Tesla Dojo Opens for AI Training

Tesla has started deploying its Dojo supercomputer that employs unusual technologies to pack AI-training performance in a smaller volume than Nvidia-based systems.

Joseph Byrne

Liberating itself from Nvidia for AI processing, electric-vehicle company Tesla is rolling out its Dojo AI systems and plans to accelerate deployment through 2024. The company expects to have 13 exaflops online this year and 100 exaflops by the end of the next (the equivalent of more than 25,000 Nvidia H100 chips), spending more than $1 billion on hardware and R&D. Its aim is to improve its cars’ driver-assistance capabilities and autonomy. Deploying Dojo, Tesla has become one of the few companies to put into production a large-scale AI system built on technology other than Nvidia’s.

Dojo employs the company’s D1 AI processor in an unusual system design built from nesting arrays of computational elements: trays contain tiles that contain chips that contain custom CPUs. Each CPU connects directly to its neighbors, even those in separate chips on adjacent tiles. A system, thus, is a large, flat two-dimensional grid of these CPUs. Clustering multiple systems together, Dojo extends the array.

In our initial coverage of the D1, we noted that Tesla was likely to encounter growing pains during system integration. An important issue was the D1 had extraordinary chip-to-chip bandwidth but no external memory interface. Competing AI processors have less on-chip memory and fast interfaces to large local DRAMs (e.g., HBM) to run large neural networks. Indeed, the company at first couldn’t move data fast enough to sufficiently utilize the D1s.

To help rectify the problem, Tesla created more custom silicon and circuit boards. Dojo Interface Processor (DIP) cards place memory near the tiles and add a plane of Ethernet connectivity. These boards employ a custom chip, as do the systems’ network-interface cards (NICs).