There’s no consensus on how best to use processing power at the edge of the network to support connected cars and autonomous vehicles (AVs). However, advanced autos are a natural use case for edge computing. Self-driving autos would consume and generate large amounts of data, rely on complex decision-making, and would be mobile, often taking vehicles far from a central data center.
Most AVs are expected to use a mix of onboard and remote decision-making. Running applications close to where data is generated can reduce round trips to the cloud, slashing end-to-end latency and easing the transmission burden on core network links.
Multi-Access Edge Computing (MEC) is one of the architectural changes behind the scenes that will complement the 5G mobile network’s advanced radio infrastructure. Others, including network slicing and virtualization of software and hardware in carrier facilities, will help make edge computing possible.
Some mobile operators have already implemented edge computing on 4G. For AT&T, positioning servers near the edge has slashed a typical 100ms roundtrip latency on LTE to 20ms, said Cameron Coursey, VP/CTO of IoT at AT&T. With 5G, AT&T wants to achieve 10ms or less to support automated driving. Verizon, Deutsche Telekom and other operators also plan to deploy MEC.
One example of an autonomous driving application that could use edge computing is automated traffic management, said Martin Beltrop, head of Nokia’s mobile networks automotive business. An application would analyze real-time data about the location, direction and speed of all the cars, pedestrians and other connected road users in a given area — up to 1,000 objects in a busy intersection. The software would use that data to build an object model of the road users, including their location and direction, which could be used to help AVs get through the area safely and efficiently.
How close is the edge?
Achieving the approximately 10ms latency needed for this service would require both 5G and an “edge cloud” located closer than the core of the carrier’s network, Beltrop said. That wouldn’t necessarily mean servers mounted on cell towers. With fiber backhaul from nearby basestations transmitting data at the speed of light, the edge cloud could be as far as 10km away, he said. In a well-architected 5G network, it could reside at the first or second aggregation point behind the basestation.
At those distances, there could be one edge cloud hosting the traffic management computing for an entire city, Beltrop said. The software would run analysis of each intersection separately but on the same computing platform. Nokia learned this in part from trials of such a data center in Munich, where it achieved a highly reliable latency of 10ms to 20ms over an LTE network, he said.
Ericsson has looked at the same problem in a slightly different way but also concluded that edge computing would eventually be necessary.
Self-driving applications can get adequate latency without edge computing, up to a certain point, said Claes Herlitz, head of Global Automotive Services. But once the number of connected cars reaches a certain level — possibly 100 million vehicles in the US — automated traffic management will become such a big problem it will need to be broken up into smaller, local chunks, he said.
Instead of one central cloud-based platform modeling traffic for all intersections, servers near each intersection would monitor and analyze only the local traffic and direct AVs in that area. As far as latency is concerned, Ericsson has already achieved 15ms roundtrip latency for AV applications with LTE and doesn’t see a need for anything lower than that, Herlitz said.
Edge computing for automotive is likely to begin in dense urban areas, running on servers shared with enterprise applications such as banking, said Maxime Flament CTO of the 5G Automotive Association (5GAA), an industry group supporting automotive uses of 5G.
There are, of course, several hurdles to overcome before edge computing can play an active role in autonomous driving at a large scale.
One area of concern is the sharing of edge computing infrastructure, which for cost reasons may be owned by one operator but used by several, Flament said. Also, edge computing may make handoffs between carriers at national borders more difficult. Instead of just maintaining a subscriber’s signal, it will be necessary to maintain edge application performance for safety. While these problems may have been solved technically, there are still unresolved organizational and governance issues, Flament said.
Edge computing will also complicate the issue of coverage, said analyst Philip Marshall of Tolaga Research. To ensure ongoing services for moving cars, there may have to be both overlapping cells and redundant edge computing infrastructure. Making sure the application and the information are in the same place, and that the car is continuously served, will require a coordination scheme — and all self-driving applications will need to be written to work with that scheme, he said.
“The orchestration is a challenge we haven’t solved yet at this level of performance and at scale,” Marshall said.
Given these hurdles, edge computing for AV applications probably won’t be built out until 2024 or 2025, he said. However, that timing matches common expectations for when fully baked 5G infrastructure for autonomous driving will come online. And as self-driving cars themselves emerge slowly through tests and trials, they might take at least that long to hit the road in significant numbers.