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 with Ethos-U NPU delegation

  • 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), MRAM, 8 MiB SRAM

Corstone-320 FVP

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

Software#

  • Linux x86_64 for the workflow below

  • Python 3.12+

  • 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 v4.4.0

Step 2: Add ExecuTorch as a Zephyr Module#

Create the submanifest, configure west to pull the modules for this demo, and update. The project filter below is for the new workspace created in Step 1; when using an existing application, retain the modules that application needs.

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
west packages pip --install

Zephyr 4.4.0 includes the Alif SoC and peripheral support used by this sample in its own source tree.

Install the Zephyr SDK (compiler toolchain):

west sdk install --gnu-toolchains arm-zephyr-eabi

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 (HP Ethos-U55 with 256 MACs):

python -m executorch.backends.arm.scripts.aot_arm_compiler \
    --model_name=mv2 \
    --quantize --delegate \
    --target=ethos-u55-256 \
    --output=mv2_ethosu.pte

If rtss_he is used instead of rtss_hp below use --target=ethos-u55-128 to match the hardware.

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

python -m executorch.backends.arm.scripts.aot_arm_compiler \
    --model_name=mv2 \
    --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_untrained instead of mv2 if you do not want to depend on torchvision pretrained weights.

These export commands retain floating-point model inputs and outputs. The Cortex-M runs the runtime, input/output processing, and boundary quantization and dequantization. An NPU operator count from Vela describes the compiled partition; it does not count all operations in the ExecuTorch program.

By default, this command calibrates using the model’s example input, which is random for mv2. For classification accuracy evaluation, supply representative calibration data and compare against PyTorch with matching input preprocessing.

Step 5: Build the Zephyr Application#

The separate build directories keep the Alif and FVP board configurations and firmware images independent. The -d option is optional when building only one target; West otherwise uses build.

For Alif E8:

west build -d build-alif -b ensemble_e8_dk/ae822fa0e5597ls0/rtss_hp \
    modules/lib/executorch/zephyr/samples/mv2-ethosu -- \
    -DET_PTE_FILE_PATH=mv2_ethosu.pte

For Corstone-320 FVP:

Use the environment from Step 3: setup_path.sh supplies the FVP executable and shared-library paths. Zephyr finds FVP_Corstone_SSE-320 on PATH and uses the board’s simulator configuration. Set the sample’s NPU MAC count and simulation limit before configuring the build; Zephyr captures ARMFVP_EXTRA_FLAGS when CMake runs.

export ARMFVP_EXTRA_FLAGS="--simlimit 10 -C mps4_board.subsystem.ethosu.num_macs=256"

west build -d build-fvp -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#

Run the image built in Step 5:

west build -d build-fvp -t run

The Ethos-U model can report NPU cycle counts, while CPU instruction timing and memory-system behavior are simplified. The <N> ms line below measures method->execute() with Zephyr’s simulated clock, rather than reading the NPU cycle counter. It does not establish Alif E8 hardware latency. Host wall time depends on the host and simulator configuration.

The sample leaves Zephyr running after printing MobileNetV2 Demo Complete. --simlimit 10 stops the FVP after ten seconds of simulated time. Confirm that the completion message appears before it exits. If you change the simulation limit or other FVP flags, rerun west build -d build-fvp -c before running again.

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 west#

west flash -d build-alif

Flash with Alif SE Tools#

If you do not want to use J-Link to write a raw image, or if west flash does not work on your board, use Alif SE Tools to program the E8 MRAM through the Secure Enclave Application Table of Contents (ATOC) flow.

Use a SE Tools release that matches the SES firmware printed on the SE UART at reset. For example, if the boot log says SES A0 v1.109.0, download and use the 1.109 SE Tools directory.

The Alif E8 sample build generates the binary consumed by SE Tools:

  • build-alif/zephyr/zephyr.bin

Create build/images/zephyr.json and build/images/app-device-config.json in the SE Tools directory:

ZEPHYR_BUILD=~/zephyr_workspace/build-alif/zephyr

cd <PATH_TO_ALIF_SETOOLS>/app-release-exec-linux_FW_1.109.00_DEV

ln -sf "$ZEPHYR_BUILD/zephyr.bin" build/images/zephyr.bin

cat > build/images/zephyr.json << 'EOF'
{
    "DEVICE": {
        "binary": "app-device-config.json",
        "version": "0.5.0",
        "signed": true
    },
    "MV2": {
        "binary": "zephyr.bin",
        "version": "1.0.0",
        "mramAddress": "0x80008000",
        "cpu_id": "M55_HP",
        "flags": ["boot"],
        "signed": false
    }
}
EOF

cat > build/images/app-device-config.json << 'EOF'
{
  "metadata": {
    "device": "AE822FA0E5597LS0",
    "project": "",
    "external_clock_sources": [
      {
        "id": "OSC_HFXO",
        "enabled": true,
        "frequency": 38400000
      },
      {
        "id": "OSC_LFXO",
        "enabled": false,
        "frequency": 32768
      }
    ]
  },
  "firewall": {"firewall_components": []},
  "pinmux": {"configurations": []},
  "miscellaneous": []
}
EOF

Make sure cpu_id matches your target: use M55_HP for ensemble_e8_dk/ae822fa0e5597ls0/rtss_hp and M55_HE for ensemble_e8_dk/ae822fa0e5597ls0/rtss_he.

zephyr.json places the image at 0x80008000, matching the board config’s CONFIG_FLASH_LOAD_OFFSET=0x8000:

Important: Keep mramAddress: "0x80008000" and CONFIG_FLASH_LOAD_OFFSET=0x8000 in sync. If these differ, SES may accept the ATOC but the M55 image will not start correctly.

Generate the ATOC package from the SE Tools directory. The executable releases expect image files under their own build/images directory, so stage the Zephyr outputs there with symlinks:

./app-gen-toc -f "build/images/zephyr.json"

Note that some versions use app-gen-toc.py.

This creates build/AppTocPackage.bin in the SE Tools directory.

The DK-E8 routes the J-Link VCOM to different UARTs using SW4:

  • Set SW4=SE for SE Tools and the Secure Enclave boot log at 57600 baud.

  • Set SW4=U2 for the Zephyr RTSS-HE console at 115200 baud.

  • Set SW4=U4 for the Zephyr RTSS-HP console at 115200 baud.

To flash set SW4=SE and run

./app-write-mram

Note that some versions use app-write-mram.py.

On some boards the running application prevents the ISP handshake. In that case, enter hard maintenance mode first. Keep SW4=SE, close any serial terminal, and use the same baud rate as the SE boot log (57600 on DK-E8):

./maintenance -c /dev/ttyACM0 -b 57600

Select:

1 - Device Control
1 - Hard maintenance mode

When the tool prints Waiting for Target..[RESET Platform], press the physical reset button on the board. Wait for the tool to return to the menu.

Verify that hard maintenance really succeeded before flashing. From the returned menu, select:

Enter - return to the top-level menu
2 - Device Information
4 - Device enquiry

Continue only if the device enquiry reports Maintenance Mode = Enabled. The maintenance tool may still show menus after Device did not respond; that does not mean the target is connected. If device enquiry also reports Target did not respond, power-cycle the board, keep SW4=SE, press reset, and repeat the hard maintenance step.

After Maintenance Mode = Enabled, exit the maintenance tool and flash without resetting out of maintenance mode:

./app-write-mram -b 57600 -nr -s

If prompted for the serial port, enter the J-Link VCOM port, for example /dev/ttyACM0.

You can verify the flash and boot if needed with SW4 set to SE, connected to the SE boot UART at 57600 baud:

picocom -b 57600 /dev/ttyACM0

Press the reset button on the E8 DevKit. The package has been accepted and booted when you see ATOC ok and a table row for M55-HP or M55-HE whose flags include B.

Connect Serial Console#

After flashing, make sure SW4 is set to U4 for RTSS-HP or U2 for RTSS-HE, open the serial terminal, and press reset:

picocom -b 115200 /dev/ttyACM0

Press the reset button on the E8 DevKit. A representative log is shown below; timing, model size, and scores vary with the model and software/hardware configuration.

*** Booting Zephyr OS build v4.4.0 ***

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

I [executorch:main.cpp] Ethos-U backend registered successfully
I [executorch:main.cpp] Model PTE at 0x800774a0, Size: 3380576 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] Input tensor: scalar_type=Float, numel=150528, nbytes=602112
I [executorch:main.cpp] Converting uint8 input (150528 elements) to float32
I [executorch:main.cpp]
--- Starting inference ---
I [executorch:main.cpp] Inference completed in 30 ms
I [executorch:main.cpp]
--- Classification Results ---
Top-5 predictions:
I [executorch:main.cpp]   [1] class 92: 3.2362
I [executorch:main.cpp]   [2] class 21: 3.0362
I [executorch:main.cpp]   [3] class 127: 2.8180
I [executorch:main.cpp]   [4] class 22: 2.5998
I [executorch:main.cpp]   [5] class 18: 2.5089
I [executorch:main.cpp]
========================================
I [executorch:main.cpp] MobileNetV2 Demo Complete
I [executorch:main.cpp] Model size: 3380576 bytes
I [executorch:main.cpp] Input: [1, 3, 224, 224] NCHW RGB tensor (150528 bytes)
I [executorch:main.cpp] Output: 1000 classes (top-5 shown)
I [executorch:main.cpp] Inference time: 30 ms
I [executorch:main.cpp] ========================================

Use mv2_untrained instead of mv2 if you want to avoid downloading pretrained weights. In that case the class scores are arbitrary and may all be close to 0.0000.

The reported inference time covers one method->execute() call, measured with k_uptime_get_32(). Model loading, input preparation, and result reporting are outside that interval. For hardware comparisons, record the software versions, model, core/NPU configuration and clocks, and measurements over repeated runs.

Troubleshooting#

Symptom

Cause

Fix

Linker: region 'FLASH' overflowed

Code and model exceed the linked image region

Check the linker map and model placement: DDR on FVP, MRAM on Alif. Changing the ATOC address does not increase linker capacity.

Linker: region 'RAM' overflowed

Pools + model copy exceed SRAM

Set CONFIG_ET_ARM_MODEL_PTE_DMA_ACCESSIBLE=y to skip the SRAM copy

No Zephyr startup banner on FVP

Incorrect boot or memory layout

Check the ELF addresses and use the Corstone-320 sample overlay with separate regions for code and writable data.

FVP exits before the demo completes

Simulation limit too short, configuration mismatch, or runtime failure

Inspect the console, check the exported NPU target and MAC count, and increase --simlimit if needed. Reconfigure after changing FVP flags.

No Zephyr serial output on Alif

SW4 still routes VCOM to SE UART

Move SW4 to U4 and use 115200 baud

No SE Tools response

Serial terminal is open, wrong SW4 setting, or app blocks ISP

Close terminal, set SW4=SE, or enter hard maintenance mode

app-write-mram fails or app does not boot

mramAddress and link address differ

Use 0x80008000, not 0x80200000 with CONFIG_FLASH_LOAD_OFFSET=0x8000

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

First 512 KiB of ISRAM at 0x31000000; boot vectors in ITCM

MRAM

.data + .bss + pools

Remaining 3.5 MiB of ISRAM at 0x31080000

Combined SRAM0/SRAM1, 8 MiB at 0x02000000

Model PTE

DDR region at 0x70000000 (16 MiB, via overlay)

MRAM (DMA-accessible)

NPU delegation

Ethos-U85 (256 MACs)

Ethos-U55 (256 MACs)

These are memory-region capacities. The method and temporary allocator pools reserve 1.5 MiB each; the linker’s RAM usage also includes other data and stacks. Use the build’s memory summary and linker map to inspect reserved space. Neither region capacity nor statically reserved RAM measures peak live allocation.

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#

On this page
Edit on GitHub
Show Source