A study by an MIT graduate student Soumya Sudhakar published by the IEEE recently made headlines for claiming that a large fleet of self-driving cars in the distant future could cause as much greenhouse gas emissions as the current global network of computer servers, cited as producing 0.3% of modern emissions.
The study makes the wrong assumptions and as such its conclusion is incorrect, but these types of studies are often held back by opponents of new technologies due to their confirmation bias and used as propaganda. Little dig into the details, so here I go.
They begin by stating that today’s car computers consume about 840 watts of power. That’s as much as a typical home or a few computers in a data center, though it’s a modest fraction of a 60mph electric sedan, which uses around 15,000 watts while driving. Additionally, they expect the amount of energy needed to increase over time rather than decrease – perhaps to a point that rivals the energy used to move the car. The sedan on city streets uses a little less, so that can start to be a sizable part of that total horsepower.
Their calculations suggest that if there are a billion or more of these cars out there, each driving an hour a day, that’s a lot of electricity. If that electricity comes from our modern fossil fuel-powered power grid, that could mean a lot of emissions. They do a lot of calculations on these invalid assumptions to generate a bad result. They try to forecast several decades to produce these numbers, and in one forecast, imagine 3 billion cars.
Let’s unpack all the dubious assumptions
- It is true that some prototype cars can consume energy like this for their calculation. There are also cars that pull much less. They imagine that the software and other sensors will not become more efficient than the prototypes, when in fact they are likely to become much more efficient.
- The paper even assumes that the computational load will increase exponentially and that the only salvation for power consumption is that AI processors also become exponentially more efficient – but not fast enough, and they should double by efficiency each year to catch up. Compute growth path in prototypes is not a predictor of future compute needs.
- The forecast is based on a 2022 architecture with many cameras constantly processing all incoming images. With other sensors, like LIDAR, that require less computation, you can also avoid a lot of camera computation on regions of the image where nothing is known to be present. This is just one of many tricks, most of which have yet to be invented, that will reduce vision’s workload – if vision is in fact the approach decades from now.
- Imagine all cars becoming self-driving. This is unlikely to happen for a very long time, although all cars have some such functionality, even as driver assistance. However, the driving assistance functions are already less powerful.
- They postulate a fossil-heavy grid like today, despite massive efforts underway to improve emissions from this grid. However, as shown below, these cars will not draw electricity from the “average” grid and will likely draw almost entirely from excess solar and other forms of excess energy (nuclear, geothermal, some hydroelectric , wind) and rarely fossil fuels. .
- Nearly all of today’s AV efforts are electric vehicles, and that’s expected to continue. AV vehicles have accelerated the transition from gas-powered cars to electric cars, a significant net gain in emissions for decades to come. VAs facilitate this change because they eliminate all existing problems with EVs in the minds of the public. You don’t care what’s under the hood of a taxi, or its range or recharge time. It is almost never necessary to use fossil fuels.
Power source of autonomous electric vehicles
Today, our grid consists of a mix of fossil fuels, nuclear, solar, wind, hydroelectric and a few other sources. Renewable energies like solar and wind energy only provide energy when the weather demands it. Nuclear power plants provide electricity all the time, whether you need it or not. Hydro and fossil power can change during the day (sometimes slowly, sometimes quickly) depending on expected and actual demand. Full grid capacity is only used on hot summer afternoons and evenings to power air conditioning, although this is expected to change in the future through the use of ice-based thermal storage.
Solar power is now the cheapest form of power plant to build in moderately sunny locations, but it only provides electricity in daylight. A ton of solar power is expected to be installed to provide electricity this afternoon. This same solar will also produce electricity from 8am to 3pm, although there may not be enough demand to use it. To avoid wasting it, the electricity companies will resell the surplus at low prices. And it’s electric cars, especially self-driving cars, that will be looking to make that deal – AVs will even be going to places to plug in during this time. They won’t want to charge during peak hours unless they have to. It is at these peak times that fossil energy is used the most – we don’t want to use it when there is enough other energy.
Nuclear power plants provide electricity all day, including all night, but people don’t want a lot of electricity at night. It doesn’t save much to shut down a nuclear power plant at night, so it’s another great time to charge up those cars. Cars will charge as much as they can when energy is cheap – their owners aren’t stupid – and that’s where the energy is excess, and that’s where the energy isn’t d fossil origin. They will top up at peak times if they have no choice but that will be the exception.
Some regions do not have or are in the process of shutting down nuclear power. They will have to rely on fossil fuels, wind and hydro at night, as well as storage. This creates an emissions risk as nighttime is certainly the most convenient time to charge cars, as most are idle at this time. Charging can be done slower, which is cheaper and better on the batteries – solar charging from 9:30 a.m. to 11:30 p.m. and 1:30 p.m. to 3 p.m., when there is both solar surplus and more cars are idle , should be done more quickly .
There are also some cars that, due to poor planning or unusual days, will need to be charged at 7 p.m., which is the fossil peak (and price peak) of the day. They will pay a high price for it, so it will be their last resort.
I’m not a fan of battery swap systems for consumer cars, although they might make sense for fleets. With battery swapping, you can top up whenever there’s a surplus of energy – green and cheap – without worrying about when cars are on the road. You need to own more batteries – the ones in the cars and the ones at the charging stations and the ones waiting to be swapped out – but you’re totally in control when you charge.
Another modest negative note is that the most efficient combined cycle fossil fuel power plants are slow to slow down in response to demand, so the ability of cars to buy excess electricity takes that slowing problem away, which which might improve the economy of these fuels, but not that much, because cars will only buy electricity if it’s cheap. It also rewards greater efficiency, which is better than punishing it.
Some predict a fully renewable grid for too long, with or even without nuclear. Obviously, there is no big emissions problem for electric cars in this case. New efficiencies, storage technologies and demand management show the way to get there.
But even when the rest of the grid is still heavily fossil fueled, autonomous electric vehicles will use low-emission power almost entirely. The author of the article could have spared himself all the calculations – it doesn’t actually matter how much energy the calculation in cars takes when it comes to emissions. This matters a lot in terms of range, because every megajoule used by computers is unusable for travel – so they’ll try to keep that low if they can, but not because they’re worried about emissions.