The SAE (Society of Automotive Engineers) International has a scale that highlights different autonomous driving capabilities. At zero you’re doing all the work but at Level 2, which is where most systems are now, there’s Partial Automation of steering and acceleration/deceleration. The driver still needs to pay attention and intervene, as the focus is on highway scenarios, performing lane centring and cruise control.
Level 3 systems are on the roads too, this is Conditional Automation, where sensors on the car monitor the environment and take full control of the vehicle, with fallback requests on the driver. No dozing at the wheel here, even Level 4 doesn’t accommodate all driving modes, so how do we get to Level 5? The ultimate destination of Full Automation.
Suitably for motorists, there is a roadmap, although, according to Prakash Kartha, Director, Market Development at Intel, the journey will take until 2025 to be realised completely.
“The networks of today are not going to cut it, when it comes to self-driving”
But it’s not simply a matter of shovelling sensor data into the cloud and waiting for a response. Kartha describes the in-vehicle compute as, “rather like a second engine within our car” and that it will function as an on-board compute cluster that will handle AI navigation decisions on the fly, path planning and environment monitoring, along with sensor processing and fusion.
“Going forward we believe that you’re going to need data centre scale processors that are obviously power efficient (our folks are on that). Combined with FPGAs (field-programmable gate array), combined with accelerators for things like computer vision,” he says.
Successfully scaling these capabilities is going to be one of the major challenges as we shift gear from Level2/3 systems running at 0.5 to 10 teraflops to a fully loaded Level 5 system operating at 50 to 100 teraflops. But like Kartha says, “our folks are on that” and vehicles equipped with Intel® Xeon® processors coupled with FPGA functions, enabling rapid customisation, is part of the plan.
Meanwhile, back in the cloud, the data centre handles model training that forms the basis of vehicle’s model scoring (or inference), namely, how its machine learning and predictive algorithms respond in the wild. The data centre also performs endpoint management and analytics, but limits need to be observed, as Kartha explains: “There’s an expectation that there’s a certain number of self-driving cars – a fleet, basically – that each data centre can manage. There is a clear correlation between what data centre compute you need, versus the number of vehicles you can manage.”
A distributed architecture of vehicle, network and cloud is what Kartha has in mind to share the burdens of autonomous driving. Rather than conveying data from car to cloud and back, the network itself would provide a vital link in terms of V2X (vehicle-to-everything) cooperative driving, vehicle-to-vehicle communications (e.g. automated overtaking), real-time map downloading and security.
“We really see 5G as the glue that brings it all together,” says Kartha. “Intel is very conscious of the fact that you can’t do this all alone. That’s one of the reasons why we’re part of a bigger consortium of companies that form the 5G Automotive Alliance (5GAA) – BMW*, Audi*, Daimler*, Ericsson*, Huawei*, Nokia*, Qualcomm* and Intel. We see 5GAA as an accelerator for cellular V2X (vehicle to everything) moving forward.” Indeed, such a partnership looks like we’re in for quite a ride come 2025.
*Trademarks are the property of their owners