Tesla D1 Chip - What is All The Fuss About?

hello my name is gary sims and this is gary explains now a tesla recently had its ai day where it talked about all the things it's trying to do particularly in for self-driving cars for automotive and then what that means for its own tools that it has to develop and then some interesting looks at the possibility of the future uh robots and so on now during all of that it announced its new d1 chip and uh there'd be quite a bit of fuss about it i really want to talk about what it is what it means all the technical specifications they'll come up on the screen as we're going along but really what is the d1 chip why is it important why are people even talking about it well if you want to find out more please let me explain now of course tesla are very famous for its range of electric cars and one of the things it's trying to do with electric cars is to turn them into autonomous self-driving cars and that means it needs to play around with machine learning and push the boundaries of artificial intelligence particularly because obviously in an outdoor situation not only have you got to follow the lane i mean i've got a video here on this channel where i drove a tesla and tested out some of its early auto pilot functions it's brilliant you know it can stay in the lane and it can break when the car in front breaks and all that kind of stuff but obviously tesla's got great plans for what it wants in the future so you've got traffic lights and pedestrians and you've got dogs running across the road and you've got you know what's happening at junctions and bicycles and buses and cars and trucks there's a whole bunch of things that go on and uh tesla are looking to develop the self-driving part of its uh of the tesla cars so that it can cope with that now machine learning has two very distinctive phases now we talk a lot about the second phase which is inference so when we talk about smartphones for example we talk about the google tensor chip when we talk about what the capabilities are of computational photography uh today on our smartphones like inference can you point your camera at something and it says oh look that's a dog that's a person that's a landscape and it realizes what it is it does object detection that's inference what that basically means you take a trained neural network it's already been set up it already recognizes certain things you present it with some new data that it hasn't seen before and it infers it guesses it it logically looks at what's there according to the training data it has and then says yes that's a dog he's never seen my dog before but it recognizes it's a dog because it's seen thousands and thousands of other pictures of dogs uh previously before that that's called inference now we want inference to run very very well on all of our devices on our smartphones in cameras in cars in anything that's going to have any kind of visual or audio input inference is very very important the other part of it of course is training that's when you start with the kind of blank neural network it's specialized of course at that point already but when you start with this blank piece of paper and you want to train it so that it can recognize a dog or of course as we get into advanced levels of machine learning it can recognize pedestrians and and cyclists and lorries and trucks and people traffic lights and and intersections and all this kind of stuff so you need to present it with hours and hours and thousands and thousands and millions and millions of examples of those things so they can start to learn what they are and start to learn about how they work now that is really really computing intensive inference is relatively easy because you say here's the input run it through this network yes or no dog or cat you know pedestrian or cyclist it's relatively simple but the training part is absolutely compute intensive now up until now a lot of the data centers around the world that are doing training are built using gpu so traditionally gpus from nvidia seem to have carved their way into this market and they build big servers with lots and lots of gpus per node and then of course each of those gpus has many many cuda cores and they are used then for training the neural network and in fact tesla talked about how their current setup is using this big sort of kind of supercomputer with all these gpu cores built into it and when you look at some of the stuff that tesla are trying to achieve through machine learning you can see that actually they're going to need better systems for training so for example they're now at the stage where they are not just taking a single frame a photo effectively and saying right can you spot the cyclist can you spot the truck can you spot the intersection it's using camera input from multiple cameras in video at the same time so as a truck goes past you can see it behind you can see it next to the car and you can see it going in front of the car and the the neural net understands that a car a truck has just gone past it or you've just gone past a truck and so that has to take multiple camera inputs and then translate that into what they're calling vector space which is the idea of what things look like logically not what they look like you know just pixels on a screen but logically right there's a car there there's an intersection there there's a turning over you know all that kind of stuff now to achieve that from multiple cameras they really need some advanced training so a few years ago they started this dojo project now dojo of course is the key thing to understand this dojo is where you do training so it's not a chip that you're going to find in a car it's not a chip you're going to find in a smartphone it's a chip having the data center for training all their modules because dojo is where you do the training it's not and then you go out into the real world to do whatever it is you've been training for so the new d1 trip is a custom chip built by tesla specifically for training machine learning neural networks and because it's a custom chip specifically designed for that it's not like a normal cpu or a gpu no it can't play crisis no it can't run android or linux or something like that of course there are things around it that need you need to be able to have ways of getting tools in there you need to be able to get ways of running things and talking to it and discs and all this kind of stuff but the actual training node as they call it is just specifically for doing machine learning and although it's a cpu in the traditional sense it takes instructions it decodes them executes them and does something with registers actually the cpu instructions are very very much geared towards machine learning statistical analysis matrices operations so they're even talking about things you know there are operations to do broadcast and gather which is the kind of things you have inside of distributed computing and they're doing that in the hardware as an instruction or a set of instructions so it's not just you kind of you know you add your multiply you divide your branch you compare as you would get in the traditional cpu it's actually got hardware instructions designed specifically for doing machine learning uh the training part of machine learning and doing that in hardware then once you have these training nodes and you could try and draw a parallel between it say a cuda core or an nvidia gpu and the training node rough parallel in in the idea it's an autonomous theme that can run its own program has its own ram and can interact with the world around it they take 354 of those and then put that onto a single chip we include all of the io how you connect over different things to talk to different things in a system because we're talking ultimately there'll be a big computer system so it needs to be able to talk to the real world and then you can take 25 of those and then put it onto a tile and then those tiles are actually built in such a way that it's got power it's got a heat sink it's got all the connections it needs and it's a combined single module that can then be built together in a cabinet with other tiles to ultimately produce a supercomputer built for machine learning for the training part of machine learning when they put all these tiles together in a cabinet it builds you a thing called an exopod that offers an extra flop of performance that's floating point operations but of course even here of course there are difference between 8-bit floating point operations 32-bit floating point operations so there are some numbers floating around that different people are analyzing but at the end of the day it's this very very big super computer which offers all these tiles built together for uh tesla to run its machine learning so what does that mean for consumers today well not a lot at the moment because our laptops and tablets and our phones will all continue to emphasize the idea of inference and a good inference engine in the hardware for consumer products and in terms of you know a tesla car then in the future of course there will be better improved or auto pilot functions self-driving functions and ultimately it shows that the machine learning really is diverging into two key areas one is of course training and the hardware needed specialist hardware needed for training and the other is the inference and the specialist hardware needed for the inference one is the consumer side and one is the producers beside the manufacturer's side of the particular products at the moment as i said gpus were really filling that gap inside of the machine learning the training part of it however now we're seeing uh hardware that's being released and if tesla are doing this you can pretty much guarantee that someone else somewhere is also working on it so maybe gpus are not going to be as significant in the data center for training as we thought they might be because now there are alternatives coming out that are more specialized better at performance better efficiency that allow greater more complex models to be trained okay so that's it that's the d1 chip the dojo where the training happens and that's from tesla which tesla will use to build the uh self-driving cars that it's trying to self-driving software that's putting into its tesla cars okay that's it my name is gary sims this is gary explains i really hope you enjoyed this video if you did please do give it a thumbs up i hope you're following me on twitter at gary explains and also i have a newsletter go to garrexplains.com type in your email address no spam just a newsletter and i think you'll enjoy it okay that's it i'll see you the next one [Music] you

Channel: Gary Explains

Views: 127,812

Rating: 4.8909583 out of 5

Keywords: Gary Explains, Tech, Explanation, Tutorial, Tesla, D1, Dojo, Project Dojo, Machine Learning, ML Training, Tesla D1, ExaPod, Training Tile, Training Node, AI, ML, Self driving cars, autonomous cars, Self-driving car, full self driving

Id: GZ1Zu44WcdA

Length: 10min 56sec (656 seconds)

Published: Mon Aug 23 2021