A new MPU optimized for high-end embedded Vision AI

As the use of embedded Vision AI is expanding, there is a need to deliver higher performance combined with greater power efficiency, and to set new benchmarks in many applications.

For offices, Vision AI will ?be used for human, face and pose recognition, and people counting will deliver a boost for security applications. Object recognition will be used in retail and factory applications, and the combination of face and pose recognition will support service robots viewing an area and doing real-time motor control and processing. Object and pose recognition will be combined in appliances, and drones will use vision AI for segmentation.

The trend is moving from traditional classification, through object recognition, and onto semantic segmentation which gives each pixel in an image a class. In addition, pose estimation will predict and track the location of a person or object, and visual simultaneous localization and mapping (vSLAM) will demand higher performance.

So what are the key challenges for designers looking to develop Vision AI?

The increase in computation needed to deliver Vision AI draws more power and generates additional heat. But the system must also aim for a small size and low cost, so adding a heat-sink or fan will be a major disadvantage. Speed is important too, so the toolchain must be easy to use for AI beginners as well as for AI experts.

To address these challenges, the Renesas quad-core RZ/V2H MPU for Vision AI integrates a DRP-AI3 accelerator and a high-performance real-time processor. This integration enables the RZ/V2H to deliver AI inference performance of up to 80 TOPS, together with real-time processing, and without the need for a cooling fan or heat-sink.

The new-generation AI Accelerator increases power efficiency, and the INT8 Dense model doubles the power efficiency of the previous generation of dynamically reconfigurable processors for AI. The INT8 Sparse AI model delivers 5x the power efficiency with pruning, and without degrading recognition accuracy.

The performance of AI inference from the RZ/V2H is around twice that of competitor MPUs, and the OpenCV Accelerator delivers up to 16.5x more performance compared to a 1.8 GHz quad CPU.

The RZ/V2H MPU also simplifies design by integrating an ISP camera up to 4k, serial servo control, motor control and high-speed communication interfaces.

The easy-to-use toolchain includes the RTK0EF0168C04000BJ evaluation board, and an AI SDK Linux development environment. Designers can also access AI applications and the e2studio GUI plug-in AI Navigator.

Altogether, the performance and power efficiency achieved by the RZ/V2H MPU Vision AI microprocessor drives the expansion of embedded high-end Vision AI into many new applications.

?

Request the free RTK0EF0168C04000BJ board to evaluate the RZ/V2H Vision AI MPU.