Topics
Latest
AI
Amazon
Image Credits:Bryce Durbin | TechCrunch
Apps
Biotech & Health
Climate
Image Credits:Bryce Durbin | TechCrunch
Cloud Computing
Commerce Department
Crypto
Image Credits:SUZANNE CORDEIRO/AFP via Getty Images / Getty Images
Enterprise
EVs
Fintech
Ganesh Venkataramanan, former senior director of Autopilot hardware, presenting the D1 training tile at Tesla’s 2021 AI Day.Image Credits:Tesla/screenshot of streamed event
fund-raise
Gadgets
Gaming
Tesla’s humanoid robot Optimus Prime II at WAIC in Shanghai, China, on July 7, 2024.Image Credits:Costfoto/NurPhoto / Getty Images
Government & Policy
Hardware
Nvidia CEO Jensen Huang and Tesla CEO Elon Musk at the GPU Technology Conference in San Jose, California.Image Credits:Kim Kulish/Corbis via Getty Images / Getty Images
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
societal
Space
Startups
TikTok
Transportation
Venture
More from TechCrunch
event
Startup Battlefield
StrictlyVC
Podcasts
television
Partner Content
TechCrunch Brand Studio
Crunchboard
touch Us
For years , Elon Musk has babble out about Dojo — the AI supercomputer that will be the cornerstone of Tesla ’s AI ambitions . It ’s important enough to Musk that in July 2024 , he said the caller ’s AI squad would “ double down ” on Dojo in the jumper lead - up toTesla ’s robotaxi reveal , which happened in October .
But what exactly is Dojo ? And why is it so critical to Tesla ’s long - terminal figure strategy ?
In short : Dojo is Tesla ’s custom - built supercomputer that ’s design to civilise its “ Full ego - Driving ” neuronic networks . Beefing up Dojo run low deal - in - hand with Tesla ’s goal to touch full ego - driving and bring a robotaxi to market . FSD , which is on hundred of thousands of Tesla vehicles today , can perform some automated driving undertaking but still command a homo to be paying attention behind the wheel .
Tesla ’s Cybercab break has come and gone , and now the party is gearing up to launch an sovereign ride - hail service using its own fleet of vehiclesin Austin this June . Tesla also said during its2024 fourth - fourth and full - year earnings callat the end of January that it plan to launch unsupervised FSD for U.S. customer in 2025 .
Musk ’s previous empty talk has been that Dojo would be the key to achieving Tesla ’s goal of full ego - driving . Now that Tesla appears close to nearing that destination , Musk has been Chrysanthemum morifolium on Dojo .
alternatively , ever since August 2024 , talk has been aroundCortex , Tesla ’s “ jumbo Modern AI breeding supercluster being build up at Tesla HQ in Austin to solve existent - world AI.”Musk has also saidit will have “ monolithic warehousing for video training of FSD & Optimus . ”
InTesla ’s Q4 shareholder pack of cards , the caller shared update on Cortex , but nothing on Dojo .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
Tesla has pose itself to spend big on AI and Dojo — and now Cortex — to reach its destination of autonomy for both cars and humanoid automaton . And Tesla ’s next success really hinges on its ability to nail this down , given the increased competition in the EV grocery store . So it ’s deserving taking a tight look at Dojo , Cortex , and where it all stands today .
Tesla’s Dojo backstory
Musk does n’t want Tesla to be just an automaker , or even a purveyor of solar panel and energy storage systems . or else , he wants Tesla to be an AI troupe , one that has cracked the code to self - driving car by mime human perception .
Most other companies build self-reliant fomite technology swear on a combination of sensors to comprehend the world — like lidar , radio detection and ranging and cameras — as well as high - definition maps to localise the vehicle . Tesla believes it can accomplish to the full autonomous drive by swear on cameras alone to capture visual data and then habituate in advance neuronic networks to process that data and make quick conclusion about how the automobile should behave .
As Tesla ’s former promontory of AI , Andrej Karpathy , said at the automaker ’s firstAI Day in 2021 , the company is basically trying to build “ a semisynthetic animal from the undercoat up . ” ( Musk had been teasing Dojo since 2019 , but Tesla formally announced it at AI Day . )
Companies like Alphabet ’s Waymo havecommercialized Level 4 self-reliant vehicles — which the SAE defines as a arrangement that can take itself without the need for human intervention under certain conditions — through a more traditional sensor and machine learning approach . Tesla has still yet to produce an sovereign system that does n’t necessitate a human behind the wheel .
About 1.8 million masses have paid the heftysubscription pricefor Tesla ’s FSD , which currently costs $ 8,000 and has beenpriced as high as $ 15,000 . The pitch is that Dojo - trained AI software package will eventually be advertise out to Tesla client via over - the - air update . The scale of FSD also mean Tesla has been capable to rake in millions of miles worth of television footage that it uses to train FSD . The idea there is that the more data point Tesla can accumulate , the nigher the automaker can get to actually achieving full ego - driving .
However , some industry experts say there might be a limit to the brute personnel approach of flip more data at a model and anticipate it to get smarter .
“ First of all , there ’s an economical restraint , and soon it will just get too expensive to do that , ” Anand Raghunathan , Purdue University ’s Silicon Valley professor of electrical and computer engineering , told TechCrunch . Further , he say , “ Some people lay claim that we might actually hunt out of meaningful information to train the models on . More datum does n’t inevitably mean more data , so it reckon on whether that information has entropy that is utilitarian to make a better model , and if the training process is able to actually distill that entropy into a just model . ”
Raghunathan say despite these dubiousness , the vogue of more data appear to be here for the short - terminal figure at least . And more data mean more compute power demand to store and process it all to train Tesla ’s AI models . That is where Dojo , the supercomputer , come in .
What is a supercomputer?
Dojo is Tesla ’s supercomputer system that ’s designed to officiate as a training ground for AI , specifically FSD . The name is a nod to the space where martial artistic creation are practiced .
A supercomputer is made up of thousands of smaller computers called node . Each of those node has its own CPU ( central processing unit of measurement ) and GPU ( computer graphic processing unit ) . The former deal overall direction of the node , and the latter does the complex stuff , like splitting undertaking into multiple parts and work on them simultaneously . GPUs are essential for political machine learning operations like those that world power FSD grooming in simulation . They also power large linguistic process good example , which is why the rise of generative AI has made Nvidia the most valuable caller on the planet .
Even Tesla buys Nvidia GPUs to train its AI ( more on that by and by ) .
Why does Tesla need a supercomputer?
Tesla ’s vision - only approach is the chief reason Tesla need a supercomputer . The neuronic networks behind FSD are trained on vast measure of driving datum to recognise and classify physical object around the fomite and then make driving decision . That means that when FSD is engaged , the neuronal nets have to collect and process visual information continuously at amphetamine that match the depth and speed acknowledgement potentiality of a homo .
In other password , Tesla means to create a digital extra of the human ocular cortex and wit function .
To get there , Tesla need to lay in and work all the video information collected from its cars around the world and feed millions of simulations to take aim its poser on the data .
Dojo picspic.twitter.com/Lu8YiZXo8c
Tesla appears to rely on Nvidia to power its current Dojo training data processor , but it does n’t require to have all its eggs in one basket — not least because Nvidia chips are expensive . Tesla also hop to make something honorable that increase bandwidth and lessen response time . That ’s why the carmaker ’s AI division decided to issue forth up with its own custom computer hardware programme that aims to educate AI models more efficiently than traditional systems .
At that program ’s core is Tesla ’s proprietary D1 chips , which the troupe state are optimise for AI workload .
Tell me more about these chips
Tesla is of a similar thought to Apple in that it believes hardware and software should be plan to work together . That ’s why Tesla is work out to move away from the standard GPU ironware anddesign its own chipsto power Dojo .
Tesla unveil its D1 potato chip , a atomic number 14 straight the size of a palm , on AI Day in 2021 . The D1 check entered into product as of at least May this class . The Taiwan Semiconductor Manufacturing Company ( TSMC ) is make up the potato chip using 7 nanometer semiconductor lymph gland . The D1 has 50 billion transistors and a large die size of 645 millimeters feather , according to Tesla . This is all to say that the D1 forebode to be exceedingly brawny and effective and to handle complex tasks cursorily .
“ We can do compute and data point transfer simultaneously , and our usance ISA , which is the instruction set architecture , is amply optimise for machine learning workloads , ” said Ganesh Venkataramanan , former senior director of Autopilot hardware , at Tesla ’s 2021 AI Day . “ This is a pure machine learning . ”
The D1 is still not as powerful as Nvidia ’s A100 chip , though , which is also manufacture by TSMC using a 7 nanometre unconscious process . The A100 contains 54 billion transistor and has a die size of 826 square mm , so it performs more or less better than Tesla ’s D1 .
To get a high bandwidth and higher compute power , Tesla ’s AI team fused 25 D1 chips together into one roofing tile to function as a unified computer arrangement . Each tile has a compute tycoon of 9 petaflops and 36 TB per second of bandwidth , and contains all the hardware necessary for might , cooling and data transfer . you may imagine of the tile as a ego - sufficient computer made up of 25 lowly computers . Six of those tiles make up one rack , and two rack make up a console . Ten cabinets make up an ExaPOD . At AI Day 2022 , Tesla said Dojo would surmount by deploying multiple ExaPODs . All of this together makes up the supercomputer .
Tesla is also working on a next - gen D2 cow chip that aims to solve info flow bottlenecks . Instead of link the individual chips , the D2 would put the full Dojo tile onto a single wafer of silicon .
Tesla has n’t confirmed how many D1 Saratoga chip it has ordered or carry to obtain . The company also has n’t provided a timeline for how long it will take to get Dojo supercomputers run on D1 chips .
In answer toa June post on Xthat said : “ Elon is build a elephantine GPU cooler in Texas , ” Musk replied that Tesla was aiming for “ half Tesla AI ironware , half Nvidia / other ” over the next 18 month or so . The “ other ” could be AMD microprocessor chip , perMusk ’s remark in January .
What does Dojo mean for Tesla?
Taking control of its own chip yield imply that Tesla might one 24-hour interval be able to apace add large sum of compute world power to AI training political program at a low toll , particularly as Tesla and TSMC scale up micro chip production .
It also signify that Tesla may not have to rely on Nvidia ’s chipping in the future , which are increasingly expensive and hard to secure .
During Tesla ’s 2nd - quarter earnings call , Musk said that demand for Nvidia hardware is “ so eminent that it ’s often hard to get the GPUs . ” He say he was “ quite interested about actually being capable to get steady GPUs when we want them , and I think this therefore requires that we put a good deal more effort on Dojo in club to ensure that we ’ve got the education capability that we need . ”
That say , Tesla is still buying Nvidia chips today to train its AI . In June , Musk post on X :
Of the roughly $ 10B in AI - related expenditures I enjoin Tesla would make this class , about one-half is home , primarily the Tesla - design AI illation computer and detector present in all of our cars , plus Dojo . For building the AI training superclusters , Nvidia hardware is about 2/3 of the price . My current undecomposed guess for Nvidia purchase by Tesla are $ 3B to $ 4B this class .
“ Inference compute ” mention to the AI computations do by Tesla cars in real time and is separate from the training compute that Dojo is responsible for .
Dojo is a risky stakes , one that Musk has hedged several times by saying that Tesla might not succeed .
In the foresighted discharge , Tesla could theoretically create a young business model ground on its AI division . Musk has said that the first version of Dojo will be tailored for Tesla computer vision labeling and training , which is great for FSD and for trainingOptimus , Tesla ’s humanoid robot . But it would n’t be utile for much else .
Musk has saidthat next versions of Dojo will be more bespoken to general - aim AI training . One potential problem with that is almost all AI software package out there has been written to work with GPUs . Using Dojo to condition worldwide - purpose AI models would require rewriting the package .
That is , unless Tesla rents out its compute , similar to how AWS and cerulean lease out cloud computing capacity . Musk also remark during Q2 earnings that he sees “ a way of life to being private-enterprise with Nvidia with Dojo . ”
A September 2023 news report from Morgan Stanley predicted that Dojo couldadd $ 500 billionto Tesla ’s market value by unlocking young tax income stream in the form of robotaxis and package service .
In short , Dojo ’s microchip are an insurance policy insurance policy for the car maker , but one that could pay dividend .
How far along is Dojo?
Reuters reportedlast yr that Tesla began output on Dojo in July 2023 , but aJune 2023 postfrom Musk suggest that Dojo had been “ online and running useful tasks for a few month . ”
Around the same prison term , Tesla said it expected Dojo to be one of the top five most hefty supercomputers by February 2024 — a effort that has yet to be in public disclosed , leaving us doubtful that it has occurred .
The companionship also said it carry Dojo ’s total compute to reach 100 exaflops in October 2024 . ( One exaflops is adequate to 1 quintillion computing gadget surgical process per sec . To strive 100 exaflops , and adopt that one D1 can accomplish 362 teraflops , Tesla would take more than 276,000 D1s , or around 320,500 Nvidia A100 GPUs . )
Tesla also pledge in January 2024 tospend $ 500 millionto build a Dojo supercomputer at its gigafactory in Buffalo , New York .
In May 2024,Musk notedthat the rear portion of Tesla ’s Austin gigafactory will be book for a “ super thick , water supply - cool down supercomputer cluster . ” Now we know that it ’s in reality Cortex , not Dojo , that is taking up that distance in Austin .
Just after Tesla ’s second - twenty-five percent net income call , Muskposted on Xthat the car manufacturer ’s AI squad is using Tesla HW4 AI computer ( renamed AI4 ) , which is the hardware that lives on Tesla vehicle , in the preparation loop with Nvidia GPUs . He noted that the breakdown is roughly 90,000 Nvidia H100s plus 40,000 AI4 information processing system .
“ And Dojo 1 will have roughly 8k H100 - equivalent of preparation online by terminal of yr , ” he continued . “ Not monolithic , but not superficial either . ”
Tesla has n’t provide updates as to whether it has gotten those chips online and running Dojo . During the society ’s 4th - quarter 2024 earnings call , no one cite Dojo . However , Tesla say it discharge the deployment of Cortex in Q4 and that it was Cortex that helped enable V13 of supervised FSD .
This news report originally publish August 3 , 2024 , and we will update it as new information develops .
