Arm/inception-v3-int8-ethos-u-executorch

Inception V3 optimized for Arm-based Edge NPU

An INT8-quantized version of Inception V3 for image classification, exported to ExecuTorch (.pte) and lowered to a single Ethos-U85 NPU partition for Arm-based Edge NPU devices such as the Alif DK-E8.

Summary

This repository contains an Arm-optimized version of inception_v3 for image classification, quantized to INT8 via static post-training quantization; weights are symmetric per-tensor, activations are asymmetric (affine) per-tensor. The model is provided in ExecuTorch (.pte) format, targeting Edge NPU systems.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on ImageNet-1k and measured performance on the Alif DK-E8.

Key results

Area	Result
Model format	ExecuTorch (.pte)
Target device class	Edge NPU
Reference device	Alif DK-E8 (Cortex-M55 + Ethos-U85-256 NPU, bare-metal)
Primary performance result	p50 latency 237.13 ms (4.22 inferences/s)
Accuracy result	Top-1 76.45%, Top-5 93.01%
Size / memory result	17.53 MB (5.93x smaller than the 103.94 MB FP32 state dict)

Original model

Field	Value
Original model	inception_v3
Original source	torchvision/models/inception.py
Original developer	Google
Original model card	torchvision.models.inception_v3
Original license	BSD-3-Clause

Model files

File	Description
inception_v3_ethosu85_optimized.pte	Arm-optimized model for deployment
config.yaml	Model I/O contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records
example.py	Corstone-320 FVP inference example
pyproject.toml	Pinned host-side dependencies and deployment metadata
uv.lock	Locked dependency versions for uv sync --frozen
install_ethos_u85_graviton.sh	Standalone installer that builds the Ethos-U85 execution stack (Arm GNU toolchain, ExecuTorch bare-metal runtime, arm_executor_runner, Corstone-320 FVP) required to run example.py. Installs into a shared, cross-model directory if sudo is available, otherwise falls back to a local install for this bundle only

Performance

Performance was measured on the reference configuration below.

Reference configuration

Field	Value
Device / platform	Alif DK-E8
CPU / accelerator	Cortex-M55 (1 core @ 400 MHz) + Ethos-U85 NPU (256 MACs/cycle @ 400 MHz)
OS	bare-metal
Runtime	ExecuTorch 1.1.0
Backend / delegate	Ethos-U85 (CMSIS-NN, Ethos-U-Driver)
Batch size	1
Precision	INT8, static PTQ — per-tensor quantization; symmetric weights, asymmetric (affine) activations
Runs	10 warmup + 100 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency	N/A	237.13 ms	N/A
p90 latency	N/A	237.13 ms	N/A
Inferences per second	N/A	4.22	N/A
Model size	103.94 MB	17.53 MB	5.93x smaller

The "Original / baseline" model size is the FP32 PyTorch state dict.

Because the Ethos-U85 NPU architecture natively processes integer workloads and does not support floating-point execution, the FP32 baseline cannot be compiled into an Ethos-U85 delegate .pte file, making NPU-accelerated baseline latency and throughput figures non-applicable.

The Graviton environment documented for example.py verifies portability of the inference flow but does not provide or imply equivalent Graviton performance.

Accuracy

Accuracy was evaluated on the ImageNet-1k validation split, comparing the FP32 baseline against the Arm-optimized INT8 model via a PyTorch FP32-vs-PT2E INT8 simulation, not on the exported .pte running on Ethos-U85.

Evaluation setup

Field	Value
Dataset	ImageNet-1k
Split	val
Number of samples	50000
Metric(s)	Top-1 accuracy, Top-5 accuracy
Evaluation runtime	PyTorch (FP32 vs. PT2E INT8)

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
Top-1 accuracy (%)	77.30	76.45	-0.85 pp
Top-5 accuracy (%)	93.45	93.01	-0.44 pp

A separate 2000-sample check on the Ethos-U85 Corstone-320 FVP emulator scored 76.35% Top-1 vs. 76.60% for the PyTorch-eager INT8 simulation on the same subsample (-0.25 pp), validating the PyTorch result as a proxy for on-device accuracy.

Accuracy was measured using the described evaluation setup. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Captured with torch.export, lowered to a single Ethos-U85 NPU partition via to_edge_transform_and_lower with Vela compilation for the ethos-u85-256 target, then exported to ExecuTorch .pte
Quantization	Yes	INT8 static PTQ (prepare_pt2e / convert_pt2e) — per-tensor quantization; symmetric weights, asymmetric (affine) activations; calibrated on 1,000 randomly sampled ImageNet-1k images
Runtime/backend selection	Yes	Ethos-U85 NPU delegate (CMSIS-NN, Ethos-U-Driver), targeting the ethos-u85-256 configuration
Graph/runtime compatibility updates	Yes	Performed as part of the Ethos-U85 export pipeline
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

On an AWS Graviton (Linux arm64) host, run the standalone installer to build the Ethos-U85 execution stack that the example needs:

The installer uses a shared /opt/ethos-u85-graviton runtime when sudo is available, reusing existing builds across model bundles. Without sudo, it automatically falls back to a local runtime/ installation. A custom location can be set with ETHOS_RUNTIME_DIR, and existing installations are detected before rebuilding.

bash install_ethos_u85_graviton.sh

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

The default run only prints predictions. To write a result without replacing the committed reference, pass an explicit destination:

uv run example.py --output /tmp/inception-v3-predictions.json

The reference predictions.json records the known-good result for the committed sample_input.jpg. The example uses the FVP to validate the Ethos-U85 runtime path; it is not a Python-native inference path for this PTE.

Note: The Python/uv example runs on AWS Graviton (Ubuntu arm64) to confirm runtime compatibility only and is intended as a guideline for building an equivalent run script for Arm-based edge systems equipped with an Ethos-U85 NPU.

Expected input

Property	Value
Input shape	[1, 3, 299, 299]
Input type	float32
Input range	[0.0, 1.0]
Preprocessing	resize shorter edge to 342, center crop to 299x299, to tensor, normalize (mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225])

Expected output

Property	Value
Output shape	[1, 1000]
Output type	Raw class logits (unnormalized), ImageNet-1k class order
Postprocessing	softmax, top-5

Intended use

This model is intended for developers evaluating image classification workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on ImageNet-1k val and may not generalize to all domains. The reported accuracy is based on a PyTorch quantized-vs-FP32 simulation pass, not a measurement of the exported .pte running on the Corstone-320 FVP or on Ethos-U85 silicon.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.
• The exported .pte has almost its entire compute graph (910 of 913 operators) delegated to a single Ethos-U85 NPU partition and cannot run on the standard ExecuTorch CPU or XNNPACK runtime — it requires real Ethos-U85 hardware (e.g. an Alif DK-E8 board) or Arm's Corstone-320 FVP.
• The example requires Linux arm64 for the included FVP installer. It targets a fixed batch-one, 299×299 input and Ethos-U85-256; it does not support dynamic shapes or other Ethos-U configurations.
• The semihosting runner uses ExecuTorch's Sram_Only memory profile with a 16 MiB temporary allocator configured by the installer.

Additional notes

The 237.13 ms latency and related PMU/SRAM figures were measured directly on physical Alif DK-E8 hardware.

The TorchVision implementation used here for loading and export is distributed under BSD-3-Clause.

Flash deployment configuration: the on-device benchmark loaded the 18,385,168-byte inception_v3_ethosu85_optimized.pte from the Alif DK-E8's OSPI1 flash at 0xc0000000. OSPI1 was configured as an 8-line Octal SPI DDR interface at 100 MHz, using XIP (linear memory-mapped) reads. The artifact occupies 17.5335 MiB (54.7921%) of the board's 32 MiB OSPI1 flash. Note: 200 MB/s is the theoretical raw interface bandwidth; measured flash bandwidth was not collected.

install_ethos_u85_graviton.sh invokes ExecuTorch's Arm setup with --i-agree-to-the-contained-eula, accepting the Arm Corstone-320 FVP EULA. Review those terms before running it.

When the FVP is run with its --fast option, it reports that timing adapters are disabled and may print Failed to initialize timing-adapter. This affects cycle-accurate timing simulation, not functional inference; the validated run completed one Ethos-U85 delegation and produced the expected output tensor.

Sample input: sample_input.jpg is derived from Samoyed on Beach by Appleinfl, via Wikimedia Commons (public domain).

About this version

Original Model: inception_v3 by Google - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to BSD-3-Clause.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.