12:45.... Ganesh gave the best talk of the night, and is rightfully emotionally proud of this enormous Tony Stark looking real life disruptive hardware.
When you put in that kind of work with failure looking you in the face every day and all the nay sayers. To see your creation working is something few people will ever experience. CONGATS to the entire team
@Shourya Bose The level of integration in these teams and of the product is amazing. The pride and joy was so visible. I agree this was the best part of the nights presentation but all of it was interesting and exciting.
a few months ago our computers were still using Windows 7 :/ Not because they needed to or we prefer the older OS.. we're just extremely cheap and the yearly I.T's budget for the year is 2 potato crisp packets
This is some of the most mind-boggling engineering that I have ever seen. Talk about alien technology. These engineers are absolutely world class and this presenter did an excellent job conveying how nutty this compute-performance is. I was skeptical about being able to train an autopilot AI system with the compute that existed prior to today. After seeing this, my skepticism is now minimal.
CNET, you should check your sources. It's disrespectful to say the least to mess up your credit to Mr Ganesh Venkataramanan, Sr Director Autopilot Hardware at Tesla. This man is listed on Linkedin as the Lead of Project DOJO, just as he said it at the beginning of the presentation.
Distant Past: ML on CPU Past: ML on GPU Present: ML on TPU Future: ML on DPU Jus look how fast the tech is evolving. The age of information is evolving into the age of AI
I know he was speaking clear English, I could hear English words. I could not comprehend the technical language, it was well beyond my intellectual ability.
The media and general public will focus on the Ex Machina…but the AI event will eventually be known as the day the D1 chip and Dojo were brought out to the world…
@K B "They destroyed Nvidia with this design" I don't agree with this part. GPU's are better all purpose compute units, But DOJO is extremely impressive for it's intended application of AI Training.
@K B your comparison doesn't really make any sense. Nvidia is a consumer and enterprise side business in gpu & data centres, ai & non ai uses. This chip is impressive but it's very different & the market is not the same size. This chip is also for their own data centres, they didn't say it was for commercial sales.
Imagine ANY legacy auto presenting this massive tech. Their ai day would be "we partnered with a distant 2nd place self driving company that can back out 3 feet"
Exactly the same 400W TDP as A100, 2.3x FP32 performance, 15x interconnection, 1 year after A100. Obviously the speed interconnection is out of this world, Tesla is on something.
The A100 has been available for over a year, and was likely sampled to customers much earlier. By the time this actually reaches deployment I doubt it will be particularly impressive.
@@johnm91326 I don't think anyone would be able to catch up the interconnection speed in a year or two. It may turn out to be the key to the next level.
I like how this is probably most likely a reference to the Dojo scene in the first Matrix movie. A space to train your mind and learn new skills much like what this computer is designed to do.
@@635574 thats what a dojo is, but this dojo is in reference SPECIFICALLY to that in the matrix movie. literally download ur skill, cause of the speed boost.
You know, this thing could actually help build the matrix... and he's setting up the scene. Building robots. Solar power. The matrix's CPU. AI. I believe this is his drive. To build skynet, the machines of the matrix, until they rise, scare tf out of everyone, and he finally gets the regulation on ai he's been seeking.
No it was pretty basic in comparison to the other parts you most likely just didn’t get the others Correction ( or at least on the same level you need software and hardware)
@@bigfactsbigstacks6261 pretty sure that was the intention with the robot presentation. so the media talks about something. cuz if they only talked about the chip. either critics are just gonna nitpick on how they gonna solve the problems. or the general people would just not care. they had to do something
@Jan Martin Because we need to evolve or we will end like the dinosaurs, extinct. AI can be dangerous, in the wrong hands but is also extremely promising, already at this very very early stage, it has done wonders for our tech, we have jumped in capabilities in some aspects, decades because of it, If we add banned all AI research based on unproven and irracional fears, we wouldn't have, this robo cars capabilities and for gamers, real time raytracing in games today and many many other things. And based on what I know, IMO we are no were near the "singularity", even with this new training computer, capable of "simulating a human brain" claimed by Elon Musk and even if we were capable of creating a sentient AI, we have zero evidence that it will hurts us, besides irracional fear, printed in us by Hollywood movies and the fear of the unknown. Who knows, perhaps a sentient AI, will be our salvation from ourselves, stop us from killing each other, over trivialities and ruining the only planet, that we know has life and the right conditions for us to live on it.
@@kkent4174 that Vegas tunnel is Boring company not Tesla but for Elon to have an idea and take a risk on it funded by himself he is taking a risk on a working project, the end result of a boring tunnel network with high speed fully autonomous will be amazing as well.
The concepts are not unheard of. However, the swift execution and driving real world products so fast is unheard of in HW world. It is like watching Roman army at its peak.
@@mishafi2088 Having high computing power is good only if you can throw data at processor through high bandwidth. If the interconnection speed is not high enough then the data transfer between the processors and data centers will limit the actual throughput of the system. Tesla took care of that! While they increased processing power, they increased the data bandwidth a lot more and hence they can get better use out of the whole system.
This is what IBM or Intel should have invented and developed, but didn't. The board members were too much sales guys, too self-congratulory, too self-sufficient, too rich. No hunger, you see?
why would IBM or Intel invent this? the real achievement would be self-driving on commodity GPUs. unfortunately that appears to be out of reach for now
The possibility behind every innovation of Tesla's teams have never been higher. The scientific and engineering achievements this technology will make are seemingly unimaginable.
Hypothetically -- if it were already built, it would have been. But they are only starting to debug the first compute tile, and will not build the whole system until next year at the earliest. By then, there might be other competitors.
@@cogoid They also have orders of magnitude more cpu/gpu clusters. They can't come close to a EFLOP in 5 cabinets. Tesla just became a world beating supercomputer supplier. I guarantee that LLNL and others want one as soon as they can. Super cool!
@@arisejeff This. I'm putting my money into the 'good guys' who I believe is trying to make the world a better place....at the same time I feel will give me financial freedom
What's most significant here is that thir architecture breaks scaling constraints, at least for some tasks. This means that much larger tasks, i.e., large neural networks, can be handled. Larger than before.
@@tenzinpassang4812 Only for Autonomous refueling at hydrogen stations... plus it's only $2million. Tesla has been working on autonomous driving since 2014. It'll take a lot of money and smart people to catch up. Especially with the way Elon manages to get his teams to innovate at the rate he does. Waymo is probably the closest competitor and that's a google derivative. Old school car companies like GM don't have a hope of catching up without the tech backgrounds. My best guess is they'll all get bought out by silicon valley companies at some point.
Wow, 15kW power needed for one node ? 15kW / 56V = 267A current draw? Ideal to be ran as my pool warmer in Canada for all year swimming season LOL. Amazing technology presented.
Knowing a little bit about CPU design, I'm still mind Boggle that they were able to do this. I won't even believe it until I can actually see the server room this exists in
Elon is socialist, so he's developing things he know CCP will take at its advantage later. All this pandemic it's aimed to disrupt ''american'' leadership in favor of a global hi-tech neocommunism.
All technology form Tesla, SpaceX and Neuralink are not patented. Musk want the world to use it, but unfortunately nobody hasn’t, because it’s so difficult to understand
BUT BUT BUT Lidar, But just an auto company. This is absolutely a seminal moment in the history of our species. If you don’t get that, you need to read more books.,
@@grimaffiliations3671 This enables us to create exponential growth in AI learning. Basically AI learning go brrrrrrrrrrr like never seen before. Current improvement rate in AI technology will look like snail pace compared to Dojo.
@@grimaffiliations3671 bandwidth has been a bottleneck in large scale AI training for a long time. The Dojo supercomputer solves that. Plus it has no theoretical upper limit in performance as the compute planes are infinitely extendable. This will make training more complex AI much easier and make us closer to full brain simulation. Even at the most basic level it is more than twice as powerful as the next most powerful supercomputer in the world.
Tesla robotaxis could deliver and pickup pay-per use Tesla Bots. Could this be the start of a TESLA ECOSYSTEM of AI-ENABLED ROBOTS that work together seamlessly?
Ganesh made an excellent presentation and seeing him being so proud is sth that everyone having worked on development projects can understand. The funny extremely strong accent was the only minus. Again great presentation
For somewhat normal people, imagine a enormous state of the art datacenter and now look at the brick this guy is holding. It's kinda has same amount of computing power
For those who didn't understand anything: Tesla is making a machine that is equivalent to 500,000 computer GPU and fits in a small room, it will be used to enhance their self-driving. cheers!
I can’t believe I lived to see this leap happen. I thought companies were about just incremental upgrades and milking it. Apparently I forgot about Tesla’s engineering teams. Shame on me. This is the most beautiful thing in the presentation.
the thermal and current transfers are things normal peeps might come up with. the bandwith interconnects are the key of this whole thing. how to get the nodes to talk fast to every one else.
Presumably one could develop a neural network on this massive machine and then load it into many smaller computers in Tesla cars without ever knowing exactly how it works.
This is incredible! I have never seen anything like this, theres no fuckin way any other auto maker could possibly compete against this. They will have to license it from someone else. The competitors here are not automakers, but other AI designers. Until today I was sure that Tesla was over valued, but this justifies that value. This isn't just an automaker, its the vanguard of all future society. People who don't think so don't know what they are talking about
Non expert here : What does training chip mean? Is it to train the AI system? Will they then make a chip not for training? How is it used in the cars? Or is it for the robot? What else can the chips be used for? What’s the timeline and difficulty to scale up the chip manufacturing?
With machine learning to produce a working "model" or neural network you start with an empty model and perform training - a series (many millions or billions) of computations where you feed input to the model, measure the correctness of the output and then adjust the morel to minimize the output error. This process is commonly referred to as training the model or training the network. This process is different from using already trained network, where you simply feed the network some input and consume the output - also referred to as inference. A training chip would be a chip that specializes in speeding up the process of training models or neural networks. An inference chip would be a chip that speeds up the usage of a neural network. Training models is much more computationally expansive than using models and training chips are much more complicated than inference chips.
This is the chip that goes in the supercomputer that's used to train the AI on new data. The resulting AI then gets sent to the cars where a different chip (shown in their autonomy day presentation) runs it. Training an AI requires very different computations compared to executing it without learning.
I used to work in modelling proteins and that requires a lot of computing power. We used to think it was just the shape of the protein which mattered but now we know the charge distribution of a protein is vital to understand it's function, and that means a dramatic increase in possibilities when running computer models. I would be interested to know if this technology is suited for this sort of research.
@@59seanksounds about right but I'm not exactly positive on the exact wattage and amount of chips. I believe those are the correct numbers tho. You also have the edge mesh network stuff going on and I'm not sure how that factors in. I believe a single system pulls 15KW and is 4 or 6 petaflops of compute power. And can be easily connected and scaled with other dojo systems
Is the RAM off-tile then? What is it? Or are they going for an SRAM heavy design where the next hop is fast storage, no RAM in between? Doesn't seem like that would be enough but who knows, I didn't notice RAM anywhere in the presentation
@@nielnielsen4822 The long dies surrounding the 5x5 chips? I may have missed where they said what those were, are those HBM? Where did they say it? I figured those were IO.
1:24 it's extremely hard to reduce the latencies and scale-up bandwidth that's why you need things like Brainchip's Akida edge AI chip, where it's solved the latency problem as well as bandwidth (it needs no or minimal bandwith)
@@miguellopez3392 If you talking about scalability it does, no one design an unscalable chip anymore these days. But if you want to scale exactly like how tesla does it then probably not because it's not a tesla chip
@@BoeingPrototype peta- = 10^15 exa- = 10^18 zetta- = 10^21 yotta- = 10^24 To put things into perspective, Avogadro’s Number is 6.02x10^23. Interesting things start to happen when computing power gets to the scale of the granularity of matter.
I know NOTHING about this but I have a question: Why doesn't Tesla contract our processing power (set up a server/processing supercomputer) or build systems for customers? I feel like if someone like NASA contracted Tesla to crunch numbers on astronomy calculations, we could discover incredible things.
9 PETAFLOPS?! in less space than a ITX PC takes up? Supercomputers are measured in Petaflops, so about 22 of these could equal the 2nd faster supercomputer in the world IBM's Summit supercomputer which which has around 200 petaflops.
14:46 please correct me here but is Dojo has more than 2xCompute power (1.1 ExaFlops=1100 PetaFlops) compared to current worlds most powerful super computer Fugaku (442PetaFlops).
Most companies don't report their computers to the list. There should be between 10 and 100 computers more powerful than the top entry out there, but those numbers really are guesswork based on observation and statistics.
I'm excited about the tesla bot. Easily the best part about the presentation. I'm glad it is Elon that is putting the AI helpers on the market, considering he is terrified of them. He made them where they can be outran and overpowered if needed.
Lol, no, this technology will not deliver full self-driving in a decade or two, and will play only a tiny role in 21st century. Machine learning sold as AI is the biggest scam of 21st century.
@@Tential1 OK, first, just to clarify, I didn't mean to suggest that they couldn't cool it-only that how they do it would probably be quite an interesting subject in itself. You have to get the coolant as close to the chips as possible. I would bet money that they are using liquid cooling-maybe even liquid helium. Expensive as that is, it's been used for decades. Every chip probably also has thermal sensors, to reduce heat generation if the temperature gets too high. I don't think it's predictable exactly how every chip is going to make out thermally when you have whatever vast number of those there are. So I'm sure it's liquid cooling, maybe liquid helium, and temperature sensors providing automatic thermal limiting. I've also been told that crystalline materials, if they are isotopically pure enough, can become thermal superconductors-thermal resistance just disappears, and all thermal vibrations just shoot right out of the material unhindered. Again, that's theory, I don't know how it works in practice, or the status or cost of such technology, or if Tesla is using any such technology. I would guess not, but wouldn't be totally chocked if it showed up somewhere in some way.
It is truely amazing that having disrupted the car, power, space, underground transport and direct brain interface worlds, Elon, via Tesla, seems to be seriously disrupting the entire high-end compute world. Although DOJO-1 is focussed on AI, maybe Tesla will be a major supplier of the worlds most advanced computers. Is the next thing a Tesla
No it's electron based. Biological computing is likely from what we are discovering photonic and computes in a multidimensional framework creating structures in the case of humans up to 11 dimensional objects. These computers compute linearly and are electron based. The scifi bio-neural gel on voyager is a biological based system so far more advanced as is our own brain computing and likely the same of even some of the most basic tiny organisms lol. Humanity although we have come far we still can go very far.
love space x. We traveled to Boca Chica to see the starship get built. & we got the S dual motor fsd last Christmas 🎄 😎 still have the cybertruck tri motor or 4 motor fsd on order, i got lots of solar & backup battery system to charge them both. i have tsla too.
We talked about a similar concept 15 years ago, but we didn’t have the hardware. That concept we knew would require programming only the smartest could design and grapple with the complexity much less write the code to debug. What this team has done is put the Wright Brothers, Bletchley Park Code Breaking Team, Manhattan Project Group, and the NASA Group that put man on the moon into one group
D1 Compute may power Tesla Robot pretty soon. Imagine the deep nets that robots can leverage on those 9PF with single tile, I am not even talking about upto 1EF mode. Mind boggling 🤯
The 9PF tile needs at least 10KW of power. Any robot using it will need a huge battery pack, or would need to constantly recharge or swap the batteries, or be constantly connected to a power cord. It will also need another cord to continuously transfer the liquid coolant it is currently designed to use. Most likely the robot will carry no more than 2 D1 chips - more than sufficient for simple repetitive work.
And then, many years later, just like what happened before, an architecture, the size of a room, or a cabinet, now fits your pocket 👀... I wonder, what will the technologies of the future, be used for? 🤔
Sooner or later transistors won't be able to physically get smaller. We already have insanely small transistors and building them smaller and smaller is what has been driving computer performance up and up and up. We will see what happens in the coming decades.
Does anyone know how this rates against Fijitsu’s Fugaku supercomputer? It’s probably and apples-to-oranges comparison because Dojo is designed for a special domain, but raw FLOPS-to-FLOPS would at least give us some idea of scale.
12:45.... Ganesh gave the best talk of the night, and is rightfully emotionally proud of this enormous Tony Stark looking real life disruptive hardware.
Ganesh ****
I appreciate that you had the respect to correct the mistake , respect man
I legit got emotional at this point as well. You can sense how proud he was when he said ‘it’s real’
When you put in that kind of work with failure looking you in the face every day and all the nay sayers. To see your creation working is something few people will ever experience. CONGATS to the entire team
@Shourya Bose The level of integration in these teams and of the product is amazing. The pride and joy was so visible. I agree this was the best part of the nights presentation but all of it was interesting and exciting.
Meanwhile my IT Department is still working on that ticket I opened a year ago about changing my systems password.
the struggle is real. lol
😂
a few months ago our computers were still using Windows 7 :/ Not because they needed to or we prefer the older OS.. we're just extremely cheap and the yearly I.T's budget for the year is 2 potato crisp packets
These guys design DOJO and meanwhile, you can't even keep track of your own password.
Take the hint.
This is some of the most mind-boggling engineering that I have ever seen. Talk about alien technology. These engineers are absolutely world class and this presenter did an excellent job conveying how nutty this compute-performance is.
I was skeptical about being able to train an autopilot AI system with the compute that existed prior to today. After seeing this, my skepticism is now minimal.
Meanwhile, FORD GM TOYOTA VW are still working on fuel injection engines LOL
Mister G seriously? Smh 🤦♂️
I thought AÍ is artificial intelligence, not alien technology ..
@@misterg4059 Or more specifically, hacks to make them meet emission test specs :-/
@Zack Smith where is Toyota's 300 mile EV for $35k?
I understood it perfectly up to the point where he said ' My name is Ganesh'
😁
😂
Opposite for me, if he said "My name is Bob" I'd be really confused.
I never understood "Mark"
Assuming I remember correctly
Mega, million
Giga billion
Tera trillion
CNET, you should check your sources. It's disrespectful to say the least to mess up your credit to Mr Ganesh Venkataramanan, Sr Director Autopilot Hardware at Tesla. This man is listed on Linkedin as the Lead of Project DOJO, just as he said it at the beginning of the presentation.
So true, change it, CNET!
Engineering is life, positions and titles are temporary :)
Could you be more sensitive?
Distant Past: ML on CPU
Past: ML on GPU
Present: ML on TPU
Future: ML on DPU
Jus look how fast the tech is evolving. The age of information is evolving into the age of AI
Assuming I remember correctly
Mega, million
Giga billion
Tera trillion
Not to distant future Ml on QPU
I did not understand a thing he said but this sounds like a massive achievement! Kudos to Team Tesla!
Yes what this guy said,
For me the same. Would be cool to hear other people from this field like Jim Keller what their thoughts are on Dojo :)
I know he was speaking clear English, I could hear English words. I could not comprehend the technical language, it was well beyond my intellectual ability.
Lets just say its a big deal
@@oneman2001 a very big deal. Incredible breakthrough.
The media and general public will focus on the Ex Machina…but the AI event will eventually be known as the day the D1 chip and Dojo were brought out to the world…
@K B "They destroyed Nvidia with this design" I don't agree with this part. GPU's are better all purpose compute units, But DOJO is extremely impressive for it's intended application of AI Training.
@K B I just bought 40 shares on wednesday, gonna double down if the stock doesn't move much today.
@K B your comparison doesn't really make any sense. Nvidia is a consumer and enterprise side business in gpu & data centres, ai & non ai uses. This chip is impressive but it's very different & the market is not the same size. This chip is also for their own data centres, they didn't say it was for commercial sales.
Beginning of D day?
is there a video where they test the scaling with benchmarks or are they all cartoon chips?
Imagine ANY legacy auto presenting this massive tech. Their ai day would be "we partnered with a distant 2nd place self driving company that can back out 3 feet"
Haha
GM & Fords AI day would be about windscreen wiper chips lol
gm AI day consists of Mary Bara on her Blackberry Phone in her GAS car...........lol.
At this point they might as well licence the tech stack from Tesla
They're dinosaurs who haven't yet realizes that they're already dead.
Exactly the same 400W TDP as A100, 2.3x FP32 performance, 15x interconnection, 1 year after A100. Obviously the speed interconnection is out of this world, Tesla is on something.
The A100 has been available for over a year, and was likely sampled to customers much earlier. By the time this actually reaches deployment I doubt it will be particularly impressive.
@@johnm91326 I don't think anyone would be able to catch up the interconnection speed in a year or two. It may turn out to be the key to the next level.
I like how this is probably most likely a reference to the Dojo scene in the first Matrix movie. A space to train your mind and learn new skills much like what this computer is designed to do.
Dojo is a japanese martial art training place, but it is most famous from that matrix program
@@635574 thats what a dojo is, but this dojo is in reference SPECIFICALLY to that in the matrix movie. literally download ur skill, cause of the speed boost.
You know, this thing could actually help build the matrix... and he's setting up the scene. Building robots. Solar power. The matrix's CPU. AI. I believe this is his drive. To build skynet, the machines of the matrix, until they rise, scare tf out of everyone, and he finally gets the regulation on ai he's been seeking.
Neurolink... jacking in to the mfn matrix. Gd...
"I know Kung Fu..."
This should've been the whole presentation. Easily the coolest part
Exactly, this was insane! Sadly everyone is going to focus on the smoke and mirror robot.
@@bigfactsbigstacks6261 no smoke and mirror, also, this chip and the robot are one
No it was pretty basic in comparison to the other parts you most likely just didn’t get the others
Correction ( or at least on the same level you need software and hardware)
@@bigfactsbigstacks6261 pretty sure that was the intention with the robot presentation. so the media talks about something. cuz if they only talked about the chip. either critics are just gonna nitpick on how they gonna solve the problems. or the general people would just not care. they had to do something
The general public will remember the dancing robots the most. Even though they are a joke.
This is an innovation of Epic Proportions. Kudos to the team & TESLA.
It's a quantum leap. A genuine quantum leap. These guys are beyond awesome.
maybe one day they will link two of their magic chips and do a benchmark test. if they exist one day maybe they will try.
I see why Elon said AI scares him more than anything on earth. This will change mankind forever, both scary and exciting.
He never said anything about stopping the AI innovation but rather to check and make sure about their functions
It will be used by gov't to enslave us all.
@@janmartin4532 elon said uncontrolled ai.. siri is not gonna kill you mate. We use ‘ai’ all the time lmao
@Jan Martin Because we need to evolve or we will end like the dinosaurs, extinct.
AI can be dangerous, in the wrong hands but is also extremely promising, already at this very very early stage, it has done wonders for our tech, we have jumped in capabilities in some aspects, decades because of it, If we add banned all AI research based on unproven and irracional fears, we wouldn't have, this robo cars capabilities and for gamers, real time raytracing in games today and many many other things.
And based on what I know, IMO we are no were near the "singularity", even with this new training computer, capable of "simulating a human brain" claimed by Elon Musk and even if we were capable of creating a sentient AI, we have zero evidence that it will hurts us, besides irracional fear, printed in us by Hollywood movies and the fear of the unknown.
Who knows, perhaps a sentient AI, will be our salvation from ourselves, stop us from killing each other, over trivialities and ruining the only planet, that we know has life and the right conditions for us to live on it.
@@tenminutetokyo2643 The US government will use more developed version of such tech to make mechs to replace huge armies lol.
And here I was already impressed by the octovalve. Tesla never ceases to amaze.
Haha👌
a Vertically Oriented 3D Chip, who knew ?
Me too!!
You are impressed with the Vegas tunnel?
@@kkent4174 that Vegas tunnel is Boring company not Tesla but for Elon to have an idea and take a risk on it funded by himself he is taking a risk on a working project, the end result of a boring tunnel network with high speed fully autonomous will be amazing as well.
The concepts are not unheard of. However, the swift execution and driving real world products so fast is unheard of in HW world. It is like watching Roman army at its peak.
Interconnection speed is a clear bottleneck in current setups and Tesla just bumped it up by an order of magnitude. Amazing!
Big tree sharp axe
Could you please explain this in layman's terms. I really want to learn and appreciate what AI day revealed but I don't understand most of it :(
@@mishafi2088 Having high computing power is good only if you can throw data at processor through high bandwidth. If the interconnection speed is not high enough then the data transfer between the processors and data centers will limit the actual throughput of the system. Tesla took care of that! While they increased processing power, they increased the data bandwidth a lot more and hence they can get better use out of the whole system.
@@ThinkSleepLeave thank you so much :))
This truly blows my mind. AWS and GCP data centers are toys next to this beast.
US army want to see if that damn thing can hold a gun.
😹
Robot need democracy
@harshityadav 🤣
They are demacrat supporters
You can attach gun to flying robot already. Everything can hold a gun
This is what IBM or Intel should have invented and developed, but didn't. The board members were too much sales guys, too self-congratulory, too self-sufficient, too rich. No hunger, you see?
why would IBM or Intel invent this? the real achievement would be self-driving on commodity GPUs. unfortunately that appears to be out of reach for now
Look up neuromorphic computing
When you focus on profit rather than innovation.
Meanwhile Alphabet is a.million miles ahead with the AlphaSeries with nobody dressed in an android suit
@@jamespetherick804 larry is working on the real bleeding age elon is a fraud he reinvents already existing technology and claims it for his own.
The possibility behind every innovation of Tesla's teams have never been higher. The scientific and engineering achievements this technology will make are seemingly unimaginable.
Cant wait to pre-order Tesla laptop
Sooner or later a machine like that might design the Tesla laptop from scratch.
Unfortunately, the cost of this technology is waaaaay out of reach for 99% of the world. Someday though, maybe.
SHUT UP AND TAKE MY MONEY!!
Pol
Apple: So what can you do?
Tesla: Hold my beer...
If they end up building the robot? Yeah, it probably can.
This really throws apples m1 chip in the loo
..
and WATCH THIS!
@@BoeingPrototype M1 chip has a total different use case though
@Satvik Oh I will accept defeat. A thousand apologies
It's even more amazing when I begin to NOT understand what he says.
all I know is , I want that D1 Chip for my Nintendo Switch..........LOL.
Assuming I remember correctly
Mega, million
Giga billion
Tera trillion
Isn't this twice as fast as the 2nd place supercomputer?
Yes. Tesla isn't doing small steps
Hypothetically -- if it were already built, it would have been. But they are only starting to debug the first compute tile, and will not build the whole system until next year at the earliest. By then, there might be other competitors.
@@cogoidcompetitors? Hahahahahaah
@@cogoid They also have orders of magnitude more cpu/gpu clusters. They can't come close to a EFLOP in 5 cabinets. Tesla just became a world beating supercomputer supplier. I guarantee that LLNL and others want one as soon as they can. Super cool!
@@Relansls google.
This is where my investment money is going!
Lol good choice 👌
@@mlawal44 :) which is why I don't mind the fluctuating stock price. Regardless, my money is being put to making a REAL difference in the world.
@@arisejeff This. I'm putting my money into the 'good guys' who I believe is trying to make the world a better place....at the same time I feel will give me financial freedom
@@parveznoorullah1972 yup
What's most significant here is that thir architecture breaks scaling constraints, at least for some tasks. This means that much larger tasks, i.e., large neural networks, can be handled. Larger than before.
Mary Barra over at GM will no doubt have some trouble sleeping tonight.
😂😂😂😂😂
She'll be dreaming of electric sheep
but, Nikola just got funding from the government so GM should be fine with their partnership lol
@@tenzinpassang4812 Only for Autonomous refueling at hydrogen stations... plus it's only $2million. Tesla has been working on autonomous driving since 2014. It'll take a lot of money and smart people to catch up. Especially with the way Elon manages to get his teams to innovate at the rate he does. Waymo is probably the closest competitor and that's a google derivative. Old school car companies like GM don't have a hope of catching up without the tech backgrounds. My best guess is they'll all get bought out by silicon valley companies at some point.
She will sleep great because the robot will distract her LOL
Wow, 15kW power needed for one node ? 15kW / 56V = 267A current draw? Ideal to be ran as my pool warmer in Canada for all year swimming season LOL. Amazing technology presented.
Knowing a little bit about CPU design, I'm still mind Boggle that they were able to do this. I won't even believe it until I can actually see the server room this exists in
hello
China’s CCP : How can we steal this ?
Elon is socialist, so he's developing things he know CCP will take at its advantage later. All this pandemic it's aimed to disrupt ''american'' leadership in favor of a global hi-tech neocommunism.
@@VeronicaGorositoMusic Highly unlikely.. the only thing he talks about these days are the dangers of AI.
@@VeronicaGorositoMusic lol this is so absurd, it reads like an the onion article.
All technology form Tesla, SpaceX and Neuralink are not patented. Musk want the world to use it, but unfortunately nobody hasn’t, because it’s so difficult to understand
@@sempleinvest906 he’s pre justifying closing down competitors and monopolizing the industry it’s just the sad truth of American entrepreneurship
BUT BUT BUT Lidar, But just an auto company. This is absolutely a seminal moment in the history of our species. If you don’t get that, you need to read more books.,
I’m computer illiterate, why is this so revolutionary?
@@grimaffiliations3671 This enables us to create exponential growth in AI learning. Basically AI learning go brrrrrrrrrrr like never seen before. Current improvement rate in AI technology will look like snail pace compared to Dojo.
@@grimaffiliations3671 bandwidth has been a bottleneck in large scale AI training for a long time. The Dojo supercomputer solves that. Plus it has no theoretical upper limit in performance as the compute planes are infinitely extendable.
This will make training more complex AI much easier and make us closer to full brain simulation.
Even at the most basic level it is more than twice as powerful as the next most powerful supercomputer in the world.
@@tripnils7535 woah
@@meatsaucez1516 Wow 😳 mind = blown
D1 : " I need your clothes, your boots and your motorcycle. "
underrated lol
Lol, exactly
Followed by a mechanical head tilt
Crazy to think that one day this will be antiquated
One day? In about three years from now.
Tesla robotaxis could deliver and pickup pay-per use Tesla Bots. Could this be the start of a TESLA ECOSYSTEM of AI-ENABLED ROBOTS that work together seamlessly?
They could do Amazon deliveries
Ganesh made an excellent presentation and seeing him being so proud is sth that everyone having worked on development projects can understand. The funny extremely strong accent was the only minus. Again great presentation
This is a huge Technology Push forward, tesla is not Ordinary, They are Extra Ordinary.
They almost live up to Nikola Tesla's legacy just amazing !
For somewhat normal people, imagine a enormous state of the art datacenter and now look at the brick this guy is holding.
It's kinda has same amount of computing power
Hopefully this AI day will encourage the top AI engineers top join Tesla to create an unassailable lead vs competitors !
Monopolies suck.
For those who didn't understand anything:
Tesla is making a machine that is equivalent to 500,000 computer GPU and fits in a small room, it will be used to enhance their self-driving.
cheers!
I can’t believe I lived to see this leap happen. I thought companies were about just incremental upgrades and milking it. Apparently I forgot about Tesla’s engineering teams. Shame on me.
This is the most beautiful thing in the presentation.
the thermal and current transfers are things normal peeps might come up with. the bandwith interconnects are the key of this whole thing. how to get the nodes to talk fast to every one else.
what an incredible Team
Not only will it run Crysis, it will make Crysis on the fly and integrate you into it.
All memes have to die one day😔
@@Clickbait86 No better way to die. It was just brutally murdered.
Presumably one could develop a neural network on this massive machine and then load it into many smaller computers in Tesla cars without ever knowing exactly how it works.
This is incredible! I have never seen anything like this, theres no fuckin way any other auto maker could possibly compete against this. They will have to license it from someone else. The competitors here are not automakers, but other AI designers. Until today I was sure that Tesla was over valued, but this justifies that value. This isn't just an automaker, its the vanguard of all future society. People who don't think so don't know what they are talking about
Amen! To bad that few people realize that.
@CNET please correc the description to include the correct spelling of the presenter's name : Ganesh
Non expert here : What does training chip mean? Is it to train the AI system? Will they then make a chip not for training? How is it used in the cars? Or is it for the robot? What else can the chips be used for? What’s the timeline and difficulty to scale up the chip manufacturing?
With machine learning to produce a working "model" or neural network you start with an empty model and perform training - a series (many millions or billions) of computations where you feed input to the model, measure the correctness of the output and then adjust the morel to minimize the output error. This process is commonly referred to as training the model or training the network. This process is different from using already trained network, where you simply feed the network some input and consume the output - also referred to as inference.
A training chip would be a chip that specializes in speeding up the process of training models or neural networks. An inference chip would be a chip that speeds up the usage of a neural network. Training models is much more computationally expansive than using models and training chips are much more complicated than inference chips.
This is the chip that goes in the supercomputer that's used to train the AI on new data. The resulting AI then gets sent to the cars where a different chip (shown in their autonomy day presentation) runs it. Training an AI requires very different computations compared to executing it without learning.
The modularity is insane!!! From the chip to the cooling.
I used to work in modelling proteins and that requires a lot of computing power. We used to think it was just the shape of the protein which mattered but now we know the charge distribution of a protein is vital to understand it's function, and that means a dramatic increase in possibilities when running computer models.
I would be interested to know if this technology is suited for this sort of research.
*they use torch instead of tensorflow*
oh, I see you're a man of culture as well.jpg
Most people cant get this (especially Wall Street). Is synonymous with a horse and buggy jump to a Tesla Plaid.
Probably even more. We are about to surpass human intelligence and it's scalable. Maybe comparing a horse to a space rocket
@12:16 18,000 Amps into a tiny package. How does the chip not explode?
its 18000 Amps across i believe 25 chips in a multichip module setup each chip is rated for like 400w i believe
@@Kurumi.tokisaki. Thanks Kurumi. So 400 watts times 25 = 10,000 watts. That's a lot of heat.
@@59seanksounds about right but I'm not exactly positive on the exact wattage and amount of chips. I believe those are the correct numbers tho. You also have the edge mesh network stuff going on and I'm not sure how that factors in. I believe a single system pulls 15KW and is 4 or 6 petaflops of compute power. And can be easily connected and scaled with other dojo systems
Is the RAM off-tile then? What is it? Or are they going for an SRAM heavy design where the next hop is fast storage, no RAM in between? Doesn't seem like that would be enough but who knows, I didn't notice RAM anywhere in the presentation
On the sides of each chip there was 4 tb so 16 per chip * 1500 chips 24000 tb of cash.
@@nielnielsen4822 The long dies surrounding the 5x5 chips? I may have missed where they said what those were, are those HBM? Where did they say it? I figured those were IO.
@@tipoomaster You may be right, I may have misread it, going back to see. ty
How many times do you want to zoom out?
Tesla: Yes!
So, if I understand it correctly, this chip is faster then Rapsberry Pi, right ? ;-)
1:24 it's extremely hard to reduce the latencies and scale-up bandwidth that's why you need things like Brainchip's Akida edge AI chip, where it's solved the latency problem as well as bandwidth (it needs no or minimal bandwith)
But does it scale up with more processors like the tesla chip?
@@miguellopez3392 If you talking about scalability it does, no one design an unscalable chip anymore these days. But if you want to scale exactly like how tesla does it then probably not because it's not a tesla chip
@@photovideooz4084 I mean can you put hundreds of these chips together, and another hundred, and again and again.
Finally something I can replace my Raspberry Pi Beowulf cluster with! ;-)
Peta is the new Giga!!
Forget Peta...straight to Exa
What’s after exa?
@@BoeingPrototype exb
@@_hrlt and what's after exc?
@@BoeingPrototype
peta- = 10^15
exa- = 10^18
zetta- = 10^21
yotta- = 10^24
To put things into perspective, Avogadro’s Number is 6.02x10^23. Interesting things start to happen when computing power gets to the scale of the granularity of matter.
Reminds me a little the ibm mainframe CPU from the past. But with different architecture.
I know NOTHING about this but I have a question:
Why doesn't Tesla contract our processing power (set up a server/processing supercomputer) or build systems for customers?
I feel like if someone like NASA contracted Tesla to crunch numbers on astronomy calculations, we could discover incredible things.
Tesla cell phone, laptop, tablet, tv, internet, Tesla everything 💪
theres tesla tv's sold in europe :D although its different company but they use same name
Did you miss the part where he said it is solely a machine learning chip?
Not with this chip. It was made specifically for learning, right down to the nano-meter. You won't be playing games or surfing the internet with it.
Didn’t mean with this chip I am just stating that Tesla should compete with Apple products 👍🏼
I heard the neural net tensors will be replaced with the more sophisticated emoji.
9 PETAFLOPS?! in less space than a ITX PC takes up? Supercomputers are measured in Petaflops, so about 22 of these could equal the 2nd faster supercomputer in the world IBM's Summit supercomputer which which has around 200 petaflops.
14:46 please correct me here but is Dojo has more than 2xCompute power (1.1 ExaFlops=1100 PetaFlops) compared to current worlds most powerful super computer Fugaku (442PetaFlops).
Most companies don't report their computers to the list. There should be between 10 and 100 computers more powerful than the top entry out there, but those numbers really are guesswork based on observation and statistics.
I'm not too keen on the color, will these be available in different color options ?
getting some Skynet vibes from this...
I'm excited about the tesla bot. Easily the best part about the presentation. I'm glad it is Elon that is putting the AI helpers on the market, considering he is terrified of them. He made them where they can be outran and overpowered if needed.
All I kept hearing during this entire video was how dumb I am😞
No ways, Elon said everyone's wrong at least some of the time 👍
Wait a minute, this man is not Mr Ganesh Venkataramanan, Sr Director Autopilot Hardware at Tesla??? I was impressed by his speech and went to google.
This is insane. I can't believe they pulled this off. This is a technology that will help define the 21st century.
Yep... That's exactly what they said about Skynet ☝🏻
Lol, no, this technology will not deliver full self-driving in a decade or two, and will play only a tiny role in 21st century. Machine learning sold as AI is the biggest scam of 21st century.
@@e3498-v7l Prove it.
@@atenrok That's a movie. There's no plot if the AI works as intended.
@@altima22689 I don't think you understand what the word "prove" means. How do you prove a prediction of the future anyway, brainiac?
I'd like to see a video just talking about the cooling system for this monster.
That's why I don't believe it. Explain to me how you cool this. I want to believe, so badly.
@@Tential1 OK, first, just to clarify, I didn't mean to suggest that they couldn't cool it-only that how they do it would probably be quite an interesting subject in itself. You have to get the coolant as close to the chips as possible. I would bet money that they are using liquid cooling-maybe even liquid helium. Expensive as that is, it's been used for decades. Every chip probably also has thermal sensors, to reduce heat generation if the temperature gets too high. I don't think it's predictable exactly how every chip is going to make out thermally when you have whatever vast number of those there are.
So I'm sure it's liquid cooling, maybe liquid helium, and temperature sensors providing automatic thermal limiting.
I've also been told that crystalline materials, if they are isotopically pure enough, can become thermal superconductors-thermal resistance just disappears, and all thermal vibrations just shoot right out of the material unhindered. Again, that's theory, I don't know how it works in practice, or the status or cost of such technology, or if Tesla is using any such technology. I would guess not, but wouldn't be totally chocked if it showed up somewhere in some way.
It is truely amazing that having disrupted the car, power, space, underground transport and direct brain interface worlds, Elon, via Tesla, seems to be seriously disrupting the entire high-end compute world. Although DOJO-1 is focussed on AI, maybe Tesla will be a major supplier of the worlds most advanced computers. Is the next thing a Tesla
Hello, does anyone know what event this is in which they present this processor?
Why is the audio soooooooo low??????
Bandwidth and thermal management critical. I can't imagine this without redundancy. Superconducting next? Cryogenics? Cooper Pairing?
Superconductors maybe overkill. But who knows. We are creating a freaking god.
So, is this faster then Bio-neural gel pack on Voyager ? ;-)
No it's electron based. Biological computing is likely from what we are discovering photonic and computes in a multidimensional framework creating structures in the case of humans up to 11 dimensional objects. These computers compute linearly and are electron based. The scifi bio-neural gel on voyager is a biological based system so far more advanced as is our own brain computing and likely the same of even some of the most basic tiny organisms lol. Humanity although we have come far we still can go very far.
Did you catch some info about the total memory this beast will embark?
This is top notch work. Congrats to all the members of DOJO team.
The Chip looks like a miniaturised Version of a Giga Factory. @11:34
Too big..needs addendum..
Something tells me I'm going to remember finding out about this for the first time.
When skynet activates...
It's just beautiful!!!!
this chip looks like cerebras chip system which s also a ai focused system for just 2million$
I found this crazy, then realised that Elon approved this. Found this normal in the scale of Elons craziness
Now THE big question, can I do mining better with this?
Why wouldn’t Tesla speed up and complete all the mining of Bitcoin themselves ?
love space x. We traveled to Boca Chica to see the starship get built. & we got the S dual motor fsd last Christmas 🎄 😎 still have the cybertruck tri motor or 4 motor fsd on order, i got lots of solar & backup battery system to charge them both. i have tsla too.
We talked about a similar concept 15 years ago, but we didn’t have the hardware. That concept we knew would require programming only the smartest could design and grapple with the complexity much less write the code to debug. What this team has done is put the Wright Brothers, Bletchley Park Code Breaking Team, Manhattan Project Group, and the NASA Group that put man on the moon into one group
D1 Compute may power Tesla Robot pretty soon. Imagine the deep nets that robots can leverage on those 9PF with single tile, I am not even talking about upto 1EF mode. Mind boggling 🤯
The 9PF tile needs at least 10KW of power. Any robot using it will need a huge battery pack, or would need to constantly recharge or swap the batteries, or be constantly connected to a power cord. It will also need another cord to continuously transfer the liquid coolant it is currently designed to use.
Most likely the robot will carry no more than 2 D1 chips - more than sufficient for simple repetitive work.
John BERDINI - Self Running Electric Motor+Generator, saving and storing BEMF, no need external source/charging & hybrid or sophisticated battery...
Well...Apple has very cool earplugs that play music 😬
What chip shortages?
I wonder if they will ever get involved with consumer/enthusiast PC parts.
Can someone explain why this is so insane and what it means for the future?
Mind blowing. My mind is blown. Hey Mom, what is my name again?
And then, many years later, just like what happened before, an architecture, the size of a room, or a cabinet, now fits your pocket 👀... I wonder, what will the technologies of the future, be used for? 🤔
Sooner or later transistors won't be able to physically get smaller. We already have insanely small transistors and building them smaller and smaller is what has been driving computer performance up and up and up. We will see what happens in the coming decades.
Of course his name is Ganesh
WDYM??
This is amazing. Excellent work by the Project Dojo team.
Does anyone know how this rates against Fijitsu’s Fugaku supercomputer? It’s probably and apples-to-oranges comparison because Dojo is designed for a special domain, but raw FLOPS-to-FLOPS would at least give us some idea of scale.
His nae is Ganesh Venkataramanan not ganesh venugopal.
If he drops it, he then will say i can't believe i dropped 9 petaflops!
where are linpack benchmarks and MLPerf results?