Tesla’s Dojo, a timeline | TechCrunch

Elon musk does not want Tesla to be just a car maker. She wants Tesla to be an AI company, discovering how to drive it.

That is important for that mission Dozo, Tesla Custom-Bilt Super Computer. Its complete self-driving (FSD) neural networks are designed for training. FSD is not really self-driving; It can perform some automatic driving tasks but still the need to focus on behind the wheel. However, Tesla thinks with more data, more calculated energy and more training, it can almost cross the edge from self-driving to complete self-driving.

And that’s where the dose comes.

The musk has been irritating the dose for some time, but the executive 2024 continues to discuss the super computers throughout the 2024. Now that we are in the 2025, another super computer chat we are in the 2025, but the importance of doses to Tesla can still exist – with EV – EV -sales reduction, investors want to assure that Tesla can gain autonomy. Below the dose has been mentioned and promised a timeline.

2019

Dozar

April 22 – This Tesla’s autonomy dayThe AI ​​team of the Automaker was on the stage to talk about autopilot and completely self-driving and strengthened the AI. The company shares information about Tesla’s custom-bilt chips, especially designed for neural network and self-driving vehicle.

During the event, the musk teases the dose, reveals that it is a super computer for AI training. He also mentions that all the Tesla cars produced at that time will have all the hardware needed for complete self-driving and only a software update was needed.

2020

Kasturi dose starts the roadshow

February 2 – Wicker Say Tesla will soon have more than one million connected vehicles worldwide, including sensors and calculations required for full self-driving-and touches the duo capacity.

“Our training will be able to process super computers, doses a lot of video training data and skillfully will be able to run hypersspace arrays between the core between the core between the lot of parameters, lots of memory and the ultra-hi bandwidth. More after this. “

August 14 – Repetition A neural network training plan called Tesla Dooso called computer development plans “called” an animal “for processing a huge amount of video data. He also says that the first version of Dozo is that “About a year away,” Which will keep the date of its introduction somewhere near August 2021.

December 31 – Elon says Does not need to do.But it will make self-driving better. “It’s not enough to be safe than human drivers, the Autopilot needs to be 10 times more safer than human drivers in the end.”

2021

Tesla makes Dozo official

August 19 – The Automaker officially announced the doso Tesla’s First AI DayAn event means to attract engineers to Tesla AI team. Tesla has also introduced her D1 chip, which automaker says that it will use it to get the doso super computer along the NVEDIA GPU. Note the Tesla will have a 3,000 d 1 chip in its AI cluster.

October 12 – Tesla has revealed Ay Doso technology Whitepaper“Tesla’s configurable floating point formats and a guide of arithmetic” “Whitepaper” WhitePaper gives a technical quality outline for new types of binary floating-pointed arithmetic and can be fully applied to software, completely hardware or software and hardware.

2022

Tesla reveals the progress of the dose

August 12 – Tesla will be called musk. “Phase on the dozo. The next year will not buy as many incremental GPUs. “

September 30 – At Tesler Second AI dayThe company has revealed that it was the first Dozo cabinet installed, testing 2.2 MW of load testing. Tesla says it was making a tile every day (which is formed by 25d 1 chip). Tesla Demos Dooso on the stage is running a stable expansion model to create an “Cybertrack on Mars” to create an AI-exposed image.

Importantly, the company sets a target date of a complete explood cluster to complete by Q1 2023 and says it has planned to create seven expans in Palo Alto.

2023

A ‘long shot bet‘

April 19 – The musk says investors in time Tesla’s first trimester earned This Dozo’s “There is the possibility of improving the expenditure of training” and “Amazon Web Services Web Services are also likely to be sold in the way we offer other companies.”

The musk further notes that he will see “the doso long shot bet,” but “but a” bet is the price “.

June 21 – Tesla AI X account Post The company that has neural networks already in customer vehicles. The thread includes a graph with a timeline of Tesla’s current and projected calculation power, which launches the doso production on July 2023, though it is not clear whether it refers to the D1 chips or supercomputer itself. Butter The same day the duo was already online and was working at the Tesla Data Center.

The company further projects that Tesla’s count will be the top five in the world by February 2024 (there is no indication that it was successful) and Tesla will reach 100 Xflop by October 2024.

July 19 – Tesla Note It has started producing doses in the second-third income report. Kasturi also said that Tesla plans to spend more than $ 1 billion at dows of dota by 2021.

September 6 – Musk At the x Tesla is limited by AI training calculations, but Nvidia and doso will fix it. He says that the Tesla video operates data from about 160 billion frame frames every day from his cars.

2024

Scale plan

24 January – During Tesla’s fourth-fourth and full-year income call, Kasturi again acknowledged that the dose is a high-risky, high-rise project. He also said that Tesla was following the “Dual Way of Nvidia and Dozo” that “dozo is working” and “training is working”. He noted that Tesla is scalling it and “Dooso 1.5, Doso 2, Dooso 3 and Plans for Whatties.”

26 January – Tesla announced plans to spend $ 500 million for a construction Doso super computer in buffaloThe The musk then puts the investment down somewhat, Post the X This is a huge digit of $ 500 million, “it is the equivalent of 10K H 100 system from Nvidia. Tesla will spend more than Nvidia hardware this year. The table parts to become competitive in AI this year are at least a few billion dollars per year”

April 30 – At the North American technology symposium of TSMC, the company says that the next generation training Tile-D2, which puts the entire doso tile on a single silicone weaver, is in production instead of attaching 25 chips to create a tile, according to the production. IEEE SpectrumThe

May 20 – Wicker Note The rear of the extension of the Giga Texas factory “will include construction of a super cube, water-winter super computer cluster.”

June 4- Ay CNBC Report Thousands of Nvidia chips stored in X and XAI for Tesla reveals diverted musk. The report was false after saying initially, Posted at the musk at X Due to the continuous construction of the southern expansion of Giga Texas, Tesla Nvidia Chips had no place to send them to turn them, “so they just sat in a warehouse.” He mentions that the extension will “House 50KH 100 House for FSD training”. ”

He Post::

“At 10B at the AI-related cost, I said that Tesla would make this year, about half the interior, mainly Tesla-designing AI inference computers and sensors are present in all our cars, plus doso. For the creation of AI training supercluster, the cost of the Nvidia hardware is about 2/3. My current best estimate for the purchase of Nvidia by Tesla is 3b to 4b this year. “

July 1 – The musk exposes to X That current Tesla vehicle may not have the right hardware for the company’s next-zenner AI model. He says that the parameter’s calculation with the next Jenner AI is about 5x increase “Vehicles assumption is very difficult to achieve without upgrading the computer.”

Challenges the Nvidia supply

July 23 – While calling for the second-third income of Tesla, Kasturi says that the demand for Nvidia hardware is “so high that GPUs are often difficult.”

“I think so we need to make much more efforts on the duo to ensure that we have the ability to training it,” Kasturi said. “And we can see the way to be competitive with Nvidia with the dose.”

A graph of Tesla’s Investor Deck has predicted that Tesla AI training capacity will ramp at about 90,000 H 100 equivalent GPU at the end of about 2024, more than 40,000 in June. Later that day X, Musk This Dozo 1’s “The end of the year will be equivalent to about 8k H 100-equivalent training online at the end of the year.” He posted Photo Super computer, which seems to use stainless steel exterior as the same refrigerator as a tesla cyber truck.

From the dozo to the cortex

July 30 – AI 5 is 18 months away from high-volume production, Kasturi says in one Answer “Tesla HW4/AI4 owners are angry about the AI5 when they come out of the way” in a post from someone who claims to start a club.

August 3- Posted at the musk at X That he made the “Giga Texas (alias Cortex) Tesla Supercomput Cluster”. He notes that it will be made with about 100,000 H 100/H 200 NVidia GPU “huge storage for FSD and Optimus video training”.

August 26 – Posted at the musk at X A video of Cortex, “He referred to” the new AI training supercluster built in Tesla HQ, Austin for the real-world AI solution. ”

2025

No updates on the dose in 2025

January 29 – Tesler Q4 and Full-Year 2024 Calls Call There is no mention of Dozo. Tesla’s new AI training supercluster cortex has created a presence at Austin Gigafactory. Tesla mentioned it Shareholder deck It has completed the cortex installation, which is made up of about 50,000 H 100 Nvidia GPU.

According to the letter, “Cortex helped to enable FSD (supervised) V13, which proudly proud of the great improvement in terms of safety and ease, 4.2X increase, higher resolution video inputs … are among the other growth,” according to the letter.

During the call, CFO Vaibhav Tenza mentioned that Tesla accelerated the Cortex buildout to speed up the Rolout out of FSD V3. He said that AI-related capital expenditure, including infrastructure, has been credited, “has been about $ 5 billion so far.” In 2025, Teneza said that he was hoping to be flat as he was related to AI.

This story was originally published on August 10, 2024 and we would update it with the development of new information.