amd/IRON

IRON is a close-to-metal programming framework that enables efficient execution on AMD Ryzen AI NPUs through Python bindings to the MLIR-AIE dialect.

The tool addresses the challenge of programming neural processing units by providing direct access to NPU hardware capabilities while abstracting away low-level complexity. It works by exposing MLIR-AIE as a Python API, allowing developers to write kernels that compile down to NPU instructions. The framework includes a library of pre-built operators covering common deep learning patterns like matrix multiplication, multi-head attention, normalization, and activation functions, with support for operator fusion to optimize performance across multiple operations.

Developers should choose this tool if they are targeting AMD Ryzen AI NPUs and need fine-grained control over execution while maintaining reasonable productivity. It suits projects requiring high-performance inference, particularly large language models, as demonstrated by the included end-to-end Llama inference example. The extensible architecture allows custom operators beyond the pre-built library, making it appropriate for specialized workloads that demand NPU acceleration.

Development activity shows consistent engagement with automated testing infrastructure in place. The project maintains active contribution guidelines and welcomes pull requests. Code quality standards are enforced through automated formatting checks. Community interaction occurs through a dedicated Discord channel for discussion and support.