Tesla SI6 Chip Promises 80x the Power of AI4
Tesla renamed its AI chips to SI and says the SI6 chip will deliver 80 times the compute of AI4, with 144GB of memory and a 2029-2030 launch. Dojo 3 returns.
Tesla has published a new roadmap for its own AI chips, and the headline number is big. The company says its sixth-generation inference chip, SI6, will deliver 80 times the compute performance of the current AI4 chip in its cars. The roadmap also renames the whole chip line from AI to SI, short for super intelligence, and brings back the Dojo training chip that Tesla cancelled last year.
From AI to SI: a new name
The rename is simple. AI4, AI5 and AI6 become SI4, SI5 and SI6. The change follows Elon Musk renaming SpaceXAI to SpaceXSI, and Tesla AI has likewise become Tesla Super Intelligence, or Tesla SI. Tesla's owner manuals already use the new wording, calling the in-car computer the SI Computer instead of the FSD Computer. The new name covers the full chip line, from the hardware already in cars to the chips still years away.
SI6: 80 times the compute
SI6 is the big promise in the roadmap. Tesla says it will offer 80 times the compute performance of HW4, the current fourth-generation chip, and nine times its memory. The chip is expected to carry 144GB of memory, use a 2nm-class process at Samsung Foundry, and keep a multi-chip-module design. It is due in 2029-2030, and Tesla has a deal with Samsung Foundry to make its sixth-generation inference chips through 2033. Tom's Hardware estimates that 80 times HW4 works out to roughly 17,280 to 46,080 TOPS. That is only an estimate, because Tesla does not officially disclose the absolute compute, memory or power figures for HW4. The supported data formats for SI6 are also still unknown, so the real-world gap could look different from the headline number. Even so, 80x is the largest jump Tesla has ever promised between chip generations.
SI5 comes first
Before SI6 arrives, there is SI5. Tesla expects a 40-fold performance improvement over HW4 from this chip, driven by an eightfold increase in compute capability and a fivefold gain from silicon optimization. It is expected to use a multi-chip-module design, be made on 3nm-class process technology at TSMC and 2nm-class technology at Samsung Foundry, and carry 96GB of memory. Expected power consumption ranges from 250W to 800W depending on the application. Tom's Hardware estimates SI5 could deliver 8,640 to 23,040 TOPS.
Dojo 3 is back
The most surprising part of the roadmap is the return of Dojo. Musk cancelled the wafer-scale Dojo project in August 2025 and dissolved the team, saying Tesla would focus on its inference chips instead. Now the third-generation Dojo, D3, is back as the company's in-house training accelerator. D3 is a different design from the original. Where D1 and D2 relied on distributed on-chip SRAM and processors joined across a single wafer, D3 is a chiplet-based design that combines 2.5D and 3D packaging with stacked high-bandwidth memory, or HBM. Tesla has not disclosed D3's compute performance, memory capacity, manufacturing process or schedule. It says only that the chip is being optimized for "space, training, and scaling." That space mention matters: if some D3 chips are meant to run in orbit, they may need radiation hardening.
Why it matters
Tesla is building its own silicon for cars, the Optimus humanoid robot, data centers and now possibly space-based compute. Owning the chip stack means Tesla depends less on Nvidia and can tune hardware exactly to its own software. The roadmap also shows how fast the company's ambitions have grown: from a chip that drives a car to chips meant to train models and run them in orbit. The caution is the timeline. SI6 is three to four years away, D3 has no public schedule, and the biggest numbers are Tesla's own targets, not measured results. But as a statement of direction, the message is clear: Tesla wants to be a chip company as much as a car company.
Sources
- Tom's Hardware - https://tomshardware.com/tech-industry/artificial-intelligence/tesla-unveils-ai-accelerator-roadmap-adopts-new-si-branding-in-super-intelligence-push-si6-to-bring-80x-more-compute-performance-dojo-3-headed-to-space - Mezha - https://mezha.net/eng/news/mezha-4e7918314775f39f25c35fe2_tesla_outlines_ai/ - AlexTech - https://www.alextech.ai/en/news/tesla-renames-its-ai-chips-and-promises-80x-more-compute/