resnet-18-int8-ethos-u-executorch

ResNet-18 optimized for Arm-based Edge NPU

An INT8-quantized ResNet-18 image classifier, compiled with the Arm Vela compiler and exported to ExecuTorch for on-device inference on the Arm Ethos-U85 NPU.

Summary

This model is an Arm-optimized version of torchvision's ResNet-18, a widely used 18-layer deep residual network for image classification originally developed by Microsoft Research. It has been quantized to INT8 using static post-training quantization (per-tensor symmetric weights, per-tensor affine activations) and compiled with the Arm Vela compiler for deployment on the Arm Ethos-U85 NPU via ExecuTorch.

The optimized model runs on the Alif DK-E8 development board (Cortex-M55 host CPU + Ethos-U85 NPU, bare-metal). The exported graph has 98.36% static op-count delegation coverage (180/183 ops) to the NPU. It targets embedded and edge vision workloads where a full application processor and OS are not available.

Accuracy was validated on the full ImageNet-1k validation split, and performance was measured directly on the Alif DK-E8 target using ExecuTorch with the Ethos-U85 backend delegate.

Key results

Original model

Model files

Performance

Performance was measured on-device on the Alif DK-E8 target running the Arm Ethos-U85 NPU via ExecuTorch.

Reference configuration

Performance results

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

Accuracy results

A separate 2000-sample check on the Ethos-U85 Corstone-320 FVP emulator scored 69.10% Top-1 vs. 69.20% for the PyTorch-eager INT8 simulation on the same subsample (-0.10 pp), validating the PyTorch result as a reasonable 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

For this release, Arm used:

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

This is necessary because the .pte cannot run on a plain Python/ExecuTorch install. The script installs, under a local runtime/ directory, the Arm GNU toolchain, ExecuTorch 1.1.0 bare-metal runtime libraries, the Ethos-U85 semihosting runner (arm_executor_runner), and the Arm Corstone-320 FVP used by example.py.

Dependencies are declared in pyproject.toml. 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 prints predictions without overwriting the reference predictions.json. To save a separate result for comparison, run uv run example.py --output /tmp/resnet-18-predictions.json.

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

Expected output

Intended use

This model is intended for 1000-class image classification on embedded and edge devices equipped with an Arm Ethos-U85 NPU, such as the Alif DK-E8 development board, in bare-metal deployments without a full application OS.

Limitations

• 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 model inherits the limitations of the original ResNet-18 architecture (Microsoft Research design, as implemented in torchvision), including a fixed 224x224 input resolution and sensitivity to preprocessing (resize + center crop) on images with extreme aspect ratios or off-center subjects. INT8 quantization introduces a small accuracy reduction relative to the FP32 baseline. Performance figures are specific to the Alif DK-E8 (Cortex-M55 + Ethos-U85) reference configuration and may differ on other Ethos-U85 configurations or MAC counts.
• The exported .pte requires Ethos-U85 hardware or the Corstone-320 FVP and cannot run on the standard ExecuTorch CPU or XNNPACK runtime.
• The example requires Linux arm64 for the included FVP installer. It targets a fixed batch-one, 224×224 input and Ethos-U85-256; dynamic shapes and other Ethos-U configurations are not supported.
• The semihosting runner uses ExecuTorch's Sram_Only memory profile with a 16 MiB temporary allocator configured by the installer.

Additional notes

Quantization follows an a8w8 scheme as compiled by Arm Vela: weights are per-tensor symmetric (zero-point fixed at 0), while activations are per-tensor affine with a histogram-observer-derived zero-point. No layers were kept in FP32 — all layers are quantized. The Vela compiler targeted the ethos-u85-256 configuration with the Ethos_U85_SRAM_MRAM system configuration and Shared_Sram memory mode. The exported graph is delegated to the Ethos-U85 NPU, with the remaining ops running on the Cortex-M55 host. Flash deployment footprint: the 7,732,576-byte resnet18_ethosu_optimized.pte occupies 7.3744 MiB (23.04%) of the Alif DK-E8's 32 MiB OSPI1 flash, which has a theoretical bandwidth of 200 MB/s.

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.

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

About this version

Original Model: torchvision resnet18 by Microsoft Research (architecture); weights from torchvision's standard training recipe - 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.