Elon Musk confirms he killed Dojo, Tesla’s supercomputer project, but the reason he claims led to the killing of the project does raise some eyebrows.
Last week, a new report from Bloomberg claimed that Tesla shut down its Dojo supercomputer program after 20 members quit to join a new startup, DensityAI.
The news came just a month after CEO Elon Musk said: “Dojo 2 is good, but Dojo 3 will be great.”
For years, the CEO has been claiming that Tesla has the “best chip design team in the world” and that Dojo’s supercomputer platform could surpass the products made by its current main training compute supplier, NVIDIA.
While Musk didn’t deny the Bloomberg report, he has now commented in a way that suggests a different story:
Once it became clear that all paths converged to AI6, I had to shut down Dojo and make some tough personnel choices, as Dojo 2 was now an evolutionary dead end. Dojo 3 arguably lives on in the form of a large number of AI6 SoCs on a single board.
If you are to believe this narrative, Musk saw that Tesla’s new AI6 chip, which has been developed for inference computer power onboard its vehicles, was so good that he decided to shut down Tesla’s chip program for training compute, Dojo, and fired the team.
Does this narrative hold up to scrutiny?
Electrek’s Take
This doesn’t add up. At best, it is partly true. Let’s say that it’s true that Musk saw the AI6 chip and was so impressed that he believes it would even replace Dojo chips. Then, why did Peter Bannon go to?
Bannon was the chip architect in charge of all custom silicon at Tesla, training compute (Dojo) and inference compute (AI6). According to the Bloomberg report, he also left Tesla amid the talent exodus to DensityAI.
However, at this point, it’s not clear whether he left on his own or if he was part of Musk’s claimed “tough personnel choices.”
If it’s the latter, it wouldn’t make sense since Musk has been so impressed with AI6. Why would he fire Bannon after he delivered the new chip?
This point supports the Bloomberg narrative that critical Dojo team members left, which hurt the program to the extent that it made sense for Musk to shut it down, especially within the context of having difficulties competing against NVIDIA and others.
And then you have to question the logic of using the same chips for training and inference computing. It’s certainly not impossible, but it doesn’t sound like an optimal solution.
They generally utilize different data types and have varying throughput requirements. Training is geared toward high numerical precision while inference needs low latency and high throughput per watt.
To be fair, energy efficiency is also essential for training computers, but it is less so than for an inference computer running in an electric vehicle with limited energy capacity.
You gotta give credit where credit is due: Musk knows how to spin a story.
There’s basically a mutiny in Tesla’s Dojo team, and Musk spins it into Tesla developed such a good inference chip that it fired its training compute chip team.
It is clearly a fake narrative pushed by Musk to justify his very recent pro-Dojo comments.
FTC: We use income earning auto affiliate links. More.
