Tesla's (TSLA 3.47%) AI chips have been called the most advanced in the world, and they'll power the company's new Dojo supercomputer set to become operational later this year. In this Backstage Pass clip from "The AI/ML Show" recorded on Jan. 12, Motley Fool contributor Trevor Jennewine talks about how these developments could shake up the AI industry in 2022.
Trevor Jennewine: At Tesla's AI Day in 2021, they introduced the D1 chip and this will power its Dojo supercomputer. I don't want to get too technical, but each chip offers 362 teraflops of compute performance. That's 362 trillion floating point operations per second. Basically, think of that as 362 trillion calculations each second, so very powerful and Tesla is going to organize those chips, 25 chips per training tile. It's going to take 12 tiles, put them in a cabinet, and it's going to connect 10 server cabinets. You're going to end up with a system that has a 120 training tiles each with 25 of those chips, which brings the compute power of Dojo to over one exa-flop. That is one quintillion floating point operations per second, that's one followed by 18 zeros, so very powerful.
The system is supposed to be operational sometime this year, and if it works out like they presented it at their AI Day, it should theoretically make Dojo the world's most powerful AI training machine. And then on the inference side, this is the in-car supercomputer. I mentioned that Tesla replaced Nvidia's chips with its own custom chips when it released autopilot hardware 3.0. Elon Musk called it the best chip in the world. He says, makes a lot of bold statements so take that with a grain of salt. But analysts took a look at it, the Nikkei Asia review. They did a tear down and they found that at the time it was six years ahead of the industry. ARK Invest took a look at Tesla's AI chips and said it was four years ahead of Nvidia at the time, which is saying something. Nvidia is a gold standard for AI, I think that's impressive.
Tesla's currently in talks with Samsung to build an FSD2 chip, so this would be autopilot hardware 4.0 and that's set to launch with the cybertruck sometime this year. That being said, Musk has said that the FSD1 chips, the one that's in the cars now, will be enough to support full autonomy. However, Tesla said the same thing when they introduced Nvidia chips in 2016 and that turned out to not be the case.
