Elon Musk shares Dojo supercomputer and Nvidia spending plans
TapTechNews June 5 news, Tesla's chief executive Elon Musk today shared some projected information about the future of the Dojo supercomputer cluster and disclosed the scale of Tesla's planned spending on Nvidia products this year.
Musk said that due to the lack of storage space, Tesla needs to transfer some of the Nvidia chips to X and xAI. Then Musk disclosed that Tesla's purchase amount from Nvidia this year is estimated to be between 3 billion and 4 billion US dollars (TapTechNews note: currently about 28.96 billion yuan).
Tesla's GigaTexas southern expansion project is about to be completed and will then accommodate 50,000 Nvidia H100 chips used for FSD training.
There are also netizens asking Musk if Tesla has a realistic path to mass-produce the Dojo supercomputer, so as to rely more on these devices rather than Nvidia hardware in training computing. In this regard, Musk explained that Tesla's demand for training computing actually accounts for only a small part, especially compared to the inferential computing required by the FSD system.
Tesla's training computing is relatively small compared to inferential computing, because the latter increases linearly with the fleet size, Musk wrote. Perhaps it is most appropriate to measure by power consumption. When Tesla's fleet reaches 100 million vehicles, the peak power consumption of the in-vehicle AI hardware will be approximately 100 GW, and the training power consumption may be less than 5 GW. These are all very rough estimates. Obviously, in the current standard, 5 GW of AI training computing is huge, but it only accounts for about 5% of Tesla's total AI computing volume.
Although Musk considers this a long-term goal, he said that it is possible that the production of Tesla's Dojo supercomputer will exceed that of Nvidia in the future. It is not impossible for Dojo to exceed Nvidia, Musk added. As I said before, this is indeed a long-term goal, but success is also one of the potential outcomes.