The calculation of the MIT researchers is not very complex, perhaps the foundation lies on the back of a beer mat in a student bar in Boston. They assumed that 1 billion autonomous cars drive an average of one hour each day with the computer using 840 watts. In that situation, the autonomous cars together already produce more CO2 than all the data centers in the world do at the moment. Data centers are now responsible for 0.3% of global greenhouse gas emissions.
Needs much more efficient hardware
Based on these results, the researchers calculated different scenarios. In 90% of these scenarios, the result was that each car would need to use less than 1.2 kilowatts of energy to prevent the cars from using more energy than the data centers do now. Reducing consumption to 1.2 kilowatts would require much more efficient computer hardware than is currently available.
Improvements in hardware efficiency have also been calculated and also paint a bleak picture. Drastic steps forward are needed in a scenario where 95% of the total number of cars in the world will drive autonomously by 2050, and the workload that the on-board computers will have to process will double every three years. In that scenario, hardware efficiency must double faster than once every 1.1 years. Otherwise, the emissions cannot be kept below the desired level.
The model is actually very complex
Sudhakar admits in an article on MIT News that the model seems very simple at first glance. “However, each of these variables contains a high degree of uncertainty because we are studying an application that is not there yet.” The researchers hypothesized that autonomous cars are actually fully autonomous. This means that the vehicles no longer really need a driver who must be ready to intervene, as is currently the case.
This can change the use of cars. For example, research—conducted by other groups of scientists—makes it likely that people will drive more car miles when truly autonomous cars become available. Precisely because it becomes possible to do other things during the journey, travel becomes less of an obstacle. The elderly, who no longer get behind the wheel themselves, can also become much more mobile with a self-driving car.
Yet other research suggests that the amount of time people spend in cars is decreasing. The reasoning behind this is that algorithms are much better able to calculate optimal routes to a destination and prevent traffic jams through better interaction between the cars. Such optimization will on average lead to shorter journey times.
Sudhakar’s statistical model incorporates all such variables. The model also had to make assumptions about the development of hardware and software that did not yet exist. The researchers have implemented the model on a multitasking deep neural network that can calculate the influence of many variables at the same time.
Energy consumption increases very quickly
According to Sudhakar, it came as a surprise to everyone how quickly the need for computing power increases with more sensors in the car. As an example, he mentions a car with ten neural networks that processes input from ten cameras and drives for one hour a day. This arrangement ensures that the on-board systems perform 21.6 million calculations per day. So a million cars do 21.6 quadrillion (ten to the power of fifteen) calculations. Currently, all Facebook data centers combined do more than a thousand times fewer calculations.
“These calculations are not immediately in everyone’s mind,” says Sertac Karaman, director of the Laboratory for Information and Decision Systems (LIDS) and one of the authors of the IEEE-Micro article. As more autonomous vehicles are used to transport goods, a huge amount of computing power is created in the distribution chain. “And the model used only takes into account the energy consumption of the on-board computer systems, not the necessary sensors such as cameras and radar.
Efficiency is already important in design
Sudhakar concludes from the study that paying close attention to hardware efficiency is very important now. “Cars typically have an economic life of 10 to 20 years, so the challenge is to develop specialized hardware that is future-proof and can also handle new algorithms.” Another challenge is that a sharp reduction in the energy consumption of computer systems still means a decrease in reliability, which is undesirable for the safety of self-driving cars.
In short, this research gives developers a lot of topics to think about when designing the systems for autonomous cars.