News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
This one tile can do 9 quadrillion floating point calculations per second, optimized for AI training. I saw it up close today. Each 15kW tile holds a 5x5 array of blocks, each with 12 D1 multi-chip-module stacks, with RF

Tesla Dojo

Tesla's custom supercomputer and D1 chip for training self-driving vision models

Photo: Steve Jurvetson, CC BY 2.0, via Wikimedia Commons

Tesla Dojo is a series of supercomputers built by Tesla to process video from its cars and train the machine learning models behind Full Self-Driving. It entered production in July 2023 around the custom D1 chip, a 7 nm design with 50 billion transistors made by TSMC. The project was disbanded in 2025 and restarted in 2026.

Share on WhatsApp

  • D1 chip: 50 billion transistors, 645 mm2 die, TSMC 7 nm
  • 354 usable cores per D1, arranged in an 18×20 mesh on one die
  • Training Tile packs 25 D1 chips and reaches 9 petaflops at BF16/CFloat8
  • ExaPOD: 120 tiles, 1,062,000 cores, 1 exaflops at BF16/CFloat8
  • A 2022 test pulled 2.3 MW and tripped a San Jose power substation

What it is

Tesla Dojo is a family of supercomputers designed and built by Tesla to handle computer vision video processing and recognition. Its job is training the machine learning models that improve Full Self-Driving, the company's advanced driver-assistance system. The design grew out of a need to chew through millions of terabytes of real driving video collected from more than four million Tesla cars, so it departs sharply from conventional supercomputer layouts. Production use began in July 2023. At its heart sits the D1 chip, a Tesla-designed processor manufactured by TSMC on a 7 nm node with 50 billion transistors across a 645 mm2 die.

How it came about

Tesla first spoke of Dojo in April 2019 at its Autonomy Investor Day, and in August 2020 Elon Musk said it was roughly a year away, held back by power and cooling problems. The D1 chip and the plan for a datacenter of 3,000 of them were shown at AI Day on 19 August 2021, alongside a first completed Training Tile. A whitepaper that October set out the CFloat8 and CFloat16 number formats. At the September 2022 AI Day Tesla showed System Trays and a Cabinet, and reported a test that pulled 2.3 MW and tripped a San Jose substation. Production power-on came in August 2023.

Inside the chip

The building block is the D1 chip. Each of its 354 usable cores is a general-purpose 64-bit superscalar CPU with simultaneous multithreading, but no virtual memory and only limited memory protection; software manages resources instead. Instructions mix RISC-V with custom ones, and a vector unit handles machine learning work in up to 16 formats at once. Cores sit in an 18×20 mesh on the die, linked by an on-chip router that moves packets and 64-byte reads and writes every cycle. Each core owns 1.25 MB of SRAM, giving 440 MB per die. Twenty-five chips form a water-cooled Training Tile, six tiles a System Tray, and ten cabinets an ExaPOD.

How it was made

TSMC manufactures the D1 on a 7 nm process. The die measures 645 mm2 and carries 50 billion transistors, running at 2 GHz for 376 teraflops in BF16 or CFloat8, or 22 teraflops at FP32. Around its edges, 576 SerDes channels move 8 TB/sec between dies. A single chip draws about 400 W. Twenty-five chips are packed 5×5 into a water-cooled Training Tile drawing 15 kW at 52 volts, and tiles are stacked vertically in cabinets to shorten communication paths. An ExaPOD holds 120 tiles, 1,062,000 cores, 1.3 TB of SRAM and 13 TB of HBM.

Tesla Dojo
A drawing made by EveryChip from what this page says. It is a diagram, not a picture of the real chip.

Where it went

Dojo exists to train Tesla's neural networks on real driving footage. Before it, Tesla's GPU clusters processed a data set of one million ten-second clips from Autopilot cameras at 36 frames per second, carrying six billion object labels with depth and velocity, totalling 1.5 petabytes. Dojo was powered on for production in August 2023 alongside a 10,000 Nvidia H100 cluster. In June 2024 Musk described construction at Gigafactory Texas for a cluster mixing Tesla AI and Nvidia hardware, planned at 130 MW of thermal design power and later above 500 MW.

Versions and family

Dojo scales in fixed steps. Twenty-five D1 chips make a Training Tile, six tiles plus a host interface make a System Tray (previously called the Training Matrix), two trays make a Cabinet, and ten cabinets make an ExaPOD with 3,000 D1 chips. Dojo Interface Processor cards sit at tile edges, each carrying 32 GB of HBM2e or HBM3 and two I/O processors, plugging into PCI-Express 4.0 x16. Tesla Transport Protocol links run over Ethernet to cut mesh hops. A new chip iteration arrived when work restarted in January 2026.

Why it matters

Dojo pushed a car maker into building its own AI silicon and supercomputer rather than relying only on merchant GPUs. Its hierarchy of tiles, trays and cabinets, and its configurable floating point formats, showed how a training system could be shaped around video data instead of general workloads. Tesla's earlier GPU cluster, expanded to 7,360 GPUs by August 2022, was claimed in 2021 to sit near the top of the world's supercomputers at about 81.6 petaflops, though that figure was disputed over precision. Dojo was designed to pass an exaflop.

Problems and reception

Musk called Dojo in January 2024 a long shot worth taking because the payoff could be very high, while rating its success as unlikely. A September 2022 test drew 2.3 MW and tripped a local San Jose power substation. Bloomberg News reported in August 2025 that the project had been disbanded, and it was restarted in January 2026 with a new chip iteration.

Questions people ask about Tesla Dojo

What is Tesla Dojo for?

It trains the machine learning models behind Tesla's Full Self-Driving system. The work is computer vision: processing and recognising video from Autopilot cameras in millions of Tesla cars so the software better understands roads.

What is the D1 chip?

D1 is the custom processor at the centre of Dojo, designed by a Tesla team under ex-AMD designer Ganesh Venkataramanan and built by TSMC on a 7 nm node. It holds 50 billion transistors on a 645 mm2 die and runs at 2 GHz.

How does Dojo scale up?

Twenty-five D1 chips form a water-cooled Training Tile. Six tiles plus a host interface make a System Tray, two trays make a Cabinet, and ten cabinets form an ExaPOD with 3,000 chips and 1,062,000 cores, rated at 1 exaflops.

What are CFloat8 and CFloat16?

They are Tesla's configurable floating point formats, described in an October 2021 whitepaper as an extension of IEEE 754. The compiler can trade mantissa and exponent precision for faster vector work and smaller storage, suiting AI training better than plain FP32 or FP16.

Is the project still running?

Bloomberg News reported in August 2025 that Dojo had been disbanded. Work restarted in January 2026 with a new chip iteration.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Tesla Dojo", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

Photo credits on this page

The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.

The weekly digest

Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.