Rate this Page
★ ★ ★ ★ ★

Zephyr: MobileNetV2 with Ethos-U NPU (Corstone FVP & Alif E8)#

Run a quantized MobileNetV2 image classifier on the Alif Ensemble E8 DevKit using ExecuTorch, Zephyr RTOS, and the Arm Ethos-U55 NPU. The same build flow also works on the Arm Corstone-320 FVP for development without hardware.

What You’ll Build#

  • A quantized INT8 MobileNetV2 model fully delegated to the Ethos-U55 NPU (110 ops, ~19 ms inference on Alif E8)

  • A Zephyr RTOS application that loads the .pte model, runs inference on a static test image, and prints the top-5 ImageNet predictions over UART

Prerequisites#

Hardware (choose one)#

Target

Description

Alif Ensemble E8 DevKit

Cortex-M55 HP core + Ethos-U55 (256 MACs), 4.5 MB HP SRAM, MRAM

Corstone-320 FVP

Virtual platform simulating Cortex-M85 + Ethos-U85 (no hardware needed, Linux only)

Software#

  • Linux x86_64 (FVP and Arm toolchain are Linux-only; macOS can export models but cannot run the FVP or flash)

  • Python 3.10+

  • Alif SE Tools for flashing (Alif hardware only)

Step 1: Set Up the Zephyr Workspace#

Create a workspace, install west, and initialize the Zephyr tree:

mkdir ~/zephyr_workspace && cd ~/zephyr_workspace
python3 -m venv .venv && source .venv/bin/activate
pip install west "cmake<4.0.0" pyelftools ninja jsonschema
west init --manifest-rev v4.3.0

Install the Zephyr SDK (compiler toolchain):

wget https://github.com/zephyrproject-rtos/sdk-ng/releases/download/v0.17.4/zephyr-sdk-0.17.4_linux-x86_64.tar.xz
tar -xf zephyr-sdk-0.17.4_linux-x86_64.tar.xz && rm -f zephyr-sdk-0.17.4_linux-x86_64.tar.xz
./zephyr-sdk-0.17.4/setup.sh -c -t arm-zephyr-eabi
export ZEPHYR_SDK_INSTALL_DIR=$(realpath ./zephyr-sdk-0.17.4)

Step 2: Add ExecuTorch as a Zephyr Module#

Copy the submanifest, configure west to pull only the modules we need, and update:

mkdir -p zephyr/submanifests
cat > zephyr/submanifests/executorch.yaml << 'EOF'
manifest:
  projects:
    - name: executorch
      url: https://github.com/pytorch/executorch
      revision: main
      path: modules/lib/executorch
EOF

west config manifest.project-filter -- -.*,+zephyr,+executorch,+cmsis,+cmsis_6,+cmsis-nn,+hal_ethos_u
west update

For Alif boards, also add the Alif HAL:

west config manifest.project-filter -- -.*,+zephyr,+executorch,+cmsis,+cmsis_6,+cmsis-nn,+hal_ethos_u,+hal_alif
west update

Step 3: Install ExecuTorch and Arm Tools#

cd modules/lib/executorch
git submodule sync && git submodule update --init --recursive
./install_executorch.sh
cd ../../..

Install the Arm toolchain, Vela compiler, and Corstone FVPs:

modules/lib/executorch/examples/arm/setup.sh --i-agree-to-the-contained-eula
source modules/lib/executorch/examples/arm/arm-scratch/setup_path.sh

Step 4: Export the MobileNetV2 Model#

Export a quantized INT8 MobileNetV2 with Ethos-U delegation. Choose the target that matches your hardware:

For Alif E8 (Ethos-U55 with 256 MACs):

python -m modules.lib.executorch.backends.arm.scripts.aot_arm_compiler \
    --model_name=mv2_untrained \
    --quantize --delegate \
    --target=ethos-u55-256 \
    --output=mv2_ethosu.pte

For Corstone-320 FVP (Ethos-U85 with 256 MACs):

python -m modules.lib.executorch.backends.arm.scripts.aot_arm_compiler \
    --model_name=mv2_untrained \
    --quantize --delegate \
    --target=ethos-u85-256 \
    --output=mv2_u85_256.pte

The --delegate flag routes all compatible ops through the Ethos-U backend. The Vela compiler converts the TOSA intermediate representation into an optimized command stream for the NPU. Use mv2 instead of mv2_untrained for meaningful predictions (requires torchvision pretrained weights).

Step 5: Build the Zephyr Application#

For Alif E8:

west build -b alif_e8_dk/ae822fa0e5597xx0/rtss_hp \
    -S ethos-u55-enable \
    modules/lib/executorch/zephyr/samples/mv2-ethosu -- \
    -DET_PTE_FILE_PATH=mv2_ethosu.pte

For Corstone-320 FVP:

west build -b mps4/corstone320/fvp \
    modules/lib/executorch/zephyr/samples/mv2-ethosu -- \
    -DET_PTE_FILE_PATH=mv2_u85_256.pte

Step 6a: Run on Corstone-320 FVP#

Set up the FVP paths and run:

export FVP_ROOT=$PWD/modules/lib/executorch/examples/arm/arm-scratch/FVP-corstone320
export ARMFVP_BIN_PATH=${FVP_ROOT}/models/Linux64_GCC-9.3
export LD_LIBRARY_PATH=${FVP_ROOT}/python/lib:${ARMFVP_BIN_PATH}:${LD_LIBRARY_PATH}
export ARMFVP_EXTRA_FLAGS="-C mps4_board.uart0.shutdown_on_eot=1 -C mps4_board.subsystem.ethosu.num_macs=256"

west build -t run

MV2 inference is cycle-accurate on the FVP and takes 10-20 minutes of wall clock. You should see output like:

========================================
ExecuTorch MobileNetV2 Classification Demo
========================================
Ethos-U backend registered successfully
Model loaded, has 1 methods
Inference completed in <N> ms
--- Classification Results ---
Top-5 predictions:
  [1] class <id>: <score>
  ...
MobileNetV2 Demo Complete
========================================

Step 6b: Flash and Run on Alif E8#

Flash with Alif SE Tools#

Use the Alif SE Tools to program the binary into the E8’s MRAM. Create a zephyr.json in the build output directory:

cat > build/zephyr/zephyr.json << 'EOF'
{
    "HP_img_class": {
        "binary": "zephyr.bin",
        "version": "1.0.0",
        "mramAddress": "0x80008000",
        "cpu_id": "M55_HP",
        "flags": ["boot"],
        "signed": false
    }
}
EOF

Important: Use mramAddress: "0x80008000" (FLASH_LOAD_OFFSET=0x8000), not the default 0x80200000. The default offset does not leave enough MRAM for the ~3.5 MB MV2 model blob.

Note: Your SE Tools install may include additional zephyr.json entries (e.g., a DEVICE block referencing app-device-config.json). Copy those from the SE Tools template if your board configuration requires them.

Generate the table of contents and flash using the SE Tools:

cd build/zephyr
python <path-to-alif-se-tools>/app-gen-toc.py
python <path-to-alif-se-tools>/app-write-mram.py
cd ../..

Refer to the Alif SE Tools documentation for installation and detailed usage.

Connect Serial Console#

Connect to UART4 at 115200 baud. On Linux:

picocom -b 115200 /dev/ttyUSB0

Press the reset button on the E8 DevKit. You should see:

Booting Zephyr OS build ff8b8697c0f5 ***

========================================
ExecuTorch MobileNetV2 Classification Demo
========================================

I [executorch:main.cpp] Ethos-U backend registered successfully
I [executorch:main.cpp] Model PTE at 0x8004b290, Size: 3490912 bytes
I [executorch:main.cpp] Model loaded, has 1 methods
I [executorch:main.cpp] Running method: forward
I [executorch:main.cpp] Method allocator pool size: 1572864 bytes.
I [executorch:main.cpp] Setting up planned buffer 0, size 752640.
I [executorch:main.cpp] Loading method...
I [executorch:main.cpp] Method 'forward' loaded successfully
I [executorch:main.cpp] Preparing input: static RGB image (150528 bytes)
I [executorch:main.cpp]
--- Starting inference ---
I [executorch:main.cpp] Inference completed in 19 ms
I [executorch:main.cpp]
--- Classification Results ---
I [executorch:main.cpp] Top-5 predictions:
I [executorch:main.cpp]   [1] class 0: 0.0000
I [executorch:main.cpp]   [2] class 1: 0.0000
I [executorch:main.cpp]   [3] class 2: 0.0000
I [executorch:main.cpp]   [4] class 3: 0.0000
I [executorch:main.cpp]   [5] class 4: 0.0000
I [executorch:main.cpp]
========================================
I [executorch:main.cpp] MobileNetV2 Demo Complete
I [executorch:main.cpp] Model size: 3490912 bytes
I [executorch:main.cpp] Input: 224x224x3 RGB image (150528 bytes)
I [executorch:main.cpp] Output: 1000 ImageNet classes (top-5 shown)
I [executorch:main.cpp] Inference time: 19 ms
I [executorch:main.cpp] ========================================

All predictions show 0.0000 because mv2_untrained has random weights. Use mv2 (with torchvision pretrained weights) for meaningful class scores.

Troubleshooting#

Symptom

Cause

Fix

Linker: region 'FLASH' overflowed

Model PTE too large for ITCM

Use the DDR overlay (FVP) or verify mramAddress (Alif)

Linker: region 'RAM' overflowed

Pools + model copy exceed SRAM

Set CONFIG_ET_ARM_MODEL_PTE_DMA_ACCESSIBLE=y to skip the SRAM copy

FVP hangs after “Ethos-U backend registered”

Cycle-accurate MV2 simulation is slow

Wait 10-20 min, or use Corstone-320 (faster than 300)

No serial output on Alif

Wrong UART or baud rate

Use UART4 at 115200 baud

app-write-mram.py fails

Wrong mramAddress

Use 0x80008000, not 0x80200000

Runtime: method allocator OOM

Pool size too small

Increase CONFIG_EXECUTORCH_METHOD_ALLOCATOR_POOL_SIZE in the Zephyr config used for the build (for example prj.conf, an OVERLAY_CONFIG, or a board .conf)

Memory Layout#

Region

Corstone-320 FVP

Alif E8

Code + .rodata

ITCM (512 KB)

MRAM

.data + .bss + pools

ISRAM (4 MB)

HP SRAM (4.5 MB)

Model PTE (~3.5 MB)

DDR (16 MB, via overlay)

MRAM (DMA-accessible)

NPU delegation

Ethos-U85 (256 MACs)

Ethos-U55 (256 MACs)

Using Claude Code with Zephyr#

If you use Claude Code, the ExecuTorch repo ships a /zephyr skill that can help with:

  • Workspace setup — scaffolds the Zephyr workspace, west manifests, and SDK install

  • Board bringup — generates DTS overlays, board confs, and linker snippets for new boards

  • Memory debugging — diagnoses linker overflow errors and runtime allocation failures, with the exact pool sizes your model needs

Type /zephyr in Claude Code while working in the ExecuTorch repo to activate it. Related skills: /export for model conversion, /cortex-m for baremetal Cortex-M builds, /executorch-kb for backend-specific debugging.

Next Steps#

  • Swap mv2_untrained for mv2 (with torchvision) to get real ImageNet predictions

  • Try other models: resnet18, or bring your own .py model file

  • Explore the hello-executorch sample for a minimal starting point

  • See the Ethos-U Getting Started tutorial for the baremetal (non-Zephyr) flow

On this page
Edit on GitHub
Show Source