The energy needed to run the powerful computers in the global fleet of autonomous vehicles in the future could generate as many greenhouse gas emissions as all the data centers in the world today.
That's one of the main findings of a new study by MIT scientists examining the potential energy consumption and associated carbon emissions if self-driving cars become widespread.
The data centers that house the physical computing infrastructure used to run applications are notorious for their large carbon footprints: they currently account for around 0.3% of global greenhouse gas emissions, roughly the amount of carbon Argentina produces each year. year. International Energy Agency. Realizing that less attention has been paid to the potential footprint of autonomous vehicles, the MIT researchers built a statistical model to study the problem. They found that a billion self-driving cars driving for an hour a day, each with a computer drawing 840 watts of power, would consume enough energy to generate the same amount of emissions as data centers today.
The researchers also found that in more than 90% of the simulated scenarios, each car would need less than 1.2 kilowatts of computing power, requiring more efficient hardware to prevent autonomous vehicle emissions from exceeding current vehicle emissions. data center. In a scenario where 95% of the world's vehicle fleet is powered by autonomous vehicles by 2050, IT workloads double every three years, and the world continues to decarbonize at its current rate, they found that the efficiency of equipment is twice as fast as before should increase. . 1.1 years to keep emissions below these levels.
“If we simply keep up with decarbonization trends and the current pace of equipment efficiency improvements, it won't be enough to limit emissions from on-board computing in self-driving cars. it becomes a big problem. But if we move forward, we can develop more efficient autonomous vehicles with a lower carbon footprint from scratch," said first author Sumya Sudhakar, a doctor of aeronautics and astronautics.
Sudhakar wrote the paper with co-authors Vivian Sze, Associate Professor in the Department of Electrical and Computer Engineering (EECS) and member of the Research Laboratory of Electronics (RLE); and Sertak Karaman, Associate Professor of Aeronautics and Astronautics and Director of the Laboratory for Information and Decision Systems (LIDS). The study was published in the January-February issue of IEEE Micro .
Emissions modeling
Researchers have created a framework to study emissions from computers running on board a global fleet of fully autonomous electric vehicles, meaning they don't need a backup human driver.
The model depends on the number of vehicles in the global fleet, the power of each computer in each vehicle, the operating hours of each vehicle, and the carbon intensity of the electricity that powers each computer.
On its own, this sounds like a deceptively simple equation. But there's a lot of uncertainty in each of these variables because we're looking at a new application that doesn't exist yet,” says Sudhakar.
For example, some studies suggest that time spent driving autonomous vehicles may increase as people are able to multitask while driving and people young and old are able to drive more. But other research suggests that time spent driving may be reduced as algorithms can find optimal routes that get people to their destinations faster.
In addition to accounting for these uncertainties, the researchers had to model modern computer hardware and software that did not yet exist.
To do this, they modeled the workload of a popular algorithm for autonomous cars, known as a multi-site deep neural network, because it can handle multiple tasks simultaneously. They investigated how much power this deep neural network would consume if it were to simultaneously process multiple high-resolution inputs from multiple high-frame-rate cameras.
When he used probabilistic models to explore different scenarios, Sudhakar was surprised at how quickly the algorithms' workload increased.
For example, if an autonomous car has 10 deep neural networks processing images from 10 cameras and the car drives for one hour a day, it will produce 21.6 million results every day. One billion cars will make allocations of 21.6 quadrillion. For comparison, all of Facebook's data centers around the world generate several trillion per day (1 quadrillion equals 1000 trillion).
“It makes a lot of sense to see the results, but it's not something that's on a lot of people's radar. These cars can actually use tons of processing power. They have a 360-degree view of the world, so even though we have two eyes, they can have 20 eyes looking everywhere and trying to understand everything that's going on at the same time," says Karaman.
Autonomous vehicles will be used to transport goods and people, he said, so vast amounts of computing power can be distributed through global supply chains. And his model only takes into account the calculations, it does not take into account the energy consumed by the car's sensors or the waste generated during production.
emission control
The researchers found that each self-driving car would need to consume less than 1.2 kilowatts of power for the calculations to prevent emissions from spiraling out of control. To make this possible, computer hardware had to become more efficient at a faster rate, doubling in efficiency every 1.1 years.
One way to increase this efficiency would be to use more specialized hardware designed to run specific driving algorithms. Sudhakar says that because researchers know the navigation and perception tasks required for autonomous driving, it may be easier to develop specialized equipment for these tasks. But vehicles typically have a lifespan of 10 or 20 years, so one of the challenges in designing specific hardware will be testing it so it can run future algorithms.
In the future, researchers can make algorithms more efficient so that they require less processing power. However, this is also difficult because giving up some precision in favor of greater efficiency can reduce the safety of the vehicle.
Now that they have demonstrated this structure, the researchers want to continue studying the efficiency of the hardware and improvements in the algorithms. Furthermore, they say their model could be improved by characterizing carbon emissions from the car's manufacturing process and emissions from car sensors – the carbon embodied in self-driving cars.
While many scenarios remain to be explored, the researchers hope this work sheds light on a potential problem that people may not have thought of.
"We hope that people will consider carbon emissions and efficiency as important parameters to take into account when developing their projects. The energy consumption of an autonomous vehicle is crucial not only for extending battery life, but also for the longevity,” says Sze.
This research was supported in part by the National Science Foundation and an MIT-Accenture grant.