Das digitale Auto wird zum vierten Bildschirm und zur nächsten Computerplattform. Eine intelligente Datenplattform ermöglicht dabei die Verarbeitung der anfallenden Daten und die Etablierung neuer datengetriebener Geschäftsmodelle.
Connected cars and autonomous driving technologies are driving incredible value for the automotive industry and have attracted significant investments to “auto tech” in both the private and public markets. Concomitant with this growth is a fundamental shift in the data platforms used to support connected car applications because of their unique requirements. If we look at connected cars and autonomous driving as the next killer app, we can learn from the platform shifts that they are precipitating as harbingers of even greater movements to come. In particular, the data platforms used to support connected car applications must support:
Real-time processing at scale. Increasingly sophisticated advanced driver-assistance systems (ADAS) need to process large volumes of data onboard the vehicle in real time to provide features such as collision avoidance, automatic braking and adaptive cruise control. The need is compounded as the car achieves greater levels of autonomous driving capabilities. An autonomous driving car gains a holistic understanding of the vehicle’s position and circumstances by combining multiple sensor outputs from devices including radar (10-100 KB/s), sonar (10-100 KB/s), GPS (50 KB/s), cameras (20-40 MB/s) and lidar (10-70 MB/s). In total, about 4 TB of data are generated and processed onboard the vehicle for every autonomous driving hour. The data platform, therefore, needs to support true real-time data processing and decision making (e.g., braking or accelerating).
Machine and deep learning. While some of the systems onboard the vehicle utilize human-curated rules that help the vehicle make decisions quickly while on the road, there is an increasing emphasis on using machine learning and deep learning to make better decisions in real time. For example, pedestrian detection is difficult to implement using a rules-based system; instead, cars use deep learning models that do semantic segmentation of real-time dashboard-mounted camera feeds in order to detect pedestrians. This shift toward using machine learning requires the use of emerging software frameworks — like Caffe2 or TensorFlow — and there will likely be many more new entrants to come. Moreover, the process of training and deploying machine learning models has precipitated new, iterative development processes that require massive volumes of training data and the close collaboration between data scientists, application developers, data engineers and governance professionals. Data platforms supporting these applications need to support an incredibly broad variety of processing engines and data types and need to facilitate a complex application development process with as little friction as possible.
Distributed computing. With increasing computational capabilities onboard the car itself coupled with internet connectivity, the modern car is the ultimate edge processing device. In addition to the real-time ADAS functions onboard the vehicle, the car sends relevant summary information a centralized fleet management application in a data center or cloud where it is aggregated across many vehicles in order to analyze fleet performance and to anticipate maintenance issues. In many instances, the data movement between the car and the data center/cloud must be bidirectional so that machine learning models can be rescored and improved over time through experiments in the data center and can be seamlessly redeployed to the vehicles. Vehicle-to-vehicle functionality will require further that the cars communicate in a peer-to-peer network that supports omnidirectional data movement. While the 2000s might have been termed the cloud era, the shift we are seeing today exemplified by connected cars is indicative of the rapid growth in total processing that is being done outside of any physical data center or public cloud environment. Consider that, compared to the 80% CAGR of computation onboard a car, industry estimates from IDC, Gartner and Wikibon put the growth of public cloud computing between 16%-19% CAGR. The data platforms that support connected car applications have to be infrastructure agnostic and have to support continuous, coordinated data flows — seamlessly moving data and compute between the data center and/or cloud and the vehicles.
