Baidu Bets on FPGAs

Article By : Sally Ward-Foxton

Production-ready Automated Valet Parking platform now built on Xilinx devices, instead of GPUs.

Baidu’s in-vehicle computing platform for automated valet parking (AVP) is now production-ready. The company’s Apollo Computing Unit (ACU) is designed specifically for AVP, which allows drivers and passengers to be dropped off before the car drives itself away to look for a parking space.

Automated valet parking is seen as an important stepping stone to fully automated driving, especially since this more achievable goal could lead to substantial short-term revenues for businesses that are investing heavily in the technology. Meanwhile, fully automated vehicles are still years from generating revenue.

Baidu’s Apollo Computing Unit (ACU) is built on a Xilinx Zynq UltraScale FPGA (specifically, the XAZU5EV) which is used for sensor fusion and processing AI workloads. This is a change from the proof of concept system shown previously, which was built on GPUs.

“By providing an adaptable hardware platform, Baidu can design flexible and scalable systems that seamlessly incorporate AI compute acceleration and functional safety through diversity in processing engines,” said Yousef Khalilollahi, Xilinx’ vice president for core vertical markets, in a statement.

The AVP platform uses a combination of cameras and ultrasonic sensors for perception. Monocular cameras (120° field of view) look forwards and backwards. Four fisheye cameras (190° field of view) detail the area around the car. Twelve ultrasonic sensors perform a radar function around the vehicle: this includes eight ultrasonic park assist (UPA) sensors with a range of 2.5m and four automated parking assist (APA) sensors with a range of 5m.

The platform is compatible with Baidu’s PaddlePaddle framework and meets automotive safety integrity level C (ASIL-C).

Baidu already has an OEM partner lined up. Weltmeister Automobile, which despite its German name is a Chinese electric vehicle manufacturer, will see its first AVP-enabled vehicles roll off the production line in early 2020. Elsewhere in the industry, Baidu’s solution is expected to target vehicle models scheduled for mass production in 2021.