ignite-xdna: models and NPU containers for AMD XDNA1

These are the quantized models and the compiled NPU containers from ignite-xdna, a bare-metal engine for the AIE2 array in AMD's XDNA1 NPU (Phoenix and Hawk Point; developed on a Ryzen 7 8700G and a Ryzen 5 8645HS) on Windows. A .ignite container is the whole program the engine dispatches: the NPU instruction stream, the packed weights and the kernel image. It runs through ignite-xdna's runtime or the Ignition SDK, not through ONNX Runtime's VitisAI EP.

The *_xint8.onnx files are AMD Quark XINT8 QDQ models, and the containers were compiled from them. The ONNX files also run on AMD's own Ryzen AI 1.7.1 stack, which is the baseline the measurements compare against.

Contents

Folder	ONNX	Container	NPU only?	Upstream licence
yolov8n/	yolov8n_cut_xint8.onnx	yolov8n_full.ignite	yes: 0 host segments	AGPL-3.0
yolov8s/	yolov8s_cut_xint8.onnx	yolov8s.ignite	yes: 0 host segments	AGPL-3.0
yolov8n-pose/	yolov8n-pose_cut_xint8.onnx	yolov8n_pose.ignite	yes: 0 host segments	AGPL-3.0
yolo11n/	yolo11n_cut_xint8.onnx	yolo11n.ignite	no: 1 host segment (the C2PSA attention core)	AGPL-3.0
yolo-world/	yolov8s-worldv2_cut_xint8_gptqcv2.onnx, yolow_text_encoder.onnx, yolow_text.npz	yolow.ignite	no: 4 host segments (the text cross-attention)	AGPL-3.0, plus MIT (CLIP)
sesr_m7/	sesr_m7_xint8.onnx, sesr_m7_fp32.onnx	sesr_m7.ignite	yes: 0 host segments	Apache-2.0
resnet50/	resnet50_xint8_c64.onnx, test_resnet50_head.onnx	resnet50_head.ignite (classifier head only)	yes: 0 host segments	Apache-2.0
modnet/	modnet_cut_xint8_calibfix.onnx	modnet_cut_dense.ignite	no: 24 host segments	Apache-2.0
fixtures/	inputs for ignite-xdna's test suite			per subfolder

A "host segment" is a region the container declares for the CPU, run through ONNX Runtime once per frame. "0 host segments" means the whole network runs on the NPU. That is read from each container's own manifest, and it is a claim about that file, not about the model family.

modnet_cut_dense.ignite is not a whole-network NPU result, and not a speed result. About half of its segments are CPU regions. In ignite-xdna's measurements, the CPU-region compute alone exceeds AMD's whole frame on the same model. It is published because its output is exact where AMD's stack is not.

resnet50_head.ignite is the classifier head alone: it takes the pooled 2048-channel features (declared input [1,2048,20,20]) and returns 1000 logits. It is not a ResNet50 backbone. resnet50_xint8_c64.onnx is the whole network, for AMD's VitisAI EP.

Every folder has a LICENSE.md naming the upstream weights, their licence, and where that licence was read. manifest.json lists every file with its sha256 and its evidence.

How these were checked

Each .ignite here was compiled by ignite-xdna at commit 394fc3d, with the flags in manifest.json. Each was then run on a Phoenix NPU (Ryzen 7 8700G) with tools/verify_engine_container.py. That check compares every layer tensor still readable after one dispatch with the integer reference built from the ONNX next to it, byte for byte. All seven read EXACT:

Container	Readable layers exact	Reused slots correct
yolov8n_full	28 / 28	38 / 38
yolov8s	28 / 28	38 / 38
yolov8n_pose	37 / 37	38 / 38
yolo11n	31 / 31	60 / 60
yolow	29 / 29	41 / 41
sesr_m7	5 / 5	4 / 4
resnet50_head	1 / 1	—

The log is results/hf/verify_published_containers_desktop2_20261013.log. Workspace reuse means a later layer overwrites some layers' tensors, so this checks the readable ones plus every reused slot's final contents; it is not all-layers coverage. modnet_cut_dense.ignite is the 2026-09-19 container from results/dense/compile_modnet_20260919.log, which records this ONNX's sha256. The verifier cannot lower dense host regions, so it was not re-run for this upload.

Accuracy

All on Desktop 2 (Ryzen 7 8700G, Phoenix NPU). COCO protocol: conf 0.001, IoU 0.7, max_det 300, per-class NMS. "AMD's stack" is the same ONNX through Ryzen AI 1.7.1's VitisAI EP.

Model	Metric	Engine container	AMD's stack	Dataset	Log
YOLOv8n	mAP@50-95 / @50	34.12 / 48.78	26.68 / 39.73	COCO val2017, 5,000	results/aie/silu_sigmoid_vs_amd/accuracy/eval_yolov8n_sigmoid_5000.log, eval_yolov8n_amd_5000.log
YOLOv8s	mAP@50-95 / @50	42.37 / 58.70	37.31 / 53.08	COCO val2017, 5,000	.../eval_yolov8s_sigmoid_5000.log, eval_yolov8s_amd_5000.log
YOLOv8n-pose	OKS mAP@50-95 / @50	44.16 / 76.39	32.64 / 66.90	COCO val2017, 5,000	.../eval_pose_sigmoid_5000.log, eval_pose_amd_5000.log
YOLO11n	mAP@50-95 / @50	34.63 / 49.59 ¹	25.82 / 38.76	COCO val2017, 5,000	results/aie/epilogue_yolo11n_yolow_20260921.log; AMD: results/eval_yolo11n_cut_xint8_npu.log
SESR M7 (x2)	PSNR-Y, dB	34.25 Set5 / 29.41 Set14 ²	34.06 / 29.32	Set5, Set14	results/aie/sesr_m7_bf16_silicon_desktop2_20260922.log; AMD: results/eval_sesr_m7_set{5,14}_npu.log

¹ Copied into that log as a summary table of the full-set run. The raw pycocotools output is not in the log. ² Identical to the same XINT8 model on ONNX Runtime CPU (34.2492 / 29.4098), as an exact engine should be.

The engine reads higher than AMD's stack on the YOLO models mainly because of --silu-sigmoid. AMD's XINT8 model replaces each SiLU's sigmoid with a straight-line HardSigmoid, and the flag computes a closer integer sigmoid on the NPU instead. Without the flag the containers read like AMD's stack (YOLOv8n 27.10, YOLOv8s 37.21, pose 32.71, YOLO11n 25.80).

Which build was measured. These figures were measured on earlier compiles of these same ONNX files with the same flags (YOLOv8n/s and pose at ignite-xdna 7700316 on 2026-09-17, YOLO11n on 2026-09-21), not on the files uploaded here. The uploaded files were checked layer-exact against the engine's integer reference (above). But the compiler's IR and reference code changed between the two builds (7700316..394fc3d), and the mAP was not re-run on the uploaded files. That is why this card carries no model-index metadata: Hugging Face would present these figures as measurements of these files, and they are not.

Other families, and why they are not in the table:

• ResNet50 (resnet50_xint8_c64.onnx on AMD's VitisAI EP): top-1 72.10 / top-5 88.10 on the repo's 1,000-image ImageNet-val evaluation set, not the full 50,000 (results/res/run_resnet50_npu.log). resnet50_head.ignite has no accuracy figure.
• YOLO-World: only a 300-image slice exists (24.7 mAP@50-95, results/aie/yolow_gptq/eval_gptqcv2_ignite300.log), and slices of this model read higher than the full set.
• MODNet: not a benchmark-standard matting set.
  • Labelled: against COCO person masks on all 2,693 val2017 images with a person, the engine scores 0.1609 person IoU, matching the XINT8 CPU reference to every digit. AMD's stack scores 0.1700 and FP32 scores 0.5030. XINT8 itself costs the model most of its quality, and neither stack wins (results/dense/acc_modnet_cut_*_20260921.log).
  • Agreement with FP32 on 50 unlabelled images: mean alpha MAD 0.1486 for the engine, 0.1672 for AMD's stack (results/dense/agreement_summary_20260921.log). On those 50 images the engine equalled ONNX Runtime CPU every time, and AMD's stack never did.

Latency

This card makes no latency claim. The engine's timings depend on the compiler version, the power mode and the machine, and they drift between sessions on the same machine. ignite-xdna's docs/BENCHMARKS.md and docs/MODEL_ZOO_BENCHMARKS.md give every measured figure with its machine, build and log. Those figures were measured on earlier compiles of these same ONNX files, not on these exact containers.

Use

Requirements: an XDNA1 NPU on Windows with AMD's NPU driver and XRT, and ignite-xdna installed in its mlir-aie-iron environment (see its docs/SETUP.md).

from ignite_xdna import InferenceSession

with InferenceSession.from_file("yolov8n/yolov8n_full.ignite") as s:
    heads = s.run_yolo_monolithic(input_data)      # every layer on the NPU
    preds = s.decode_yolo_predictions(heads)

To put everything where ignite-xdna's tests and pipelines expect it (models/ and build/), pinned and checked by sha256, run python tools/fetch_hf_artifacts.py from a checkout.

To rebuild a container from its ONNX:

ignite-compile --engine graph --input yolov8n/yolov8n_cut_xint8.onnx --output yolov8n_full.ignite --silu-sigmoid

Not here, and why

• BiSeNetV2: its weights are trained on Cityscapes, whose terms are non-commercial and restrict redistributing derivatives.
• YOLO26n: the only source ONNX on the build machine postdates its container, so the two could not be shown to pair.
• Upstream .pt, .pth and .ckpt weights: fetch them from their owners (Ultralytics, timm, AMD, ZHKKKe).
• Calibration and evaluation data (ImageNet, COCO, Cityscapes): never redistributed.

Measurement logs

Every number in ignite-xdna's documents is backed by a log in jdominick05/ignite-xdna-results.