Elon Musk would not need Tesla to be only a carmaker. He desires Tesla to develop into a man-made intelligence firm, one that may make automobiles drive themselves.
Essential to this process is Dojo, Tesla’s custom-built supercomputer designed to coach its Full Self-Driving (FSD) neural community. FSD shouldn’t be really totally autonomous; it might carry out some autonomous duties however nonetheless requires an attentive driver. However Tesla believes that with extra knowledge, extra computing energy and extra coaching, it might cross the edge from virtually autonomous driving to completely autonomous driving.
That is the place Dojo is available in.
Musk has been teasing Dojo for a while, however the government has been ramping up discuss of supercomputers in 2024. Guarantee Tesla can obtain autonomous driving. Under is a timeline of Dojo mentions and commitments.
2019
First point out of Dojo
April twenty second – At Tesla’s Autonomous Driving Day, the automaker’s synthetic intelligence staff took to the stage to speak about Autopilot and Full Self-Driving, and the unreal intelligence that powers each. The corporate shared details about Tesla’s {custom} chips designed particularly for neural networks and self-driving automobiles.
Throughout the occasion, Musk teased Dojo and revealed that it’s a supercomputer used to coach synthetic intelligence. He additionally identified that each one Tesla automobiles produced at the moment can have all of the {hardware} required for full self-driving and can solely require a software program replace.
2020
Musk begins dojo street present
February 2nd – Musk stated that Tesla will quickly have greater than 1 million related automobiles world wide, geared up with the sensors and computing energy required for totally autonomous driving, and vigorously promoted Dojo’s features.
“Our coaching supercomputer Dojo will be capable of course of massive quantities of video coaching knowledge and effectively run hyperspace arrays with massive numbers of parameters, ample reminiscence, and ultra-high bandwidth between cores. Extra on this later.
August 14th – Musk reiterated Tesla’s plans to develop a neural community coaching pc referred to as Dojo “to deal with actually massive quantities of video knowledge,” calling it a “beast.” He additionally stated that the primary model of Dojo might be “a few 12 months away,” which might imply its launch date can be round August 2021.
December 31 — Elon stated Dojo shouldn’t be wanted, however it can make autonomous driving higher. “It is not sufficient to be safer than a human driver. Autopilot in the end must be 10 occasions safer than a human driver.”
2021
Tesla formally declares Dojo
August 19 – The automaker formally introduced Dojo at Tesla’s first-ever Synthetic Intelligence Day, an occasion designed to draw engineers to hitch Tesla’s synthetic intelligence staff. Tesla additionally unveiled the D1 chip, which the automaker stated it can use together with Nvidia’s GPUs to energy the Dojo supercomputer. Tesla famous that its synthetic intelligence cluster will home 3,000 D1 chips.
October 12 – Tesla releases one Dojo Know-how white paper, “Tesla Configurable Floating Level Format and Arithmetic Information.” The white paper outlines a technical commonplace for a brand new sort of binary floating level arithmetic for deep studying neural networks that may be carried out “solely in software program, solely in {hardware}, or any mixture of software program and {hardware}.”
2022
Tesla declares Dojo progress
August 12 – Musk stated Tesla will “progressively introduce Dojo. There might be no want to purchase as many incremental GPUs subsequent 12 months.
September 30 – At Tesla’s second AI Day, the corporate revealed that it had put in its first Dojo cupboard and carried out a 2.2-megawatt load take a look at. Tesla says it produces one tile (consisting of 25 D1 wafers) day-after-day. Tesla demonstrated Dojo on stage, operating a secure diffusion mannequin to create a man-made intelligence-generated “Cybertruck on Mars” picture.
Importantly, the corporate has set a goal date for completion of a full Exapod cluster within the first quarter of 2023 and stated it plans to construct a complete of seven Exapods in Palo Alto.
2023
A ‘long-term guess’‘
April 19 – Musk informed traders throughout Tesla’s first-quarter earnings name that Dojo “has the potential to extend the price of coaching by an order of magnitude” and “has the potential to develop into a marketable service that we might supply to different corporations in the identical approach.” “Providers” Amazon Net Providers offers community companies. “
Musk additionally famous that he “views Dojo as a low-risk guess” however a “guess value making.”
June 21—— The Tesla AI X account posted that the corporate’s neural community is already being utilized in buyer automobiles. The submit features a chart with a timeline of Tesla’s present and projected computing capabilities, which places Dojo’s manufacturing begin date in July 2023, though it is unclear whether or not this refers back to the D1 wafer or the supercomputer itself. Musk stated that on the identical day, Dojo was on-line and operating duties in Tesla’s knowledge heart.
The corporate additionally expects Tesla’s computing energy to be among the many high 5 on the planet by round February 2024 (there isn’t a indication that this plan is profitable), and by October 2024, Tesla’s computing energy might be 100 exaflops.
July 19 – Tesla famous in its second-quarter earnings report that it has begun manufacturing of Dojo. Musk additionally stated Tesla plans to spend greater than $1 billion on Dojo by 2024.
September 6 – Musk posted on X that Tesla is proscribed by synthetic intelligence coaching calculations, however Nvidia and Dojo will remedy this drawback. He stated this can be very troublesome to handle the roughly 160 billion frames of video knowledge Tesla obtains from its automobiles day-after-day.
2024
Plans to broaden
January twenty fourth – Throughout Tesla’s fourth quarter and full-year earnings name, Musk as soon as once more admitted that Dojo is a high-risk, high-reward venture. He additionally stated that Tesla is pursuing a “twin path of Nvidia and Dojo” and that “Dojo is taking part in a job” and is “doing coaching work.” He famous that Tesla is increasing and “plans to launch Dojo 1.5, Dojo 2, Dojo 3, and so on.”
January 26 – Tesla introduced plans to spend $500 million to construct the Dojo supercomputer in Buffalo. Subsequently, Musk downplayed the funding to a sure extent. He posted on X that though US$500 million is a large sum of money, it’s “solely equal to Nvidia’s 10k H100 system.” Tesla will spend greater than that on Nvidia {hardware} this 12 months. At present, staying aggressive within the area of synthetic intelligence requires not less than billions of {dollars} per 12 months.
April 30 – Based on IEEE, at TSMC’s North American Know-how Symposium, the corporate stated that Dojo’s next-generation coaching tile, D2, is already in manufacturing, putting all the Dojo tile on a single silicon wafer, quite than connecting 25 wafers. Create a tile.
Could 20 – Musk notes The again finish of the Texas Gigabit Manufacturing facility growth will embrace constructing “an ultra-dense water-cooled supercomputer cluster.”
June 4 – A CNBC report revealed that Musk moved 1000’s of Nvidia chips reserved for Tesla to X and xAI. After initially calling the report false, Musk posted on X As a result of ongoing development of the southern extension of Giga Texas, Tesla has no place to ship Nvidia chips to energy them up, “so they only sit within the warehouse.” He famous that the growth venture will “accommodate 50,000 H100s for FSD coaching.”
He additionally post:
“Of the roughly $10B of AI-related spending that I might say Tesla will do that 12 months, about half is inside spending, totally on Tesla-designed AI inference computer systems and sensors which are Testers are in all our automobiles, and Dojo. For constructing an AI coaching supercluster, NVidia {hardware} accounts for about 2/3 of the price. My present finest guess for Tesla to accumulate Nvidia this 12 months is $3 billion. $4 billion.
July 1 – Musk revealed on X that present Tesla automobiles could not have the {hardware} for the corporate’s next-generation synthetic intelligence mannequin. He stated that the variety of parameters of the following era of synthetic intelligence has elevated by about 5 occasions, “which is troublesome to attain with out upgrading the car inference pc.”
Nvidia provide challenges
July 23 – Throughout Tesla’s second-quarter earnings name, Musk stated demand for Nvidia {hardware} is “so excessive that it is typically troublesome to get GPUs.”
“So I believe it requires us to place extra effort into Dojo to verify we’ve the coaching capabilities we want,” Musk stated. “We do see a path to competing with Nvidia via Dojo.”
A chart on the Tesla investor platform predicts that Tesla AI coaching capability will improve from about 40,000 in June to about 90,000 H100 equal GPUs by the tip of 2024. Later that day, Musk posted on X that Dojo 1 will “obtain roughly 8k H100 equal on-line coaching by the tip of the 12 months.” He additionally posted photographs of the supercomputer, which seems to make use of the identical refrigerator-like chrome steel exterior because the Tesla Cybertruck.
XXX
July 30 – Musk stated in a press release that AI5 continues to be about 18 months away from mass manufacturing. reply Somebody’s submit claimed to type a membership, “Tesla HW4/AI4 homeowners are offended that AI5 was lagging behind within the launch.”
August third – Musk posted on X that he carried out a drill on the “Tesla supercomputing cluster situated in Giga Texas (often known as Cortex).” He famous that it will likely be composed of roughly 100,000 H100/H200 Nvidia GPUs, with “large storage for FSD and Optimus video coaching.”
