Ryzen AI XDNA2 Hardware
This page covers the hardware needed for FastFlowLM / Lemonade NPU. NPU inference uses XDNA2 and is independent of the ROCm GPU stack: GPU tutorials use the iGPU (gfx1151 on Strix Halo); NPU tutorials use /dev/accel/accel0.
See also: Lemonade Linux + FLM · FastFlowLM Linux
1. Supported processors
FastFlowLM NPU LLMs on Linux require XDNA2. XDNA1 cannot load the NPU2 weights used in these tutorials.
| Family | Codename | NPU | This tutorial |
|---|---|---|---|
| Ryzen AI Max 300 | Strix Halo (e.g. MAX+ 395, gfx1151) | XDNA2 | Primary; sample host is GMKtec EVO-X2 |
| Ryzen AI 300 | Kraken Point, Strix Point | XDNA2 | Supported; official Gemma 4 benches often use Kraken |
| Ryzen AI 400 | Gorgon Point | XDNA2 | Supported |
| Z2 Extreme | Handheld | XDNA2 | Supported |
| Ryzen AI 7000 / 8000 / 200 | Phoenix / Hawk Point | XDNA1 | Not supported for NPU LLMs |
The Strix Halo iGPU is gfx1151. Lemonade llamacpp:rocm / llamacpp:vulkan use that iGPU and do not occupy /dev/accel.
2. Runtime stack
| Item | Requirement |
|---|---|
| NPU firmware | 1.1.0.0 or later (validated at 1.1.2.65) |
| Kernel + driver | Kernel 7.0+ with in-tree amdxdna, or amdxdna-dkms on Ubuntu 24.04 |
| Userspace | libxrt-npu2 + FastFlowLM (flm) |
| memlock | ulimit -l = unlimited |
| Device node | /dev/accel/accel0; user in render (also add video) |
| IOMMU | cmdline has both amd_iommu=on and iommu=pt |
3. Check the machine
bash
lspci -nn | grep -iE 'processing|display|npu'
ls -l /dev/accel/accel0On Strix Halo the NPU PCI ID is commonly 1022:17f0. After flm is installed:
bash
unset ROCM_PATH HIP_PATH LD_LIBRARY_PATH
flm validateYou should see the NPU device, firmware 1.1.x, and memlock infinity / unlimited.
4. Relationship to ROCm GPU
- They can coexist: ROCm for GPU (vLLM / Ollama / llama.cpp), FastFlowLM for NPU.
- Do not mix env vars:
source /opt/rocmthenflmoften becomesNo such device with index '0'. UnsetROCM_PATH HIP_PATH LD_LIBRARY_PATHfirst. - Unified memory: NPU and large GGUF share system RAM. Gemma 4 E4B NPU is ~9 GiB; Qwen3.6-35B-A3B NPU is ~29 GiB. Keep one large model resident.
- One NPU client at a time: if Lemonade has loaded an FLM model, do not also
flm serve.