New AI Embedded Accelerator Unveiled for Robotics at ISSCC 2022

Renesas, a leading semiconductor company, recently introduced a groundbreaking AI accelerator designed for real-time processing of lightweight algorithms in robots. The company showcased this cutting-edge technology at the prestigious International Solid-State Circuits Conference held in San Francisco.

The AI accelerator developed by Renesas is specifically tailored to meet the demands of modern robots. It combines conventional processing capabilities for planning and control with AI processing power of approximately 100Top/s peak for vision-based environmental recognition. Remarkably, the accelerator achieves this high performance while maintaining low power dissipation, eliminating the need for additional cooling mechanisms such as a fan. The power consumption of the 14nm integrated circuit stands at around 5W.

One of the key features of Renesas' AI accelerator is its heterogeneous microprocessor architecture, which includes a dynamically reconfigurable component dedicated to AI processing. This accelerator can deliver an impressive 130Top/s at an efficiency of 23.9Top/s/W when operating at 0.8V with INT8 (8-bit integer) data.

According to Renesas, the performance of the AI accelerator was evaluated using various AI models, including ResNet50, YOLOV2, and deeplabV3. The company reported achieving efficiency levels ranging from 9 to 11Top/s/W with these models. The processor's unique architecture enables it to dynamically prune calculations involving zero values in the weight matrix, a technique known as 'N:M pruning.' This approach significantly reduces the number of computations required for sparse matrices, resulting in 80-90% fewer calculations compared to traditional methods.

The processor boasts 216 processing elements that can be reconfigured within a single clock cycle, allowing for optimization at each stage of a multi-step algorithm. Renesas demonstrated the effectiveness of this technology in applications such as SLAM (Simultaneous Localization and Mapping), showcasing a substantial increase in operational speed and power efficiency compared to using the embedded CPU alone.

Looking ahead, Renesas plans to integrate this innovative technology into its RZ/V series of microprocessors tailored for vision applications. The AI accelerator's impressive performance metrics and energy efficiency make it a promising solution for real-time robot processing, setting a new standard in the field of artificial intelligence and robotics.