Tesla’s neural network training supercomputer “Dojo” has been hinted at a couple of times by CEO Elon Musk. Dojo is supposed to be a powerful computer processing huge amounts of data so as to achieve the famous vision-only autonomous driving.
Dojo itself is however still in development but Tesla seemed to have moved a little closer to their goal when they revealed a new supercomputer that will serve as a predecessor to Dojo.
The new supercomputer was revealed at the 2021 Conference on Computer Vision and Pattern Recognition last week. Andrej Karpathy, Tesla’s head of AI revealed that the new supercomputer’s processing power will allow Tesla to ditch lidar and radar sensors for quality optical cameras.
A vision only approach is akin to the way human beings respond to new environments. For computers, however, it requires a vast data set and a very powerful supercomputer that will utilize this data to train Tesla’s neural net-based autonomous driving technology.
Tesla’s new supercomputer is rumoured to be the fifth most powerful in the world with 1.8 FLOPS. It has 10 petabytes of “hot tier” storage and runs at a whopping 1.6 terabytes per second.
A vision-only approach is however not a new idea, Elon Musk has been advocating for his neural technology for some time now as it is faster than radar or lidar.
Tesla’s Model 3 and Model Y for instance are being built without radar. Instead, they are going to rely on quite a number of cameras and a powerful machine learning system for its advanced driver system and autopilot.
Lidar is the method of choice for most autonomous driving companies. They require high-definition maps and incredibly detailed maps of the places they operate in.
The vision-only approach on the other hand assesses the immediate environment and according to Karpathy “..can in principle function anywhere on earth.”
“Replacing a “meat computer,” or rather, a human, with a silicon computer results in lower latencies (better reaction time), 360-degree situational awareness and a fully attentive driver that never checks their Instagram,” he added.
He went on to explain that the supercomputer collects video from eight cameras installed all around the vehicle. At 36 frames per second, the supercomputer will capture incredible amounts of data that will enable it to understand the car’s surroundings.
The vision-only approach is a far superior system for autonomous driving compared to radar or lidar systems. It is however more of a challenge to build owing to the large amounts of data that have to be captured and processed at speeds that match that of a human.
With a new supercomputer however and an arguably over-ambitious CEO, Tesla is surely on the road to building the first-ever vision only autonomous vehicle.