Edge AI
Ahead of exhibiting for our first time at Embedded World, we wanted to create a demo that would attract interest to our stand and show off our expertise. As we recently became a Toradex Proven Partner, we felt it appropriate to base our work on their i.MX95 Verdin Evaluation Kit. We set out to demonstrate practical use of the eIQ Neutron Neural Processing Unit (NPU) and show how to make efficient use of processing across its heterogeneous platform.
Our Work
We created a ‘Doodle Translator’, a touchscreen-based device that allows you to draw a doodle and see real-time classification results (it tells you what you’ve drawn). To achieve this we used Tensorflow lite with a custom convolutional neural network trained on Google’s Quick, Draw! dataset, accelerated by the NPU. The user interface was built with Python’s pygame.
To highlight energy-efficient design, we suspended the Cortex-A application cores when the device wasn’t in use. An ultrasonic sensor running on an otherwise idle Cortex-M core detected presence of people, waking the A cores only when needed. This approach reduced overall power consumption and showed effective partitioning of workloads between heterogeneous cores.
Services Performed
- Distribution Development: Yocto on Toradex i.MX 95 Verdin Evaluation Kit / Toradex 10.1” LVDS Capacitive Display
- Application Development: Tensorflow Lite, custom convolutional neural network trained on Google’s Quick, Draw! dataset, pygame
- Low Level: Customisation of Cortex-M core firmware
A Successful Outcome
The demo attracted lots of interest at Embedded World and demonstrated our capabilities in application development, hardware acceleration, AI, and low-level firmware on the i.MX95 platform. Watch the video below to see the demo in action.
Find out More
Reach out to us to find out how we can assist you.
Contact Us