Humanoid & robot atlas
Dojo D1
Photo: Steve Jurvetson, CC BY 2.0, via Wikimedia Commons
AI acceleratorsUnited States

Tesla · ai accelerator

Dojo D1

Tesla's training chip for its Dojo supercomputer (50 billion transistors). The project was reported disbanded in August 2025 and restarted in 2026; it trains models rather than running in robots.

Compute
354 custom cores per die; training chip
Power
about 400 W
Process
7 nmTransistor node
Made by
TSMCFoundry
First year
2021
Price
—Not published

Robots that use it

No robot in the atlas is publicly named as using this chip yet.

The Tesla family

Every Tesla chip in the atlas, oldest first.

  1. —AI6Not published
  2. 2019FSD Chip (Hardware 3)72 TOPS2
  3. 2021Dojo D1354 custom cores per die; training chip
  4. 2023AI4 / FSD Computer 2 (Hardware 4)Not published1
  5. 2027AI5Not published (Tesla claims about 10x AI4)1

About Tesla

United States · chip designer

Designs its own self-driving chips (FSD chip in HW3, AI4 in HW4, AI5 next) and the Dojo D1 training chip; Samsung and TSMC make them. The same compute family runs the Optimus humanoid.

Sources

Specs come from the maker and reference pages.

Tesla / Hot Chips Tesla Transport Protocol over Ethernet (TTPoE) — D1 architecture

Statistics only, not advice. No affiliation, sponsorship or endorsement. How we work · Terms