50 GOPs performance consuming only 1-10 mW of power in a 1 mm2 die area.
Dual Arm Cortex-M55 cores: one high-performance (up to 400 MHz) for ML workloads and one efficient (150 MHz) for always-on, low-power tasks.
Innovative architecture eliminates reliance on external DRAM, reducing both cost and power consumption.
Leveraging Arm for High-Speed Image Processing
The Arm Cortex-M55 CPU enables high-speed vision processing with the “big” core running at up to 400MHz for ML workloads and the “LITTLE” core running at 150MHz for always-on tasks at very low power. This delivers significantly faster inferencing versus using just a microcontroller.
The chip uses Arm Helium instructions for image processing before AI inference, as well as image cropping, resizing, and color adjustment.
- Scalability: The Cortex-M55 is compatible with Arm ML frameworks like CMSIS-NN, enabling flexible software development. For example, the same code can run on Cortex-M55 or other Arm cores.
- Hardware integration: Integrating the processor with the Arm Ethos-U55 accelerator and vision IP enables tightly coupled processing between the CPU, hardware accelerators, and peripherals.
- Ecosystem support: Using the Cortex-M55 allows tapping into the broad ecosystem of Arm tools, software, and support. This was a key factor in Himax’s decision.
Optimized for Vision Applications at the Edge
The WE2 is optimized for running vision AI models at the edge. Himax has demonstrated models for:
- Face detection and landmarks: It goes beyond basic face detection to also perform facial landmark detection, identifying key facial structures like eyes, nose, and mouth.
- Human pose estimation: The WE2 can estimate the pose of the human body by detecting key joints and limbs. It identifies body parts like arms, legs, torso, and head to determine the overall body position and movements. This has applications in fitness tracking, gesture control, and identifying actions.
- Object detection: Object detection YOLO v8n models can run on the WE2 to detect and classify multiple objects in a scene. It can identify common objects like people, cars and car plates, animals, household items, parcels left at doors, and more across over 80 categories.
Designed for Flexibility
With support for external flash, the WE2 allows large AI models to be stored off-chip. The chip can interface with image sensors and Himax adopted the Arm ecosystem for software compatibility. Himax leverages open-source tooling — including Keil MDK, VS Code, Pytorch, and TinyNeuralNetwork — to enable users to develop custom AI applications optimized for low power. Himax provides reference models to help bootstrap development.
Discover the Real-World Impact of Edge AI
Explore in-depth use cases that show how edge AI is powering the next generation of IoT, solving real-world problems, driving faster decisions, and lowering costs through smarter operations right at the device level.