At Autonomy's investor day, Elon Musk and his team showed off a lot of numbers. They talked a lot about the growth in fleet size and the performance of the new FSD neural network chip. However, he forgot to include a number. It is the product of these two factors multiplied.
If you look at TOP500.org, which lists the world's 500 most powerful supercomputers, you'll see a lot of fast machines. One of the measures of these systems is Pflops (which corresponds to one quadrillion (one thousand trillion) floating point operations per second).
The fastest supercomputer on the list is Summit, at the Department of Energy in Oak Ridge, Tennessee. It has a capacity of approximately 144 Pflops. And if we add up all the processing power of the first 500 computers, we get a total processing power of approximately 1,415 Pflops.
In about a year, Tesla will have put around a million vehicles on the road equipped with its new “Full Self Driving” chip. About half of them will come with this chip preinstalled, and the other half, which is in circulation today, will be upgradeable.
For the entire fleet, Tesla will have processing power of approximately 117,000 TOPs. Typically, a supercomputer's processing power is measured in floating point operations, while Tesla is only interested in integer multiplication for autonomy. So let's assume that FLOPs are 10 times more “powerful” than OPs (the conversion is not exact, but it should give us an order of magnitude).
This means that Tesla's combined vehicle network has orders of magnitude more processing power than the world's fastest supercomputer, and an order of magnitude more processing power than the fastest 500 computers combined. . When comparing supercomputers, PFlops isn't the only thing that matters, but it's an easy metric to use. It is clear that with the very limited amount of bandwidth available between cars and the cloud, the architecture of the application must be very different from that of a typical supercomputer, and will give it both advantages and benefits. disadvantages.
Tesla's distributed machine will have very interesting characteristics.
- It will be distributed throughout the world
- It will be extremely mobile
- It will be connected to a set of sensors (vision, radar, ultrasound).
- It will operate the world's largest neural network based on AI deep learning and focused on image recognition.
- Its size will continually increase (at an exponential rate over the next few years).
- It will be optimized to consume little energy and will increasingly be powered by solar energy.
- It will automatically and dynamically update.
The objective of this IT behemoth will be limited to a single task: driving vehicles equipped with evolving “artificial intelligence”. This AI will be based on deep learning by neural networks that process billions of kilometers of visual vehicle data per month, as well as other sensor data. The goal of this system will be to perfect “autonomous driving” of cars to a level that no human could hope to achieve.
Although there are other applications distributed with massive numbers of “nodes” which can be seen in bitcoin mining, XBox, Whatsapp clients, etc…. Self-driving juggernaut Tesla is different in that it is a system that improves as the number of nodes increases, and overtime will do so more and more automated manner due to its increasing reliance on neural networks and software engineering algorithms.
As Stewart Bowers pointed out during his Autonomy investor day presentation: “Not only can we look at what's happening around the vehicle, but we can look at how humans have chosen to interact with that environment. We start with a single neural network, detections around it. We then build it all together in multiple neural networks and multiple detections. We then integrate all the other sensors and convert everything into what Elon calls “vector space”, that is, an understanding of the world around us. As we get better, we transfer more and more of this logic into the neural networks themselves. The obvious end goal is for the neural network to look at all the cars, gather all the information and ultimately produce a source of truth for the world around us.”
In reality, the Tesla neural network creates both a model of the world as it is and a probabilistic model of how different things in the world behave based on their appearance. Take the example of a pothole or a bag on the road. If it's a pothole, the car will see it, then notice that it never moves. And if you ride on it, your suspension will register some impact. If it's a bag, it can move (if it moves, it's not a pothole), and if you hit it on the ground or in the air, there is no major impact on the vehicle.
He continues: “We have a neural network that runs on our wide fisheye camera. This neural network doesn't make a single prediction about the world, it makes many distinct predictions, some of which verify each other. These predictions combine to give us a better idea of what we can and cannot offer in front of the vehicle and how to plan for that. (…) We can use this both to learn future behaviors that are very precise, but we can also construct very precise predictions about how things will continue to happen in front of us.”
After Tesla's self-driving rollout, the next largest “dedicated” and naturally scalable application for a computing system could be NOAA's recently upgraded weather forecasting system, which runs on a pair of 8-digit supercomputers. 4 PFlops combined, and forecasts weather for the entire United States. However, while NOAA's latest update represented about 50% more power over three years, Tesla's self-driving AI system is evolving at a much faster pace. As Elon notes: “When things are changing at an exponential rate, it's very difficult to get a handle on it, because we're used to extrapolating on a linear basis. But when you have massive amounts of hardware on the road, the cumulative data grows exponentially and the software improves at an exponential rate.”
In 4 years, in addition to its new 3x faster chip architecture under development, Tesla is expected to have over 5 million cars on the road, with a combined processing power of something close to 2 million POPs ( 22,500 more processing power than NOAA's current pair).
It is also possible that in the next few years Tesla will be able to use high-speed satellite internet (at least outside of dense urban areas) to provide high-speed bandwidth to the network, which is currently severely limited by inconsistent 4G speed connectivity, opening up greater potential for rapid improvement.
However, within 4 years, the competition could turn to Tesla. By then, it is very likely that other major automakers will find a way to update the software on the majority of their new vehicles.