Tesla Full Self Driving FSD 4D to Get Boost from Dojo - The NN Training Supercomputer

 

In recent years, companies have increasingly sought to develop autonomous driving systems. While others rely on Lidar, Tesla has taken its own unique path, seeking to create a "Smart Autopilot" which interprets data from four dimensions.

An expensive Lidar, which will make any car ugly, determines the environment around the car and, if it detects an object in the way, slows down. At the same time, neatly placed around the Tesla car, cameras and radars, using sophisticated software, not only determine obstacles but are able to understand what kind of object it is and determine the trajectory of its movement.

In February 2019, Tesla filed a patent for 'Generating ground truth for machine learning from time series elements', which disclosed a machine learning training technique for generating highly accurate machine learning results. In fact, it reveals the methodology of the work of the system in 4D.

When we look at images of an event, we see only separate images, but when we watch a video, this gives us a complete understanding of what happened and we can correctly and objectively assess it. The same thing happens with Tesla cars, which are "see" in 4D. When Tesla cameras capture images, they combine them with time (4th dimension) to create surround video.

This is the key point in order to correctly recognize and take into account the trajectory of dynamically occluded objects, which is especially important in places with the dense vehicle and pedestrian traffic.

Good explanation. 4D is essential for dynamically occluded objects, especially in large intersections with dense vehicle & pedestrian traffic. Frame rate & latency from frame to wheel vector change also important.

— Elon Musk (@elonmusk) November 8, 2020

For now, behind all the impressive superpower of Tesla's FSD are huge computers that process billions of miles of data using machine learning via neural networks (NN). But now Tesla is developing a new supercomputer, the Dojo. Its goal is to increase the speed and accuracy of training at least 10 times over the current computer.

According to Musk, its development and creation should take about a year before version 1.0 is released and at the moment it is known that Dojo will be more than twice the speed of the current most powerful supercomputer - Fugaku.

Version 1 is about a year away

— Elon Musk (@elonmusk) November 8, 2020

Tesla rewrote all labeling software for 4D and within a year (or so) the Dojo will begin contributing to NN training. At the moment, it takes Tesla about three days to pass one training model, but this is a very long time. Thus, increasing the power by 10 times will reduce the time to about seven and a half hours. Thus, several trainings can be done in one day, significantly accelerating the trajectory towards Level 5 autonomy.

We rewrote all labeling software for 4D. Very different from labeling single photos. Dojo won’t contribute for about a year. It’s mostly a generalized NN training computer, but benchmark we’re tracking is frames/second. Must beat next gen GPU/TPU clusters or it’s pointless.

— Elon Musk (@elonmusk) November 8, 2020

© 2020, Eva Fox. All rights reserved.

_____________________________

We appreciate your readership! Please share your thoughts in the comment section below.

Article edited by @SmokeyShorts, you can follow him on Twitter

Follow @EvaFoxU