Elon Musk commits $1 billion toward building Tesla's Dojo supercomputer

It should be ready by the end of 2024 but can we take Musk's word at face value when it comes to deadlines?
Ameya Paleja
Tesla's Dojo supercomputer
Tesla's Dojo supercomputer


Tesla CEO Elon Musk has committed to spending a billion dollars to build the Dojo supercomputer over the next year, Bloomberg reported. Unveiled in 2019, Project Dojo aims to train machine learning models to improve the Full Self Driving (FSD) feature in Tesla's cars.

The FSD on Tesla is offered as a feature in "beta mode", meaning it is far from being the advanced driver-assist system that Musk has declared it to be. Tesla has occasionally provided updates on the software, but many have pointed out the flaws in the design and how it cannot be trusted.

Tesla cars carry as many as nine cameras for driver assistance features. So far, the company has collected 300 million miles of video data that it plans to use to train its machine-learning models to improve the FSD. To do so, however, it intends to use its supercomputer, dubbed Dojo by Musk.

What is Project Dojo?

Tesla has been using clusters powered by Nvidia's A100 chips for training its Autopilot but in 2019; Musk mentioned that the electric vehicle maker would switch to using self-designed chips. It took the company another two years to officially announce a supercomputer powered by its D1 chips and dubbed Project Dojo.

Last year, Tesla claimed that it had moved forward from just a chip and training tile and built a cluster, referred to as an Exapod, after the exaflop or 1 quintillion (​​1018) floating-point operations per second capability of the computer.

So powerful was the setup that when the company tested the infrastructure, it tripped the substation in Palo Alto, Electrek had then reported.

Can Musk's deadlines be trusted?

Like other Tesla products, taking Musk's predictions at face value is hard. Promised in 2021, the first Cybertruck has recently rolled out of Tesla's factory.

Elon Musk commits $1 billion toward building Tesla's Dojo supercomputer
Tesla has designed its own chips to build its Dojo supercomputer

One could argue that building and training a supercomputer is very different from assembling cars at the production level. While that might be true, one cannot dismiss that the central to these promises is Musk, who is known to set highly ambitious deadlines and not deliver on them.

At Tesla AI Day in October last year, Musk expected the cluster to be ready by early 2023. Last month, Tesla confirmed that Dojo was online and training was scheduled to begin in July. Musk's latest statement confirms that Dojo needs further investments, and the supercomputer might not begin delivering results well into the next year either.

Wall Street reacted negatively to the news about the expense, and Tesla's stock price slid down. Tesla's approach to vertical integration now requires it to work on vehicle hardware, software, and training of models internally to solve the problem of vehicle autonomy at scale.

Musk's approach to building its supercomputer to train its models may benefit him in his other ventures, such as xAI, which he recently announced. But is it slowing down Tesla's rollout of the FSD as well?

As recent examples have shown, training models are becoming increasingly simpler for enterprises, and one does not have to start from scratch to be a pioneer in a field. As a former investor in OpenAI, Musk would know better.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board