The promise of reducing traffic congestion and human-made accidents has caught our interest in autonomous vehicles for a decade. While we are still far from having fully autonomous vehicles hit the road, the idea of having self-driving cars steer us towards a new futuristic normal is undeniable. However, according to a new study from MIT, it might come at a massive cost.
In a startling revelation, MIT researchers found that the energy needed in the future to power just the computers on a worldwide fleet of autonomous vehicles could produce as much greenhouse gas emissions as all of the data centers in the world right now. The researchers discovered this while investigating the potential energy usage and associated carbon emissions in the likelihood that autonomous vehicles will receive wide adoption in the future.
It is well known that data centers that contain physical computing infrastructure that power online applications have a huge carbon footprint. The International Energy Agency estimates that data centers used about 200 terawatt-hours (TWh), or close to 1% of the world’s electricity demand, and contributed 0.3% of global CO2 emissions in 2018.
Noting insufficient research devoted to studying the possible carbon footprint of autonomous vehicles, MIT researchers developed a statistical model to investigate the issue. According to their findings, 1 billion autonomous vehicles, each powered by a computer using 840 watts, would be using enough energy to produce almost the same amount of global data center emissions from 2018. As a point of reference, there are presently about 1.5 billion automobiles on the planet’s roadways.
The researchers also discovered that in over 90% of the modeled scenarios, EV computers would need to consume less than 1.2 kilowatts (kW) of computing power just to stay within the existing range (below 2018 levels) of data center emissions, which is something we just cannot achieve with present hardware efficiencies. For instance, a different statistical model that examines a scenario in which 95% of all vehicles are autonomous by 2050 and computing workloads double every 3 years reveals that for emissions to remain at the current levels, hardware efficiencies in automobiles would need to double every 1.1 years. Meanwhile, the Moore’s Law rate, which has been widely recognized in the industry for decades, indicates that computer power doubles approximately every two or more years. To make things worse, this rate is anticipated to slow down rather than accelerate eventually.
Soumya Sudhakar, lead MIT researcher on the study, said that though the findings are only projections, they should urge those working on self-driving cars to understand that doing things “as usual” is insufficient and that computer efficiency should be a top priority. This is crucial to minimize the emissions from computers onboard autonomous vehicles. Soumya adds, “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.”
The MIT team built a framework to study the operational emissions from the computers onboard a global fleet of autonomous electric vehicles. This model is dependent on several factors, including the total number of cars in the global fleet, the computing capacity of each computer on each vehicle, the hours driven by each vehicle, and the carbon intensity of the electricity that powers each computer. Because the MIT team is considering a future application that is not yet available, Soumya stated that each variable in the function equation involves a great deal of uncertainty. This was important because, according to some prior research, people might spend more time driving in autonomous vehicles since the hands-off steering wheel means multitasking! Autonomous vehicles could also encourage more younger and older drivers to get behind the wheel. At the same time, other research suggests that the amount of time spent driving can reduce as a result of algorithms discovering the fastest routes to destinations. Another problem is attempting to model for cutting-edge hardware and software technology that doesn’t yet exist.
As a result, researchers used a multitask deep neural network, a well-known method for autonomous cars, to mimic the workload of the algorithm. According to MIT, semi-autonomous vehicles currently use multitask-deep neural networks to navigate their surroundings by continuously receiving real-time data from several high-resolution cameras. Next, the research team explored several situations using the probabilistic model. They were surprised to learn how rapidly the workload of the algorithms increased.
According to one estimate, if a self-driving car employed ten deep neural networks to analyze video from 10 cameras for one hour of driving, it would produce 21.6 million inferences daily. Now consider how many inferences would be generated if one billion cars were used. 21.6 quadrillion conclusions! To put it into context, every Facebook data center in the world generates a few trillion inferences per day, reveals MIT. Imagine how energy-hungry autonomous vehicles are when put together!
Researchers suggested that to improve efficiency, engineers should create specialized hardware that would power navigation and perception tasks as well as run specific driving algorithms. However, Soumya noted that there is a problem with that, as cars frequently have lifespans of 10 to 20 years, and creating specific hardware now would create an additional challenge of making it “future-proof” so that it can support new algorithms.
The authors of the study suggested that researchers could also attempt to develop algorithms that are more effective and use less processing resources. However, that would mean compromising accuracy for effectiveness and possibly endangering vehicle safety.
After establishing this framework, the MIT teams intend to continue to investigate hardware efficiency and algorithm improvements in autonomous vehicles. Furthermore, they believe that characterizing embodied carbon from autonomous cars (i.e., the carbon emissions produced during the production of a car) and emissions from a vehicle’s sensors can improve their model.
The research, which has been published in IEEE Micro‘s January-February edition, was funded by the National Science Foundation and the MIT-Accenture Fellowship. The paper was co-authored by Soumya and her co-advisors, Sertac Karaman, an associate professor of aeronautics and astronautics and the director of the Laboratory for Information and Decision Systems (LIDS), and Vivienne Sze, an associate professor in the Department of Electrical Engineering and Computer Science (EECS) and a member of the Research Laboratory of Electronics (RLE).
Vivienne hopes that this research will motivate automakers to integrate emissions and carbon efficiency metrics into their designs.