Generate and validate faces on the DevKit

Open the application console

Keep SW4 set to UART4. In Visual Studio Code (VS Code), open Serial Monitor on the serial port exposed through PRG USB and use the following settings:

• Baud rate: 115200
• Data bits: 8
• Parity: None
• Stop bits: 1

If the port was open before you changed SW4, close and reopen it.

Build and debug the application

To build and debug the application:

1. In the CMSIS view, confirm that DevKit-E8 is active.
2. Select Build, then select Debug. Keil Studio starts the J-Link GDB server over Serial Wire Debug (SWD) and loads the application into magnetoresistive random-access memory (MRAM).
3. The debugger stops at main. Press F5 to continue.

The firmware initializes the display and NPU, loads the two ExecuTorch methods, and generates a repeatable boot image. The output is similar to:

    

        
        Ethos-U version info:
    Arch:       v2.0.0
    MACs/cc:    256
ExecuTorch pico-faces (m3_long_cfg, a16w8/a8w8)
Methods:
  dit_step  2 input(s), 1 output(s)
  decode    1 input(s), 1 output(s)
Generating: seed 3, 4 steps, class 1, w 4.0
Image: 128x128x3, CRC32 <export-dependent-value>
Test_result: PASS

        
    

The LCD displays the 128 x 128 result scaled to 384 x 384 pixels in the center of the screen. The Cyclic Redundancy Check (CRC) depends on the host-side export, so use Test_result: PASS and the displayed image as the functional checks.

Confirm NPU execution

Inspect the two performance lines printed after generation. They report NPU cycles, NPU active time, multiply-accumulate array activity, and AXI reads for dit_step and decode. These counters come from the Ethos-U performance monitoring unit and confirm that both methods executed on the NPU.

The following is a reference DevKit run of the source project:

These instructions were verified on a Mac using the following versions:

Treat the timing as a reference measurement rather than a pass threshold. Tool versions, the exported program, and board conditions can change the result.

Generate more faces

Use the SW2 joystick after the boot image appears:

• Press toward the left to generate one face with the next seed.
• Press toward the right to start continuous generation. Press in the same direction again to stop.

Each completed joystick generation updates the LCD and prints the following to the console:

• Image number
• Seed
• Class
• Guidance
• Total time

The boot demo prints a CRC. Joystick log lines don’t include one.

To check standalone boot, end the debug session and press RESET (SW1) while the application is idle. The boot image and Test_result: PASS should return.

View an example result

The following image shows a generated face displayed on the DevKit LCD:

Generated face displayed on the Alif Ensemble E8 DevKit

Troubleshoot common problems

Use the following checks if the application doesn’t reach the pass marker:

(Optional) Use the Corstone-320 FVP

The same CMSIS solution also includes the SSE-320-U85 target for the Corstone-320 Fixed Virtual Platform (FVP). On Linux and Windows, the managed tools environment supplies the native FVP. Before selecting Run or Debug, select SSE-320-U85 in Manage Solution, select Apply, and recreate the AI layer.

On macOS, the project runs the FVP in Docker. To use the FVP, follow these additional steps:

1. Install Docker Desktop for Mac and keep it running.

2. In cmsis-executorch.csolution.yml, find the debugger entry under the target-set of the SSE-320-U85 target. Replace its model value with the Docker wrapper for the project:

       
   
           
           
   model: "${workspaceFolder}/.vscode/fvp.sh"
   
       
   

3. Select SSE-320-U85 in Manage Solution and select Apply. After the extension generates the debug configuration, open a terminal in the CMSIS-Executorch directory and copy the macOS configuration:

       
   
           
           
   cp .vscode/launch.json.mac .vscode/launch.json
   
       
   

4. Recreate the AI layer, build, and select Run or Debug. The first invocation builds the Docker image and downloads the simulator, so it takes longer than subsequent runs.

The FVP writes the raw image to out/fvp_image.bin. To return to the board, select DevKit-E8, select Apply, and recreate its AI layer before building again.

What you’ve accomplished

You’ve exported a generative model through ExecuTorch, loaded it on the Alif Ensemble E8 DevKit, and generated faces with both neural-network methods delegated to Ethos-U85. You also validated the run through the LCD, pass marker, and hardware performance counters.

You can extend the Learning Path workflow to deploy other Generative AI models on a microcontroller with Arm Ethos-U85 acceleration.