Popular Mechanics — 2023-03-08
Automotive Industry
The growing visibility of climate disasters around the world, such as flooding and wildfires, have made it impossible to ignore the impact that daily luxuries, such as plastic waste, have on our planet. Yet despite progress towards tackling climate change’s most conspicuous culprits, another prominent source may soon be lurking under our noses—or right in our hands.
Whether we’re saving family photos or Googling pictures of suspicious rashes on WebMD, our devices are using energy to reach into “the cloud” and pull out our tiniest whims. The cost of maintaining the data center brain behind all this knowledge—such as keeping the computers from overheating in addition to simply powering them—is a whopping 1% of global electricity consumption, which generates 0.3% of global emissions, according to a 2022 report from the International Energy Agency.
Other invisible sources like digital currencies, including Bitcoin, have made headlines for eating up large chunks of energy, but a paper published by MIT researchers at the end of 2022 suggests that a new source of digital emissions could emerge by 2050 to rival data-center emissions: autonomous vehicles (AVs).
“We wanted to ask the question, in a future scenario where now our daily computing activities involve driving in an [AV], how do we model this new source of emissions from running the computers onboard the [AV]s?” Sudhakar said.
In their paper, Sudhakar and her colleagues consider a number of factors that might contribute to the emissions of AVs, including the number of vehicles on the road, the time those cars spend driving, and the computing power they carry. In particular, it’s the smart sensors built into AVs that the researchers think could seriously add up.
Part of what makes future AVs potentially safer on the road than human-driven cars is that while human drivers at most only have two eyes to survey the road, an AV could have 10 cameras positioned 360-degrees around the car. Each of these cameras is composed of a neural network that is constantly analyzing visual input to ensure the vehicle’s safety. As the researchers outline in their paper, this means that a single AV driving for one hour could make 21.6 million inferences per day and that an estimated fleet of one billion AVs could make 21.6 quadrillion inferences per day.
For comparison, “Facebook runs trillions of DNN inferences per day across its data centers,” the authors write. This effectively makes future AVs “data centers on wheels.”
There are a few caveats to be made for this prediction, for example Sudhakar and colleague’s model is based on level four and five AVs which can drive without human supervision. As of right now, these types of advanced AVs don’t exist. It also doesn’t yet take into account the impact of renewables on this problem, which Sudhakar says would have a positive impact.
The Energy Strain of Cryptocurrency
The prediction by the MIT team may sound futuristic, but this same scenario has already played out with the energy demands of cryptocurrency mining. According to the New York Times, mining one Bitcoin used to take only the computing power of a single household computer, but as the digital currency’s value grew so did its energy consumption. By 2022, the global electricity consumption of Bitcoin mining was between 0.4% and 0.9% of global consumption. This puts it just ahead of the current 0.3% global energy consumption of data centers and the predicted 2050 energy consumption of AVs, and on par with the country of Argentina.
The primary way to address this problem right now, she says, is to focus on developing both hardware and software that help AVs make these inferences more efficiently.
That’s work that Asaf Cidon, assistant professor of electrical engineering and computer science at Columbia University, and Srabanti Chowdhury, associate professor of electrical engineering at Stanford University, are looking to tackle in their work.
One of the problems facing the hardware side, Chowdhury tells Popular Mechanics, is that the silicon materials traditionally used to create sensor technology, like computer chips, have reached their limits.
“Technology is so energy hungry because the world is becoming extremely digitally connected, and it’s a trend that everything we use needs to be smart or even emotional,” Chowdury says. “What that means in electronics is more computing and sensing capabilities and more demand for power.”
To improve the amount of energy a single chip can handle, Chowdury says that new approaches to computing such as mixed-material silicon chips and light-based computation should continue to be implemented. In her lab, Chowdury is looking at diamonds as a way to improve silicon performance. Another pathway is the continued development of 3D computer chips, which have the potential to handle increased input more efficiently than their 2D counterparts.
For his part, Cidon isn’t quite as convinced. For one thing, new cars today are already filled with smart sensors. “Blind spot detection, the automatic braking system, and lane detection… are basically using AI,” Cidon says. “I don’t know if I buy the idea that computation on the car is going to be a big problem.”
That said, Cidon does agree that smart sensors in general create a large carbon footprint. And while building more efficient hardware solutions has been an active area of research for many years, he says that until recently much less attention has been given to improving the efficiency of software. In many cases, there isn’t a clear way for developers to track, for example, how much energy an individual web application is using, he says.
Some ways that software could help address the emissions problem is to reduce the amount of carbon used during the training process for AIs as well as looking for “bloated” computations that use energy while providing little benefit to the overall computation. Another approach is something called “memory pooling” that would let devices share unused space on a computer’s CPU.
Time will tell whether such innovations will really tip the scale to change gloomy predictions on future AV emissions, but the researchers are hopeful that they will contribute to a more efficient and effective computing landscape.