NPU experiments.