Skip to main content
Edge AI (Edge Artificial Intelligence) means that artificial intelligence operations are performed on the device where the data is collected (at the edge), not in the cloud. That is, data is processed at the endpoint such as a sensor, camera, drone, IoT device; it is not constantly sent to the cloud for analysis. The aim is to reduce latency, increase privacy, and reduce network traffic. In artificial intelligence and image processing projects, large amounts of data are used, and high processing power is required. The AM67A processor contains MMA (Matrix Multiply Accelerator), a special hardware unit to accelerate such computations. This unit processes image data and AI models quickly. MMA works connected to the processor and lightens the load on the main CPU, making computations more efficient. The board’s processor also works together with the C7x Digital Signal Processor (DSP) to perform different types of calculations (tensor, vector, scalar) with high speed and low energy consumption. MMA has its own memory and data transfer system, so data is processed quickly, and energy savings are achieved. In this section of the documentation, we will explain the concept of “AI Capability” on the Development Board, discuss which tools will be used, where, and for what purpose, and describe example projects. The fundamental AI capability and usage logic on the development board is based on the execution of various pre-trained models in an accelerated manner using the MMU/C7x modules to perform tasks, allowing the CPU/MCU to access these results. It contains 2 Deep Learning Accelerators and has a processing power of 4 TOPS (Trillion Operations Per Second). The table below shows the estimated processing power required for which technologies.
Use CaseApproximate Processing Power
Object detection (e.g., YOLOv5s)1-2 TOPS
Face recognition1 TOPS
Anomaly detection (with vibration data)0.5 TOPS
License plate recognition2 TOPS
Voice command processing1 TOPS
Optical character recognition (OCR)1.5-3 TOPS