High performance transformer-based AD perception software stack

Ambarella, Inc. and Plus have announced that Plus's PlusVision – a high-performance transformer-based AD perception software stack with industry-leading power efficiency – is now available on Ambarella's CV3-AD AI domain controller family of SoCs, which provides industry-leading AI performance per watt. This state-of-the-art perception offering will help Tier-1 suppliers and carmakers accelerate the launch of next-generation vehicles with L2+/L3 autonomy.

PlusVision is the transformer-based, deep neural network AI perception software that combines Plus's L4 perception capabilities with an extensive and diverse dataset. Its AD stack has accumulated millions of real-world driving miles in the U.S., Europe and Australia. This data collection has made software smarter, allowing incredibly information-rich output that helps OEMs develop and fine-tune detailed autonomous driving policies. Implementing uniform perception software is key to allowing these data-driven, custom policies across different vehicle models and trim levels.

"We are delighted to collaborate with Ambarella in offering PlusVision's leading-edge perception capabilities on their CV3-AD SoCs," said David Liu, CEO and co-founder at Plus. "This combination will provide Tier-1s and OEMs with a flexible, high-performance and power-efficient perception solution, enabling them to create a future-proof platform that can be adapted for their evolving autonomy needs."

"Our collaboration with Plus offers Tier-1s and OEMs additional cutting-edge transformer-based AI perception software solutions to help accelerate their next-generation vehicle development," said Fermi Wang, president and CEO of Ambarella. "This is a pivotal time for the automotive industry, and solutions like this give Tier-1s and OEMs the flexibility to efficiently differentiate their offerings and maximise customer value."

The stack, running on a mid-range CV3-AD655 SoC, delivers peak performance of 11 cameras being simultaneously processed at 30fps. This solution provides more than 200m of front-camera detection range, allowing the long-distance perception essential in high-speed driving conditions.

The CV3-AD SoC family provides industry-leading low power consumption while offering advanced image processing and superior AI performance per watt. The companies' teams collaborated to optimise the transformer-based PlusVision model for the CVflow AI engine inside the CV3-AD SoCs, which supplies an optimal solution for AD systems with high performance, low latency, and low-power AI processing. Also, the CV3-AD family's efficiency helps increase battery range in new EVs, and its optimised processing performance provides OEMs with extra headroom to integrate other software and create differentiated features.