8 Platform one February/March 2024 | Uncrewed Systems Technology Ambarella has developed a modular software stack for autonomous vehicles that runs on a range of different hardware, including its latest chip (writes Nick Flaherty). The stack has been demonstrated on the new low-power CV3-AD chip, providing self-driving capabilities up to SAE Level 3, while also running on other hardware. The stack runs entirely on the CV3-AD chip, built in a 5 nm CMOS process to provide environmental perception and route planning to create HD maps in real time from SD resolution data. The stack is carefully optimised to minimise the processing load and the need for external memory accesses to keep the power consumption of the chip under 30 W, compared with 300 W for other chips for autonomous driving. It means the electronic control unit does not have to be liquid-cooled. The chip has up to eight ARM Cortex A78AE CPU cores with a dual-core, lockstep pair of Cortex-R52 cores, which support functional safety to the ASIL-B standard. There are up to six custom AI accelerator cores that support 8 bit and 4 bit data in the neural networks. It also includes a H-264 codec for video conversion from long-range, ultra-high-definition cameras and HD short-range with stereo redundancy, as well as the 4D imaging radar developed by Oculii, which is part of Ambarella. The CVflow software stack includes route planning, and it has been widely tested for corner cases such as narrow roads, construction and areas with a high density of pedestrians. A key element of the hardware design is that there are spare cycles on the ARM processors in the chip for additional software and the proprietary networks. “We are using SD maps and generated HD maps in real time for path planning,” said Chris Day, vice-president of business development at Ambarella. “As developers have control of all the data from the raw data, they can decide on the best place to solve the problem – for example, in the ISP and not in the algorithmic part – so the image is better and the algorithm can be simpler. “This leads to better management of the overall signal chain. This was an epiphany for a software developer.” Deep-learning algorithms have been used in all the modules and developed with the chip in mind. This requires support from the backend tools. An automated annotation pipeline automates the training of the sensor fusion, and there is a PCIe card for development and validation of the software. The CV3 has been used by Applied Intuition for a closed-loop implementation in a system-in-the-loop (SIL) simulation system that is porting proprietary neural networks from other customers. Both the chip and the software stack have been benchmarked with the Resnet34 neural network, and they show a 5.7x increase in performance over the previous generation, along with a 4.3x increase in frames-per-second per watt to boost power efficiency. Autonomous vehicles Cooler stacks and chips Providing self-driving capabilities up to SAE Level 3 The CVflow software stack includes route planning
RkJQdWJsaXNoZXIy MjI2Mzk4