Skip to content
Catalyst Note CN markCNCatalystNote

Noting the changes that set others in motion.

AI hardware

Tesla outlines Dojo D3 chiplet design and SI6 inference roadmap

Tom's Hardware reports Tesla is shifting from wafer-scale SRAM to stacked HBM for training chips while renaming accelerators to super intelligence.

Tesla outlines Dojo D3 chiplet design and SI6 inference roadmap
Depth
1 outlet · newest source 239h ago

Key points

  • Tesla plans a Dojo D3 training accelerator using chiplets and stacked HBM instead of SRAM.
  • The upcoming SI6 inference chip aims for 80 times the performance of the fourth-generation model.
  • The company has adopted "super intelligence" branding for its AI hardware and related entities.

Tom's Hardware reported that Tesla has disclosed a revamped roadmap for its custom AI accelerators, introducing new branding and architectural shifts for both training and inference silicon.

Evidence: reportedTesla unveils AI accelerator roadmap, adopts new 'SI' branding in 'Super Intelligence' push

aifleet-tomshardware

The claim

Tesla expects its sixth-generation inference accelerator, designated SI6, to deliver 80 times more compute performance than the existing fourth-generation chip. The company also described its next-generation Dojo D3 training accelerator as a chiplet-based design that combines 2.5D and 3D packaging with stacked high-bandwidth memory.

Evidence: reportedTesla unveils AI accelerator roadmap, adopts new 'SI' branding in 'Super Intelligence' push

aifleet-tomshardware

Evidence: reportedTesla unveils AI accelerator roadmap, adopts new 'SI' branding in 'Super Intelligence' push

aifleet-tomshardware

The response

The report noted that this marks a re-emergence of the in-house Dojo project after Elon Musk previously cancelled the wafer-level Dojo initiative in August 2025 and stated that SpaceX and xAI would exclusively use Nvidia hardware.

Evidence: reportedTesla unveils AI accelerator roadmap, adopts new 'SI' branding in 'Super Intelligence' push

aifleet-tomshardware

Why this matters now

The adoption of HBM suggests Tesla intends to substantially increase memory capacity for modern AI models that require large amounts of storage for parameters and activations, moving away from the capacity limitations of SRAM-centric architectures.

What buyers should weigh

Enterprise IT buyers should note that the shift to HBM introduces different latency, bandwidth, packaging and power trade-offs compared to previous designs. The report states it remains unclear whether Tesla has abandoned wafer-scale integration altogether, as stacking custom HBM on logic dies using wafer-scale packaging remains a technical possibility.

Why it matters

  • The move to chiplets and HBM may alter supply chain dependencies for organizations considering custom silicon alternatives to standard GPU offerings.
  • The 80X performance claim for SI6 inference hardware could impact cost-per-inference calculations for large-scale AI deployment.
  • The renaming of AI accelerators to "super intelligence" and SpaceXAI to SpaceXSI indicates a broader branding strategy that may affect vendor identification in procurement.
  • Uncertainty around wafer-scale integration abandonment leaves open questions about the long-term viability of Tesla's custom training infrastructure.

Topics: Tesla, AI accelerators, chiplets, HBM, custom silicon

Updated

Then / now / next

Then

Tesla relied on a wafer-scale system-on-wafer architecture with SRAM for its Dojo D1 training accelerator.

Now

The company is developing Dojo D3 with chiplets and stacked HBM while rebranding AI hardware as super intelligence.

Next

It remains unclear whether Tesla will fully abandon wafer-scale integration or continue using it for select use cases.

Story timeline

  1. Then Tesla relied on a wafer-scale system-on-wafer architecture with SRAM for its Dojo D1 training accelerator.
  2. 06 Nov 2026 Tesla unveils AI accelerator roadmap, adopts new 'SI' branding in 'Super Intelligence' push
  3. Now The company is developing Dojo D3 with chiplets and stacked HBM while rebranding AI hardware as super intelligence.
  4. Next It remains unclear whether Tesla will fully abandon wafer-scale integration or continue using it for select use cases.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Terms in this story

HBM

High-bandwidth memory, a type of memory packaging that stacks multiple memory chips to increase bandwidth and capacity.

Chiplet

A small semiconductor die that can be combined with other chiplets to form a larger package, offering modular design benefits.

SRAM

Static random-access memory, a type of memory that retains data as long as power is supplied, often used for cache.

Wafer-scale integration

A technique that uses an entire semiconductor wafer as a single chip, rather than cutting it into individual dies.

Questions readers ask

What is the new name for Tesla's AI accelerators?

Tesla has adopted the term "super intelligence" and renamed its AI accelerators with SI monikers, such as SI6.

How does Dojo D3 differ from Dojo D1?

Dojo D3 uses a chiplet-based design with stacked HBM, whereas Dojo D1 relied on a system-on-wafer architecture with distributed on-chip SRAM.

Did Tesla cancel its custom silicon efforts?

Elon Musk cancelled the wafer-level Dojo project in August 2025, but the new roadmap indicates a continuation with a different architectural approach.

Reading receipt

  • Tesla plans a Dojo D3 training accelerator using chiplets and stacked HBM instead of SRAM.
  • The upcoming SI6 inference chip aims for 80 times the performance of the fourth-generation model.
  • The company has adopted "super intelligence" branding for its AI hardware and related entities.

Still open: Has Tesla completely abandoned wafer-scale integration for all future projects?

Still open: Will the new SI6 inference chip meet the claimed 80X performance increase in real-world deployments?

Time spent: 0:00 · Depth: Full

Sources

  1. Tom's Hardware

Related stories

Background reading