At the recent All-In Summit 2024, Tesla CEO Elon Musk shared crucial information about the next-generation AI chip, Dojo2. He stated that Dojo2 chips will begin mass production by the end of 2025 and will play a critical role in Tesla's AI infrastructure.
Specifically, the Dojo chip is responsible for model training, while the in-vehicle chip handles inference. This implies that Tesla will introduce multiple generations of Dojo chips in the future.
Musk mentioned that the performance of the Dojo2 chip is expected to be comparable to NVIDIA's B200 AI training system to some extent. He believes that technology usually requires three major iterations to truly achieve excellence. Therefore, to truly assess the superiority of the Dojo series chips, we will have to wait for the release of Dojo3, possibly in late 2026.
It is worth noting that Tesla's first-generation Dojo chip is already in mass production. This chip utilizes TSMC's InFO_SoW technology, with a 5x5 array of D1 chiplets on a single training module, demonstrating Tesla's advancements in chip design and manufacturing. Additionally, the initial Dojo system also features a V1 interface processor for network interconnectivity, laying the groundwork for future AI development.
Key Points:
🌟 Dojo2 chip expected to start mass production by the end of 2025, comparable to NVIDIA's B200.
🚀 Musk believes technology needs three iterations to achieve excellence, looking forward to Dojo3's performance.
💻 The first-generation Dojo chip is already in mass production, using TSMC's advanced technology, laying the foundation for future development.