Arm/deit-tiny-int8-ethos-u-executorch

DeiT-Tiny optimized for Arm-based Ethos-U NPU

DeiT-Tiny image classification optimized as a quantized ExecuTorch .pte model for Arm-based Ethos-U NPU systems (Ethos-U85). Most quantized layers use INT8 activations and INT8 weights (a8w8), while selected output.dense layers use INT16 activations and INT8 weights (a16w8) to preserve accuracy within acceptable ranges. Quantization is per-tensor throughout: weights use symmetric quantization (zero-point fixed at 0) across all layers, while a8w8 activations use an asymmetric scheme (affine, histogram-observer-derived zero-point); the a16w8-overridden output.dense activations are symmetric (zero-point fixed at 0), still histogram-observer-derived.

Summary

This repository contains an Arm-optimized version of facebook/deit-tiny-patch16-224 for image classification. The model is provided in ExecuTorch .pte format, targeting Arm-based Ethos-U NPU systems.

This version demonstrates efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm evaluated the model on ImageNet-1k and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ExecuTorch .pte
Target device class	Ethos-U NPU
Reference device	Alif DK-E8 (Cortex-M55 + Ethos-U85, bare-metal)
Primary performance result	81.59 ms p50 latency (12.26 inferences/s)
Accuracy result	Top-1 67.95% / Top-5 88.47% (PyTorch quantized simulation, not measured on-device — see Limitations)
Size / memory result	4.71 MB (.pte), 4.65x smaller than FP32

Original model

Field	Value
Original model	facebook/deit-tiny-patch16-224
Original source	Hugging Face
Original developer	Facebook AI Research (Meta)
Original model card	facebook/deit-tiny-patch16-224
Original license	Apache-2.0

Model files

File	Description
deit-tiny_ethos_ethosu_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	Minimal inference example
pyproject.toml	Pinned runtime dependencies and deployment metadata
uv.lock	Locked dependency versions, sources, and hashes 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	CMSIS-NN, Ethos-U-Driver
Batch size	1
Precision	Mixed quantized precision: primarily a8w8 (INT8 activations / INT8 weights); output.dense uses a16w8 (INT16 activations / INT8 weights) for accuracy recovery
Quantization method	Static PTQ, per-tensor; symmetric weights, activations asymmetric except where widened to a16w8, with SmoothQuant activation smoothing
Runs	10 warmup + 100 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency	N/A	81.59 ms	N/A
p90 latency	N/A	81.59 ms	N/A
Model size	21.89 MB	4.71 MB	4.65x 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	50,000
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	72.13%	67.95%	-4.18 pp
Top-5 Accuracy	91.13%	88.47%	-2.66 pp

A separate 2,000-sample check compared the PT2E-quantized model executed in PyTorch eager mode against the same model exported and run on the Ethos-U85 Corstone-320 FVP: Top-1 67.90% (PyTorch) vs. 68.30% (FVP), a +0.40 pp difference — 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

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	Converted to ExecuTorch .pte and ahead-of-time compiled with Arm's Vela compiler for the Ethos-U85-256 target
Quantization	Yes	Static PTQ, per-tensor; symmetric weights throughout, activations asymmetric (a8w8) except symmetric on the a16w8-overridden output.dense layers. SmoothQuant activation smoothing was applied before observer insertion.; calibrated on 2,000 random ImageNet-1k images
Runtime/backend selection	Yes	CMSIS-NN + Ethos-U-Driver via the ExecuTorch Ethos-U delegate
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

Most quantized layers use INT8 activations and INT8 weights (a8w8). Selected output.dense layers use higher activation precision (a16w8) to preserve accuracy within acceptable ranges.

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

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 — it requires Ethos-U85 hardware or the Corstone-320 FVP (see Limitations). The script installs the Arm GNU toolchain, the ExecuTorch 1.1.0 bare-metal runtime libraries (built from source), the Ethos-U85 semihosting runner (arm_executor_runner), and the Arm Corstone-320 FVP that example.py uses to simulate the Cortex-M55 + Ethos-U85 target.

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 prints predictions without overwriting the reference predictions.json. To save a separate result for comparison, pass an explicit destination:

uv run example.py --output /tmp/deit-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, 224, 224]
Input type	float32
Input range	[0.0, 1.0]
Preprocessing	Resize (shorter side 256, bicubic) → center crop to 224x224 → 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	float32 (raw class logits, unnormalized, ImageNet-1k class ordering)
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 .pte requires Ethos-U85 hardware or the Corstone-320 FVP to execute; it cannot run on a standard desktop/CPU ExecuTorch Python runtime because the Ethos-U backend delegate is not registered there.
• The model uses a fixed input size of 224x224 and batch size 1; dynamic shapes 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

• SmoothQuant activation smoothing was applied before observer insertion to reduce activation outliers ahead of static PT2E quantization. Most quantized layers use INT8 activations and INT8 weights (a8w8), while output.dense (the MLP-down projection in each transformer block) uses INT16 activations and INT8 weights (a16w8) to preserve accuracy within acceptable ranges.
• Flash deployment configuration: the on-device benchmark loaded the 4,935,568-byte .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 4.7069 MiB (14.71%) of the board's 32 MiB OSPI1 flash. Note: 200 MB/s is the theoretical raw interface bandwidth; measured flash bandwidth was not collected.
• EULA notice: install_ethos_u85_graviton.sh installs the Arm Corstone-320 Fixed Virtual Platform (FVP) by invoking ExecuTorch's upstream examples/arm/setup.sh with --i-agree-to-the-contained-eula. Running the installer accepts Arm's End User License Agreement (EULA) for the Fixed Virtual Platforms on your behalf. Review Arm's FVP EULA terms before running the script if you have not already agreed to them.
• Sample input: sample_input.jpg is derived from Samoyed on Beach by Appleinfl, via Wikimedia Commons (public domain).

About this version

Original Model: facebook/deit-tiny-patch16-224 by Facebook AI Research (Meta) - 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 Apache-2.0.

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.