- Tesla’s “Cortex” supercomputer cluster is being inbuilt Austin, Texas, to assist autonomous automobiles and humanoid robots.
- The supercomputer might be powered by over 100,000 Nvidia H100 and H200 chips.
- The power seems just like different knowledge facilities, with thick cables and loud cooling programs.
As its lead within the electrical car arms race begins to wane, Tesla’s answer for staying forward sooner or later is to deal with autonomous automobiles. However the way it supposedly goals to make automobiles drive themselves is completely different from different gamers on the market.
It entails gathering colossal quantities of video footage from hundreds of thousands of its electrical automobiles worldwide, processing that data at AI-powered knowledge facilities after which packaging that into software program updates earlier than sending them over-the-air with the aim of creating Teslas drive like people. To make this a actuality, Tesla has began constructing large knowledge facilities at its Gigafactories.
One in every of these amenities is being constructed on the south aspect of the corporate’s headquarters in Austin, Texas. The system at this advanced known as Cortex, a “supercomputer cluster” that can run on greater than 100,000 Nvidia H100 and H200 graphics processing items (GPUs) for video coaching of the Optimus humanoid robots and the Full-Self Driving (FSD) system.
CEO Elon Musk shared a video of Cortex this morning. Although the power seems just like different knowledge facilities, the entire surroundings offers off a high-tech, futuristic vibe. It is considered one of our first-ever appears on the supposed know-how behind the know-how.
There seem like numerous aisles of black server racks with tightly stacked {hardware} positioned behind glass doorways and casings. Thick bundles of pink and blue cables appear to snake between the racks and there is a loud hum, presumably emanating from the overhead cooling programs.
Talking of cooling, the power has gargantuan followers on the roof, in all probability the dimensions of an plane propeller, and even greater—Tesla is utilizing a patented cooling system for a similar. Furthermore, the Texas supercomputer will apparently require 500 megawatts of energy sooner or later. A Duke College research mentioned a sports activities stadium can devour 5 mW of electrical energy throughout a sport. So 500 mW might energy about 100 such stadiums without delay, which highlights the power-hungry nature of AI knowledge facilities.
This is not Tesla’s solely supercomputer. The corporate can also be constructing its “Dojo” supercomputer at its New York Gigafactory, additionally for video coaching of the AI-based FSD programs in its future EVs. The pinnacle honcho claimed early this 12 months that Tesla is investing $500 million in New York to construct out this facility, which can also be producing the NACS adapters for house owners of Ford, Rivian and different manufacturers.
“5 hundred million, whereas clearly a big sum of cash, is simply equal to a 10k H100 system from Nvidia,” Musk mentioned on the time. “Tesla will spend greater than that on Nvidia {hardware} this 12 months. The desk stakes for being aggressive in AI are not less than a number of billion {dollars} per 12 months at this level.” (Nevertheless, Musk’s choice to prioritize sending Nvidia chips to social media platform X—which is privately held and technically unrelated to Tesla—raised some eyebrows earlier this 12 months.)
One other one in Memphis is already operational. It is being expanded in the intervening time and when full it could be “essentially the most highly effective AI-training cluster on the earth,” in line with Musk.
Since asserting a pivot to AI and autonomy, the corporate’s passenger automobile and public charging companies have taken a again seat. Other than the refreshed Mannequin 3 and the Cybertruck, the remainder of Tesla’s line-up is getting older quick and that is considerably mirrored in its gross sales slip. Musk needs Tesla to be a tech firm constructing robotaxis, humanoid robots and AI.
For now, Tesla’s FSD has proven numerous flaws in real-world driving circumstances and it would not fairly advantage the FSD label but. The automaker additionally hasn’t handled who or what can be liable within the occasion of a robotaxi crash, and it faces quite a few investigations—together with a U.S. Division of Justice probe—into the way it has bought this know-how to traders and the general public.
We could get some solutions at Tesla’s robotaxi debut occasion on Oct. 10.
Are you aware something about Tesla’s newest plans? Contact the writer; we’re blissful to talk securely and anonymously. suvrat.kothari@insideevs.com
