The dark side of self-driving vehicles
Many visions of the future envision self-driving electric vehicles that will whizz through cities picking up passengers. But is it possible? ?
MIT researchers have found that the energy required for running computers in autonomous cars worldwide could cause as much greenhouse gas emissions as all of the data centres worldwide.
This research revealed how much computing is required to keep billions upon billions of self driving vehicles on the roads – up to (1 quadrillion = 1,000 trillion).
According to the International Energy Agency, data centres account for 0.3 percent of global greenhouse gases emissions. That’s about the same amount of carbon that Argentina produces each year.
Researchers at MIT created a statistical model in order to study the problem.
They found that 1,000,000 autonomous vehicles could drive for just one hour per day using a computer consuming 840 Watts. That would produce enough energy to emit about the same amount of carbon dioxide as all current data centres.
The researchers suggest that faster technological advancements may be necessary to reduce the impact.
Soumya Sudhakar is a graduate student of aeronautics and astronautics. She says that if we keep the current trends in decarbonisation and the current rate at which hardware efficiency improvements are being made, it doesn’t seem like it will be enough to limit the emissions from computing onboard autonomous vehicle.
“This has the potential to become an enormous problem. But if we get ahead of it, we could design more efficient autonomous vehicles that have a smaller carbon footprint from the start.”
Researchers created a framework to examine the operational emissions of computers onboard a global fleet electric vehicles that are fully self-driving, which means they don’t need a human driver.
The model is dependent on the number of vehicles in the global vehicle fleet, the power of each car’s computer, the driving hours and the carbon intensity of electricity used to power each computer.
This equation looks simple enough to be deceptive. Sudhakar says that each variable contains a lot more uncertainty as we are looking at an emerging application.
Sudhakar was amazed at the speed with which the algorithms were able to handle different scenarios when they used the probabilistic approach.
An example: If an autonomous vehicle has 10 deep neural network processing images from 10 camera lenses, and it drives for 1 hour per day, it will make 21.6million inferences daily.
1 billion vehicles would yield 21.6 quadrillion conclusions.
This is a small example of the number of trillions of inferences that all Facebook data centres around the world make each day.
Sertac Karaman is an associate professor of astronautics and aeronautics.This makes sense after seeing the results. However, it is not something people are likely to be interested in. These vehicles could be running on a lot of computer power. They see the world from every angle. We may only have two eyes but they might have 20.
Watch: Car makers confront challenges with humanoid bots