Tesla published an updated AI accelerator roadmap on November 8 with SI6 compute reaching 80 times that of HW4
The sixth-generation Tesla inference accelerator is projected to deliver 80 times the compute of the current fourth-generation chip, while Dojo D3 moves away from SRAM toward chiplet design with stacked HBM.
On November 8, Tesla published an updated AI accelerator roadmap revealing more of the company's chip development direction for the coming years. According to the roadmap, the sixth-generation Tesla inference accelerator is projected to reach 80 times the compute of the existing fourth-generation chip, while the Dojo D3 training accelerator will abandon an SRAM-centric architecture in favor of chiplet design with stacked high-bandwidth memory.
Tesla also began replacing the term artificial intelligence with super intelligence, giving its AI accelerators the SI series naming. For the same reason, Elon Musk recently confirmed that SpaceX AI will be renamed SpaceX SI. The most notable change in this roadmap update may be that Tesla has restarted its in-house Dojo training accelerator program.
Musk cancelled the entire wafer-scale Dojo project in August 2025 and said in August 2026 that SpaceX and xAI would use only NVIDIA hardware across their ground and space data centers. The third-generation Dojo D3 will nevertheless differ clearly from the earlier D1 and D2. The roadmap also lists InFO_SoW training compute units built on TSMC 3nm and Samsung 2nm processes, with 440MB of SRAM per chip and 11GB of SRAM per wafer.