2023/09/18 03:26:39
Reported by Economic Daily reporter Yin Huizhong/Taipei
Tesla is sprinting towards the layout of its supercomputer "Dojo". It is rumored that it will expand its cooperation with TSMC (2330) . Its supercomputer chip "D1" is produced using TSMC's 7nm family process combined with advanced packaging. Next year, the volume of chips produced by TSMC is expected to be higher. This year it has doubled to 10,000 pieces, and orders will continue to increase in 2025, fueling TSMC’s high-speed computing (HPC)-related order momentum.
Regarding related rumors, TSMC said it does not comment on market rumors; Tesla did not immediately respond to reporters’ questions. It is reported that Tesla and TSMC have been cooperating for a long time. When TSMC previously held a North American technology forum, Tesla senior executives appeared to share successful cases of cooperation with TSMC, which has convinced the industry of the relationship between the two companies in cooperating on advanced chips. .
Tesla Dojo rumored to expand cooperation with TSMC
Dojo is Tesla’s own supercomputer released in 2021. It is used to train self-driving AI models, speed up the development of Tesla’s self-driving system and develop more advanced self-driving functions. The latest introduction on Tesla’s official website shows that Dojo belongs to AI and robots It is developing and deploying autonomous operation systems on a large scale in vehicles, robots and other fields.
Tesla clearly states on its website that it is convinced that vision and planning based on advanced AI, with the effective use of inference hardware, is the only way to realize the popular application of fully automatic assisted driving, the development of bipedal robots and other applications.
It is understood that Tesla’s supercomputer Dojo mainly uses TSMC’s 7nm family process and combines it with InFO-level system-on-wafer (SoW) advanced packaging InFO-SoW to provide customized high-speed computing needs, and Without the need for a PCB carrier board, the relevant chips can be integrated into the heat dissipation module to speed up the production process.
The basis of Tesla's supercomputer Dojo is the company's self-designed D1 chip, which is designed to have 120 training units and 3,000 chips. Tesla plans to expand its supercomputer computing power lineup at its global engineering headquarters in Palo Alto, California. According to industry conversion, according to the expansion of computing power plan, it will produce nearly 10,000 12-inch wafers next year. wafer.
Tesla has previously been reported to have purchased Huida H100 chips in large quantities, and has recently made a major effort to expand its own supercomputer lineup, leaving the outside world with room for imagination about subsequent application development. It is reported that Tesla will produce about 5,000 supercomputer chips at TSMC this year. If it increases to 10,000 next year, it will double its growth. If it continues to expand chip usage in 2025, the volume will continue to increase.
According to legal analysis, TSMC's 7nm family has previously been impacted by factors such as the slump in non-Apple mobile phones and poor personal computer market conditions. The capacity utilization rate is not as good as expected, which has become the main factor dragging down TSMC's second financial forecast revision this year. Driven by the growth of new AI/HPC-related applications and increasing automotive demand, and driven by the fact that the worst of the personal computer market has passed, TSMC's 7nm family capacity utilization is expected to rebound, which is conducive to the gradual recovery of operations in 2024, but the subsequent recovery still needs to be strong. Observe market conditions and customer base needs.