Elon Musk's Ambitious Plan for Tesla Dojo AI Supercomputer Unveiled

For years, Elon Musk has been vocal about Dojo, the AI supercomputer at the heart of Tesla's AI ambitions. In July 2024, Musk emphasized its importance, announcing that Tesla's AI team would intensify efforts on Dojo ahead of the much-anticipated robotaxi reveal in October.

So, what exactly is Dojo, and why is it so crucial to Tesla's long-term strategy?

In essence, Dojo is Tesla's custom-built supercomputer designed to train its "Full Self-Driving" (FSD) neural networks. The enhancement of Dojo is closely tied to Tesla's goal of achieving full self-driving capabilities and launching a robotaxi service. Currently, FSD, installed in hundreds of thousands of Tesla vehicles, can handle certain automated driving tasks but still requires a human driver's attention.

Following the Cybercab reveal in October, Tesla is set to launch an autonomous ride-hail service in Austin this June, using its own fleet of vehicles. During the 2024 fourth-quarter and full-year earnings call in January, Tesla announced plans to introduce unsupervised FSD for U.S. customers in 2025.

Despite Musk's earlier claims that Dojo was the key to Tesla's full self-driving ambitions, he's been relatively quiet on the subject as the goal approaches. Since August 2024, the conversation has shifted to Cortex, Tesla's new AI training supercluster at its Austin headquarters. Musk has described Cortex as having "massive storage for video training of FSD & Optimus."

While Tesla's Q4 shareholder deck provided updates on Cortex, it was silent on Dojo. Tesla is clearly investing heavily in AI, including Dojo and now Cortex, to achieve autonomy in both vehicles and humanoid robots. With the EV market becoming increasingly competitive, Tesla's future success depends heavily on mastering these technologies.

Tesla's Dojo Backstory

Elon Musk's vision for Tesla extends beyond just manufacturing cars, solar panels, and energy storage systems. He wants Tesla to be an AI powerhouse, pioneering self-driving technology by emulating human perception.

Unlike many other autonomous vehicle tech companies that rely on a mix of sensors like lidar, radar, and cameras, along with high-definition maps, Tesla aims to achieve full autonomy using cameras alone. These cameras capture visual data, which is then processed by advanced neural networks to make real-time driving decisions.

At Tesla's first AI Day in 2021, former head of AI, Andrej Karpathy, likened the company's goal to building "a synthetic animal from the ground up." Musk had been hinting at Dojo since 2019, but it was officially unveiled at AI Day.

While companies like Alphabet's Waymo have successfully commercialized Level 4 autonomous vehicles, Tesla's system still requires a human behind the wheel. Over 1.8 million people have subscribed to Tesla's FSD, which can cost anywhere from $8,000 to $15,000. The promise is that Dojo-trained AI will be rolled out to customers via over-the-air updates. The extensive data collected from FSD vehicles—millions of miles of video footage—is used to train and refine the system.

However, some industry experts question the efficacy of this data-heavy approach. Anand Raghunathan, a Purdue University professor, told TechCrunch, "First of all, there’s an economic constraint, and soon it will just get too expensive to do that. Some people claim that we might actually run out of meaningful data to train the models on. More data doesn’t necessarily mean more information, so it depends on whether that data has information that is useful to create a better model, and if the training process is able to actually distill that information into a better model."

Despite these concerns, the trend towards more data seems to be sticking around, at least for now. And with more data comes the need for more computational power to store and process it all, which is where Dojo, the supercomputer, comes into play.

What is a Supercomputer?

Dojo is Tesla's supercomputer system, designed specifically to train AI models, particularly for FSD. The name "Dojo" is a nod to martial arts training spaces.

A supercomputer consists of thousands of smaller computers called nodes, each with a CPU and GPU. The CPU manages the node, while the GPU handles complex tasks like splitting operations and working on them simultaneously. GPUs are crucial for machine learning tasks such as those used in FSD training simulations. They also power large language models, which explains why Nvidia has become the most valuable company on the planet due to the rise of generative AI.

Tesla currently uses Nvidia GPUs to train its AI, but it's looking to diversify its hardware sources due to the high cost and limited availability of Nvidia chips.

Why Does Tesla Need a Supercomputer?

Tesla's vision-only approach necessitates a supercomputer. The neural networks that power FSD are trained on massive amounts of driving data to recognize and classify objects around the vehicle and make driving decisions. When FSD is in use, these networks must continuously collect and process visual data at speeds that rival human perception.

In essence, Tesla aims to create a digital replica of the human visual cortex and brain functions. To achieve this, Tesla needs to store and process vast amounts of video data from its global fleet of vehicles and run countless simulations to train its AI model.

Dojo pics pic.twitter.com/Lu8YiZXo8c

— Elon Musk (@elonmusk) July 23, 2024

Tesla relies on Nvidia for its current Dojo training computer but is working on its own custom hardware to increase efficiency and reduce costs. The goal is to develop hardware that boosts bandwidth and reduces latencies, moving away from standard GPU hardware.

Tell Me More About These Chips

Tesla, much like Apple, believes in designing hardware and software that work seamlessly together. This philosophy led Tesla to develop its own chips for Dojo.

At AI Day in 2021, Tesla introduced the D1 chip, a palm-sized silicon square. Production of the D1 began by at least May 2024, with manufacturing handled by the Taiwan Semiconductor Manufacturing Company (TSMC) using 7 nanometer semiconductor nodes. The D1 boasts 50 billion transistors and a large die size of 645 millimeters squared, promising high power and efficiency for complex tasks.

"We can do compute and data transfers simultaneously, and our custom ISA, which is the instruction set architecture, is fully optimized for machine learning workloads," said Ganesh Venkataramanan, former senior director of Autopilot hardware, at Tesla's 2021 AI Day. "This is a pure machine learning."

While the D1 is powerful, it doesn't quite match Nvidia's A100 chip, which also uses a 7 nanometer process and contains 54 billion transistors with a larger die size of 826 square millimeters.

To achieve higher bandwidth and compute power, Tesla's AI team combined 25 D1 chips into a single tile, functioning as a unified computer system. Each tile offers 9 petaflops of compute power and 36 terabytes per second of bandwidth, and includes all necessary hardware for power, cooling, and data transfer. Six tiles make up one rack, two racks form a cabinet, and ten cabinets create an ExaPOD. At AI Day 2022, Tesla revealed plans to scale Dojo by deploying multiple ExaPODs, forming the supercomputer.

Tesla is also developing a next-generation D2 chip to address information flow bottlenecks. The D2 would integrate the entire Dojo tile onto a single silicon wafer.

Tesla hasn't disclosed how many D1 chips it has ordered or expects to receive, nor has it provided a timeline for when Dojo supercomputers will run on D1 chips.

In response to a June post on X about building a giant GPU cooler in Texas, Musk said Tesla aims for "half Tesla AI hardware, half Nvidia/other" over the next 18 months. The "other" could include AMD chips, according to Musk's January comments.

What Does Dojo Mean for Tesla?

By controlling its chip production, Tesla could potentially add significant compute power to its AI training programs at a lower cost, especially as production scales with TSMC.

This move could also reduce Tesla's reliance on Nvidia's increasingly expensive and scarce chips. During Tesla's second-quarter earnings call, Musk expressed concern about securing steady GPU supplies, emphasizing the need to focus more on Dojo.

Despite these efforts, Tesla is still purchasing Nvidia chips for its AI training. In June, Musk posted on X:

Of the roughly $10B in AI-related expenditures I said Tesla would make this year, about half is internal, primarily the Tesla-designed AI inference computer and sensors present in all of our cars, plus Dojo. For building the AI training superclusters, Nvidia hardware is about 2/3 of the cost. My current best guess for Nvidia purchases by Tesla are $3B to $4B this year.

— Elon Musk (@elonmusk) June 2024

"Inference compute" refers to the real-time AI computations performed by Tesla vehicles, distinct from the training compute handled by Dojo.

Dojo represents a risky yet potentially rewarding venture for Tesla. Musk has acknowledged that Tesla might not succeed with Dojo, but he envisions it as a cornerstone for Tesla's AI division, initially tailored for computer vision labeling and training for FSD and Optimus, Tesla's humanoid robot.

Future versions of Dojo could be adapted for general-purpose AI training, though this would require rewriting software currently designed for GPUs. Alternatively, Tesla could rent out its computing power, similar to AWS and Azure.

During Q2 earnings, Musk mentioned seeing "a path to being competitive with Nvidia with Dojo." A September 2023 report from Morgan Stanley suggested that Dojo could add $500 billion to Tesla's market value by opening new revenue streams in robotaxis and software services.

In essence, Dojo's chips serve as an insurance policy for Tesla, with the potential for significant returns.

How Far Along is Dojo?

Reuters reported that Tesla began production on Dojo in July 2023, but a June 2023 post from Musk indicated that Dojo had been "online and running useful tasks for a few months."

Tesla anticipated that Dojo would rank among the top five most powerful supercomputers by February 2024, though this hasn't been publicly confirmed, casting doubt on its achievement. The company also projected that Dojo's total compute would reach 100 exaflops by October 2024. (One exaflops equals 1 quintillion computer operations per second. To achieve 100 exaflops, assuming one D1 can achieve 362 teraflops, Tesla would need over 276,000 D1s, or around 320,500 Nvidia A100 GPUs.)

In January 2024, Tesla pledged $500 million to build a Dojo supercomputer at its gigafactory in Buffalo, New York.

In May 2024, Musk mentioned that the rear portion of Tesla's Austin gigafactory would be dedicated to a "super dense, water-cooled supercomputer cluster," which turned out to be for Cortex, not Dojo.

Following Tesla's second-quarter earnings call, Musk posted on X that Tesla's AI team is using Tesla HW4 AI computer (renamed AI4), the hardware found in Tesla vehicles, in the training loop with Nvidia GPUs. He noted that the breakdown is roughly 90,000 Nvidia H100s plus 40,000 AI4 computers.

"And Dojo 1 will have roughly 8k H100-equivalent of training online by end of year," he continued. "Not massive, but not trivial either."

Tesla hasn't confirmed whether these chips are now operational and running Dojo. During the company's fourth-quarter 2024 earnings call, Dojo wasn't mentioned, but Tesla did report completing the deployment of Cortex in Q4, which enabled V13 of supervised FSD.

This story originally published August 3, 2024, and we will update it as new information develops.