Tesla is confidently doubling its Dojo D1 supercomputer chip orders with TSMC, according to a report published by Taiwan’s Economic Daily. The custom Tesla application-specific integrated circuit (ASIC) is designed to power its Dojo supercomputer, to train driver assistance and self-driving AI models, but is also expected to find wider use in robo-taxis and associated services.
The report from Taiwan says that Tesla will double its Dojo D1 ASIC order with TSMC to 10,000 units in the coming year. It adds that the order volume will continue to increase into 2025. The Economic Daily frames this story partially from TSMC’s perspective, saying that the chipmaker’s HPC-related order momentum has increased thanks to Tesla.
Neither Tesla nor TSMC have issued statements about the rumored Dojo D1 chip order increase. The source report characterizes the Tesla / TSMC as something of an open secret. Moreover, Economic Daily sources claim to know that the Tesla supercomputer ASIC “mainly uses TSMC’s 7nm family process and combines it with InFO-level system-on-wafer (SoW) advanced packaging.”
We reported on the Dojo D1 supercomputer chip back in August 2021. Yes, exciting tech can be a long time coming, but the official news release gave us some specs to dig our teeth into. For example, it is claimed that the D1 ASIC packs 50 billion transistors and is capable of 362 TeraFLOPs at FP16/CFP8 precision, or about 22.6 TeraFLOPs of single-precision FP32 tasks.
Currently, Tesla is a big consumer of Nvidia GPUs. In 2021 it is claimed Tesla was using a primary unnamed cluster using 5,760 Nvidia A100 graphics processing units (GPUs) to develop its Autopilot ADAS. Last month we reported that Tesla was due to switch on a powerful new Nvidia-based supercomputer, packing in 10,000 Nvidia H100 GPUs. Demand for Nvidia GPUs is outstripping supply, but hopefully the upcoming success of Dojo will alleviate this.
A central feature of Project Dojo is scalability, hence Tesla can keep ordering ASICs from TSMC to build out multiple ‘ExaPODs’ – each with 1,062,000 cores (housed in 3,000 D1 chips) and capable of up to 20 ExaFLOPs.
Tesla will likely start to leverage its AI computing power outside of its automotive-related business, Morgan Stanley analysts suggested to CNN. They foresee the Dojo supercomputer addressing new markets, benefitting the business in a similar way to what AWS does for Amazon.