This article comes to us courtesy of EVANNEX, which makes and sells aftermarket Tesla accessories. The opinions expressed therein are not necessarily our own at InsideEVs, nor have we been paid by EVANNEX to publish these articles. We find the company's perspective as an aftermarket supplier of Tesla accessories interesting and are happy to share its content free of charge. Enjoy!

Posted on EVANNEX on July 12, 2021 by Iqtidar Ali

Tesla's Elon Musk recently announced via his Twitter feed that the automaker is going to be holding a long-awaited Tesla AI Day sometime soon. The purpose of Tesla's upcoming AI Day is to recruit fresh talent to take Full Self-Driving software to the next level. In turn, Musk and his team are expected to give a detailed account of ongoing progress surrounding the development of Tesla's self-driving artificial intelligence efforts.

Above: Looking ahead to Elon's overview of artificial intelligence coming soon via Tesla's AI Day (Source: Tesla / E3 Expo)

Tesla recently transitioned to a new approach in the application of Full Self-Driving in its cars. The upcoming versions of Tesla FSD software are just going to be vision-based, the forward-facing radar is now removed from the algorithm. Tesla already started abandoning programming logic for Neural Net-based decision-making last year. Now it’s boldly moving ahead with a “pure vision” approach.

With ongoing changes in the works, what will Musk actually discuss at Tesla's upcoming AI Day? We received a few clues when Andrej Karpathy, Director of AI at Tesla, revealed a workflow diagram (see below) of the Tesla Neural Net processing architecture when he was addressing the annual computer vision event CVPR ’21.

Tesla Neural Net Architecture.
Above: Tesla Neural Net Architecture (Source: Andrej Karpathy via CVPR'21)

Karpathy also revealed during his talk that Tesla has added the third cluster to its in-house neural net supercomputer, which makes it now one of the world’s largest supercomputers. According to Karpathy, Tesla's 1.5 petabytes of data requires a huge amount of computing power — the reason for scaling the Tesla neural net supercomputer.

In short, there's a huge amount of Autopilot video data to process from around 1.5 million cars globally (8 cameras on each vehicle) so Tesla is building this powerful supercomputer to train the Neural Net. According to Karapathy, Tesla has adopted the vision-based FSD approach in lieu of Lidar as that approach remains difficult to scale.

Tesla supercomputer
Above: Glimpse of the Tesla vision AI supercomputer (Source: Andrej Karpathy via CVPR'21 / Tesla)

“It’s actually quite unscalable to collect, build, and maintain these high-definition Lidar maps. It’s incredibly expensive to keep this infrastructure up to date. So we took the vision-based approach, which of course is much more difficult because you actually have to get in your neural networks that function incredibly well based on the videos,” Karapathy explained at the CVPR’21 event.

Tesla is scaling up its Autopilot FSD and asking computer science students and interested developers to join the team if they want to be part of the next disruption in the automotive industry — self-driving cars and (eventually) robotaxis.

In advance of Tesla's AI Day, software engineers who might be interested in an opportunity to work on one of the world’s most powerful and advanced supercomputers can apply for a job at Tesla by emailing their interest to supercomputing@tesla.com.

===

Written by: Iqtidar Ali. An earlier version of this article was originally published on Tesla OracleRevised update edited by EVANNEX.

Got a tip for us? Email: tips@insideevs.com