Elon Musk showed off the Dojo supercomputer for training Tesla’s autopilot – it’s equivalent to 8,000 Nvidia H100 AI accelerators

After launching the Memphis Supercluster, “the world’s most powerful artificial intelligence training cluster,” Elon Musk also shared a photo of another supercomputer from one of his companies. This is the Dojo system, built on Tesla-developed Dojo D1 accelerators, which will train autopilot for electric vehicles. During the quarterly report, Musk also said that he would redouble efforts to develop and deploy Dojo due to high prices for Nvidia products.

Image source: x.com/elonmusk

Musk promised to launch Dojo D1 by the end of the year. The performance of this cluster is equivalent to 8,000 Nvidia H100 accelerators, which, according to the businessman, “is not very much, but not a small thing either.” For comparison, the xAI supercomputer opened in Tennessee for AI training will eventually operate with 100 thousand Nvidia H100 accelerators.

Musk first unveiled the giant Dojo D1 chips in 2021, with a target performance of 322 teraflops. Last August, Tesla began searching for a senior data center engineering program engineer, one of the first steps an organization typically takes when planning to launch its own data center. Tesla also increased its orders for the Dojo D1 in September, indicating the company’s confidence in the product. In May it became known that their mass production was already underway.

It looks like these accelerators have now arrived in the US, and Musk has already shared pictures of the Dojo supercomputer. The Dojo D1 chip is a system-on-wafer processor in a 5 × 5 array. That is, 25 ultra-high-performance crystals are made on a single wafer and interconnected using TSMC InFO (Integrated Fan-Out) technology – they work as a single processor and turn out to be more efficient than similar multiprocessor machines. The Tennessee facility is owned by xAI and is used primarily for training the large Grok language model, while the Dojo chips are focused on video training and will be used to work on autopilot technology.