Jordan Dominick / Ignition

Ignition compiles a model into an .ignite container that runs on the NPU of a Windows 11 PC with an AMD Phoenix processor.

Model	On the NPU	Against AMD's stack
YOLOv8n, YOLOv8n-pose	Every layer	Faster
YOLO11n, YOLO26n	All but their attention blocks	Faster
YOLOv8s	Every layer	Faster, until AMD's compiler is tuned
YOLOv8m, l, x	Every layer	Slower
SESR M7 super-resolution	Every layer	Slower

1. Install

Paste into PowerShell (again later to update):

irm https://raw.githubusercontent.com/jdominick05/Ignition/main/install.ps1 | iex

It sets up everything under %LOCALAPPDATA%\Ignition. To also install AMD's NPU driver:

& ([scriptblock]::Create((irm https://raw.githubusercontent.com/jdominick05/Ignition/main/install.ps1))) -InstallDriver

2. Get a model

Download YOLOv8n's published COCO weights, export them and cut off the decode head, which runs on the CPU:

Set-ExecutionPolicy Bypass -Scope Process -Force; . "$env:LOCALAPPDATA\Ignition\ignition-models.ps1"
New-Item -ItemType Directory -Force models | Out-Null
iwr https://github.com/ultralytics/assets/releases/download/v8.3.0/yolov8n.pt -OutFile models\yolov8n.pt
python pipelines/yolov8n/1_export.py
python pipelines/yolov8n/1b_cut_head.py

3. Quantize

The NPU runs 8-bit integers. Quark calibrates on sample images; COCO's first 128 work:

iwr https://github.com/ultralytics/assets/releases/download/v0.0.0/coco128.zip -OutFile coco128.zip
Expand-Archive coco128.zip data
python pipelines/yolov8n/3b_quantize_cut.py --calib-dir data/coco128/images/train2017 --out models/yolov8n_cut_xint8_coco128.onnx

This took 5 minutes and 13.5 GB of temporary disk on the test PC; Quark's custom-op warnings are harmless. This is the YOLOv8n recipe's quickstart. The badges measure its benchmark model.

4. Compile and run

In the same window:

. "$env:LOCALAPPDATA\Ignition\ignition-env.ps1"
ignite-compile --input ..\ignite-xdna\models\yolov8n_cut_xint8_coco128.onnx --output ..\ignite-xdna\build\yolov8n_quickstart.ignite --silu-sigmoid
python live_ignition.py --model ..\ignite-xdna\build\yolov8n_quickstart.ignite

--silu-sigmoid computes SiLU more accurately than AMD's quantizer. A window opens on webcam 0; press q to quit, or add --source picture.jpg for an image.

More

• Performance against AMD's stack, and where AMD is ahead
• Other models, flags, Python API and limitations
• Designing and quantizing models for this NPU
• Open work

Licensed under AGPL-3.0-or-later.