CAMES

A tremendous investment in automotive embedded systems has been observed in recent years. And the trend continues to grow considering the emerging self-driving vehicles [2,6]. Those vehicles are equipped with complex sub-systems like multi-view camera system, lane-keeping assistant system, pedestrian avoidance system, etc. The systems are parts of the embedded system in the vehicle which contains hundreds of endpoints (cameras, millimeter-wave radars, GPS, brake system, etc.) and form a sophisticated network [12]. The state-of-art network of automotive embedded systems tends to be deployed in a distributed manner, where various sub-systems are divided into domains [6,12]. In each domain, a Domain Control Unit (DCU) is responsible for computing tasks. The DCU can be seen as a SoC which contains a CPU and sometimes a FPGA/GPU [5] for remote acceleration. The endpoints within/between each domain are connected via specialized network/bus like LIN, CAN or FlexRay, which provides deterministic network in the loops between sensors, DCUs and actuators. However, the emerging trend of self-driving in automotive brings at least three new challenges that the current systems find hard to address. 1) high-bandwidth and deterministic interconnection. Currently, 60-100 endpoints (sensors/actuators) are deployed per vehicle. But recent industry figures suggest the number to reach over 200 per vehicle [1]. In addition, inter-domain traffic grows rapidly as domains are increasingly integrated. The ever-scaling system desires lightweight and high-bandwidth network rather than sophisticated and low-bandwidth buses. Moreover, deterministic features should also be preserved for real-time and security-relatedapplications.2) high-performance computing for AI-powered inference real-time AI models (e.g., convolutional neural network) has been widely adopted for autopilot functionalities [2,3]. However, low-end computing platforms [5] have difficulties deploying large scale AI models [2,9]. For instance, Tesla requires >50 Tops neural network performance while maintaining the power under40W per chip (cooling reasons). This requires high-end computing node being implemented. Considering the budget and cooling constraints in automotive systems, deploying those computing nodes in a distributed manner would be infeasible.3) Over-the-air(OTA) updates for automotive software/firmware. Due to the fast evolution of software/firmware on vehicles, a number of manufacturers tend to update software/firmware via OTA techniques. This requires DCUs to be networked with the cloud from time totime. The distributed positioning increases the complexity of the remote updates and may incur security issues [6]. To address the challenges above, we present CAMES, a TSN-enabled centralized architecture for automotive embedded systems. In the center of CAMES, the central computing platform takes the advantage of high-end FPGAs for remote acceleration. DCUscan be largely simplified as no power-hungry computing tasks running locally. Between DCUs and the centralized computing platform, TSN (time-sensitive network) [11] is implemented to provide high-bandwidth and deterministic L2 connection for both real-time and non-real-time traffic. On top of TSN, we proposed a light weight transport layer protocol for data delivering between sensors/actuators and the central computing platform. Since software/firmware of DCUs/sensors/actuators are updated by the central computing platform instead of exposing directly to the remote, security and simplicity feature can be better maintained during OTA. This demo demonstrates the advantages and feasibility of such a design with two use cases which are typical in automotive scenarios.