The Real Reason Elon Musk is Building the DOJO Supercomputer 🤖💥 | AI Revolution - video Dailymotion

Ai Revolution

Elon Musk is changing the AI game once again with Tesla’s secret weapon — the Dojo Supercomputer 🧠⚡. Built entirely in-house, Dojo is designed to train Tesla’s Full Self-Driving system with unmatched speed and precision 🚘📈. By cutting out external suppliers like Nvidia and unleashing Tesla’s own D1 chips, Musk is betting big on full autonomy and AI dominance 🤯💻. This isn’t just about cars — it’s about the future of intelligent machines 🌐🚀.  Discover why Dojo might be the most powerful AI system you’ve never heard of — and why it could reshape the world as we know it 🔍🌍.  #ElonMusk #TeslaDojo #AIRevolution #Supercomputer #FullSelfDriving #AutonomousVehicles #TeslaFSD #MachineLearning #NeuralNetworks #TechInnovation #FutureOfAI #TeslaTechnology #ArtificialIntelligence #DojoSupercomputer #D1Chip #NextGenAI #SelfDrivingCars #ElonMuskUpdates #AIInfrastructure #AITraining

Transcript

00:00One of the biggest gambles that Elon Musk and Tesla have made in recent years is the design

00:08and production of their own bespoke supercomputer, Dojo. Musk has said repeatedly that Dojo is a

00:16risky move, but it's also a technology that Tesla might not be able to live without. Let's talk

00:23about why that is. What's really fascinating is that when Elon Musk started talking about Tesla's

00:33Project Dojo back in 2019, the demand for graphics processing units or GPUs was still limited

00:40primarily to the video game and digital video rendering industry. Supercomputers were pretty

00:45much only thought of as machines that solved medical mysteries and made climate predictions.

00:51Yet at Tesla's first autonomy day, we saw the company's AI team standing on stage telling the

00:58world that they were going to build their very own supercomputer and use it to train artificial

01:03intelligence. Very few people at the time would have known how significant this was. Flash forward

01:11five years and now GPUs and supercomputer clusters have become the single most important piece of

01:17infrastructure in the AI revolution that has swallowed the world. The rapid shift in the current of the

01:25tech industry has catapulted Nvidia to the top of the stock market, making the world's biggest GPU

01:31supplier into the world's most valuable company. Ironically, Elon Musk and Tesla have had a major role

01:38to play in the rise of Nvidia. Tesla has spent billions and billions of dollars on Nvidia AI hardware.

01:45This is most evident in the company's latest construction project, a gigantic addition built

01:51onto the south end of the Texas Gigafactory that is now home to a supercomputer cluster that Elon Musk

01:57has named Cortex. About half of the chips that power this data center will be made by Nvidia,

02:04amounting to about 50,000 of their H100 GPUs. But a significant amount of the computing power in

02:11Cortex will also be made by Tesla. And somehow, at the same time as Cortex was being built,

02:17Elon also found the time to construct an even bigger AI training computer for his XAI startup

02:23company at a location in Memphis, Tennessee. That one is known to house up to 100,000 of Nvidia's

02:31current H100 and H200 GPU chips, with Elon already threatening to buy an additional 300,000 units of

02:39Nvidia's next generation chip, the B200. So, all that to say, Elon Musk knows the value of Nvidia

02:46hardware better than most, which is why he has been pushing his own staff at Tesla to develop something

02:53that can compete with what Nvidia is doing. And this is about more than just trying to cut out some of

02:59Nvidia's market share. It's more about preserving Tesla's own ability to remain at the front of the AI

03:06race. During Tesla's Q2 earnings call this year, Elon said that demand for Nvidia hardware is so high

03:13that it's often difficult to get the GPUs. He continued by saying, quote,

03:18I am quite concerned about actually being able to get steady GPUs when we want them,

03:23and I think this therefore requires that we put a lot more effort on Dojo in order to ensure that

03:29we've got the training capability that we need. So, one significant reason that Tesla needs Dojo

03:35is simply because no matter how big Nvidia grows, one company can't possibly quench the world's thirst

03:41for AI hardware. As Elon Musk would say, the best supplier is no supplier. Of course, this is also

03:49about money. Economics 101 says that when the demand is high and the supply is limited, the price will go

03:55up. And when a company finds themselves in a position like Nvidia does today, where the wealthiest

04:00companies in the world are demanding more and more hardware than you could possibly produce,

04:06then you can pretty much charge any price you want. Now, that's not to say that designing and

04:12building your own bespoke supercomputer from the ground up is going to be any cheaper than buying

04:16hardware on the open market, not in the short term at least, but Dojo is a long game. Going back to

04:23Tesla's Q1 earnings call 2023, Elon told investors that Dojo has the potential for an order of magnitude

04:30improvement in the cost of training and also has the potential to become a sellable service that we

04:37would offer to other companies in the same way that Amazon Web Services offers web services. He

04:43followed that up by saying, quote, look at Dojo as kind of a long shot bet, but a bet worth making.

04:50In the summer of 2023, Tesla's Dojo D1 chip goes into mass production at TSMC in Taiwan.

04:57It's probably not a coincidence that this timeline matches up with a sudden rise of Nvidia's stock

05:03price. On June 21st, Elon Musk confirms that an early version of the Dojo computer rack has already

05:10been online and running useful tasks at the company for a few months. In January 2024, we get our first

05:17indication of what a functional Dojo training cluster will look like, while Elon also reiterates

05:22that Dojo is a high-risk, high-reward project. He also says that Tesla was pursuing the dual path

05:29of Nvidia and Dojo, that Dojo is working and is doing training jobs. He notes Tesla is scaling up

05:36and has plans for Dojo 1.5, Dojo 2, Dojo 3, and whatnot. Now, Dojo V1 is a $500 million installation

05:44at Tesla's production facility in Buffalo, New York. We are not exactly sure why Tesla chose an

05:50obscure location, but as someone who lives in the same part of the world, we know that it gets cold

05:55during the winter, which probably helps with the cooling. And there's also an abundant stable supply

06:01of green energy from hydroelectric dams. So in many ways, Buffalo is probably a better location for a

06:08data center than Austin or Fremont. Either way, Elon continues to temper expectations on Dojo, reminding us

06:17that $500 million is actually not very much money, saying it's only equivalent to a 10K H100 system from

06:25Nvidia. Tesla will spend more than that on Nvidia hardware this year. The table stakes for being competitive

06:31in AI are at least several billion dollars per year at this point. In June 2024, Elon puts these

06:39numbers into an even more specific context, writing on X. Of the roughly 10 billion in AI-related

06:46expenditures I said Tesla would make this year, about half is internal, primarily the Tesla-designed

06:52AI interface computer and sensors present in all of our cars, plus Dojo, for building the AI training

06:58superclusters, Nvidia hardware is about two-thirds of the cost. My current best guess for Nvidia purchases

07:04by Tesla are $3 to $4 billion this year. He follows that up by reiterating, there is a path for Dojo to

07:12exceed Nvidia, it's a long shot, as I've said before, but success is one of the possible outcomes.

07:18So even among Tesla's own AI hardware spending, Dojo is a lower-level priority than inference computing

07:25hardware. This is the autopilot computer that is integrated into every Tesla vehicle. It is the

07:31hardware that runs the in-car hardware, that takes input from the eight cameras and uses that to make

07:36real-time driving decisions. Tesla is currently producing version 4 of their inference AI computer,

07:42with version 5 expected to arrive sometime late next year, or maybe in 2026, given Elon timelines.

07:49But we know now that this in-car processing power is quickly becoming more important than the data

07:54center training hardware, and that the autopilot V5 chip will be pretty much necessary to achieving

08:00full autonomous driving. Elon wrote on X, our next-gen AI model after this has a lot of promise,

08:07around 5x increase in parameter count, which is very difficult to achieve without upgrading the

08:12vehicle inference computer. Okay, so what does Dojo actually do? That's the tricky bit to explain,

08:19because AI training is about the closest thing to digital alchemy that we've seen so far,

08:24but Elon recently specified that Dojo 1, which is the Buffalo cluster that's being installed right

08:29now, will have the equivalent to roughly 8,000 NVIDIA H100s of training power online by the end of this

08:37year. Elon shared a couple of pictures that show the front and back of the Dojo cabinets, but no specific

08:43details. So we can gather that Dojo V1 is not massive, but not trivial either. It's less than

08:49150 power of Cortex at Giga Texas, and less than one-tenth the power of the XAI Gigafactory of Compute

08:56in Memphis. But in the case of Dojo, raw power shouldn't matter much, keeping in mind that this

09:02is a custom-designed AI training system that was developed entirely by Tesla. So the hardware is

09:09custom-tailored to Tesla-specific applications, which should make it significantly more efficient

09:16at those particular tasks than a generic multi-purpose chip like those from NVIDIA.

09:23Musk has said that the first version of Dojo will be tailored for Tesla computer vision labeling

09:28and training, which is great for FSD and for training Optimus, Tesla's humanoid robot.

09:34But it wouldn't be useful for much else, and for right now, that's totally fine. We know that the

09:41company's main push right now is getting the Robotaxi vehicle ready to be unveiled in October

09:46of this year. And it would make sense that the acceleration of AI training hardware with Cortex

09:52and Dojo is being done specifically to support the Robotaxi project. This is one of the rare vehicles

09:58that will be judged entirely on its functionality, its ability to drive autonomously. So the software

10:05absolutely needs to deliver. Now that does leave an open-ended question on what the future of Dojo

10:12will actually be. If Tesla has already decided that they are going to solve full self-driving

10:17within the next year, or at least in the relatively short term, then what do they really need Dojo for?

10:23Well, many things, really. So first off, Tesla is pushing out the Robotaxi in America and potentially

10:30China, but for FSD to work in every part of the world, the AI will need to be trained on road data

10:36that Tesla hasn't even collected yet. If you think about the quantity of video that Tesla has collected

10:41from the US and China, their two biggest markets, they need equivalent data from Brazil, India, Thailand,

10:48every language, every different variety of road sign or lack of road sign, there is no end in sight for

10:56Tesla's driving AI. And beyond that, we know that Tesla will move into a more generalized AI model

11:02approach for their humanoid robot project. Right now, they are training the robot's vision, navigation,

11:08and object detection networks in pretty much the same way that they train FSD. But for the Tesla bot to

11:14break out into the consumer market and fulfill Elon's bold expectations of bots outnumbering

11:20people, then they're going to need a pretty staggering variety of AI models that go way beyond

11:26vision, which is where Dojo is currently constrained. All we have to go on is a very limited reply from

11:33Elon. He wrote in June 2023, Dojo V2 will address these limitations. So stay tuned. Thank you to

11:42Drift for partnering with us on this video. Check out the link down below in the description.

The Real Reason Elon Musk is Building the DOJO Supercomputer 🤖💥 | AI Revolution

Category

Transcript

Be the first to comment

Recommended