xdna

A close-to-metal Python API for programming AMD Ryzen™ AI NPUs (AI Engines), built on an open-source MLIR-based compiler toolchain.

useful Gentoo overlay Curated ebuilds, AI, tools & science

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.

Practical applications for the AMD XDNA NPU on Linux — webcam effects (working) + embeddings/audio/ggml roadmap. Built on open-xdna.

Run LLMs and Ollama Qwen across AMD XDNA1 NPU, CPU, and GPU on Ryzen AI Linux systems.

Stable Diffusion image generation on AMD Ryzen AI NPUs for Linux

A lightweight TUI monitor for AMD Ryzen AI NPUs

Unlocking the power of the Halo NPU: NPU + iGPU live verification, compression, and routing research on AMD Strix Halo

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).

Open, reproducible AMD XDNA1 + Strix Point XDNA2 compute on Linux: pinned builds, CPU-reference checks, persistent C/Python runners, and an LLM roadmap.

Fix for AMD XDNA NPU driver on Ryzen AI 300 (Strix Point/Halo) — SMU init bypass patch, systemd auto-loader, full root cause analysis

Experiments, notes, and benchmarks for AMD Ryzen AI (Krackan Point NPU + Radeon iGPU) on Linux

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

Whisper ASR on the AMD Ryzen AI NPU — zero-dependency Bash wrapper over FastFlowLM, plus a RAPL power/latency benchmark harness. RTF ≈ 0.17.

Research-grade DSP, SDR, transforms, and the retained M32/M33 PQC foundation on AMD Phoenix XDNA1 using MLIR-AIE, IRON, Peano, and XRT.

Functional AMD XDNA1 (Phoenix/Hawk Point) NPU emulator for QEMU, targeting stock amdxdna + XRT guests.

Meeting-Ops UC-1 — a self-contained on-premise meeting appliance (record, transcribe, diarize, summarize) for the AMD Ryzen 9 8945HS / Radeon 780M / XDNA1 box. Includes honest AMD NPU bring-up findings. AGPL-3.0.

all-MiniLM-L6-v2 on the AMD Ryzen AI NPU (XDNA2), from hand-written MLIR-AIE kernels. C++ runtime, no Python, OpenAI-compatible /v1/embeddings. Faster than the CPU at half the cores and 1.9× less energy.

OCI container toolboxes for AMD Strix Halo — ROCm, HIP, llama.cpp, FLM, moonshine, kokoro, and other inference runtimes prebuilt for the iGPU/NPU.