How Supercomputers ACTUALLY Run The World

the world's richest man Mr Elon Musk is building his very own supercomputer but why is he doing that on the surface level the machine named Dojo is under development at Tesla to help the company train their artificial intelligence and eventually make their self-driving cars more reliable but if we pull back and try to see the bigger picture there is much more at play here supercomputers are quickly becoming some of the most valuable infrastructure on the planet behind closed doors these things have essentially been guiding the course of human civilization since the 1990s so in a disturbingly real sense having control over supercomputers can equal control over the entire world before we get into the weeds of how supercomputers run the world we are going to to need to establish a baseline understanding of just what these things are how they work and what they do the first level is knowing that a supercomputer is not just a super powerful version of a regular computer you can't just load up your own PC with upgrades until it goes super computer it doesn't work like that supercomputers achieve their performance by making use of parallel Computing so a supercomputer is really a very large number of smaller computers all functioning as one to accomplish a task faster and more efficiently the classic example used here is an F1 pit crew when the race car comes to a stop in the pit Lane all four tires are removed and four new tires are installed this all happens in a period of about 2 and 1/2 seconds and that is only possible because a team of highly trained individuals work together in unison each performing their own simple task one person loosens the wheel nut another person removes the old Tire a third person places the new tire and this is done on all four corners of the vehicle at the same time along with a flurry of other important tasks performed by a total of 22 individuals when the efforts of a pit crew are combined in parallel they accomplish orders of magnitude more work than any single person could ever dream of even the most superhuman Hercules and Godlike mechanic in the world would never come anywhere close to the efficiency of 22 mere mortals working together so now all we have to do is replace the pit crew members with computer processors and we're beginning to see how parallel Computing works this is not the kind of system that is going to be used for any of the same things that an average person would use a computer for these aren't for playing video games or even making video games or editing YouTube videos or anything exciting like that supercomputers are almost exclusively used for scientific calculation which is also often referred to as high performance Computing or HPC obviously any computer is able to make calculations a lot of calculations but the difference is that the normal computer will try to cram all of those calculations through one single processor in series while a supercomput cluster will distribute those calculations evenly across hundreds or even thousands of processors in the words of Tim Carroll the director of high performance Computing at Microsoft If one computer can solve a problem then 5,000 computers can solve it even faster this is obviously going to require an incredible amount of Hardware to accomplish the most powerful supercomputer in the world is named Frontier and it's based in the United States at the Oakridge National laboratory in Tennessee There's a list of the 500 most powerful supercomputers in the world at least the ones that we know about and it gets updated twice a year but it's unlikely that Frontier will be unseated from the top spot on that list anytime soon the previous most powerful supercomputer in the world the Japanese fugaku topped out in June 2023 at 442 pedop flops while Frontier reached a maximum of 1,19 94 pedop flops and stick with me we'll get this all explained but for now just know that Frontier is more than double the performance of the next best supercomputer that's a pretty wide margin so what the hell is a pedop flop we measure computers by how many float Point operations per second they can accomplish these are called flops basically how many calculations they can process in 1 second one pedop flop is one quadrillion floating Point operations per second and Frontier was the first computer to pass 1,000 pedop flops meaning it has reached the exoscale if you were to write out one exaflop it would be a one with 18 zeros behind it just an unfathomably large number of calculations and that's just in one second of operation to accomplish this Frontier uses 9,472 AMD epic CPUs which are 64 core processors that are optimized for HPC each CPU is paired with four Radeon Instinct gpus and each of these CPU GPU groupings are called a node and this is what you'll find inside of those giant supercomputer cabinets Frontier is spread out over 74 cabinets all working together as one system at Peak Performance this entire system draws over 40 megaw of power power which is equivalent to about 50,000 average homes to manage the amount of heat generated by all of that energy the system uses 30,000 L of water per minute for liquid cooling with all of that combined each cabinet weighs in at around 4 metric tons the cost to build this system in 2021 was $600 million which is not really a whole lot of money in today's world when you think about the value of the product I mean Elon Musk paid $42 billion just to be in charge of Twitter but the true cost of something like this is in the operation just try to imagine all of that electricity and water and routine maintenance for the entire system so what exactly do these things do that would justify all of that money in short supercomputers can predict the future that's the primary selling point for this technology at least but on a grander scale supercomputers allow us to see the unseeable know the unknowable they give us the power to transcend many of the limitations within the human condition let's start off with something simple during the course of research for this script we came across an Australian news report from 1973 that claimed a super computer had predicted the end of civilization obviously that pequ our interest and if we start off with how an ancient supercomputer made a prediction about the future then all we need to do is scale that up to understand how significantly more powerful supercomputers shape our modern world so the idea was to create a complex model of the world and the effects that human beings were having on it it was the first time that we'd really tried to visualize the entire world as one system which means that there were a number of complex variables that needed to be taken into consider consideration the output of that particular computer model was essentially just a series of graphs that were all overlaid on a timeline that extended from the year 1900 to 2100 the portion of the graph from 1900 to 1973 was based on known data and then everything after that was a result of the computer model identifying various Trends and cause and effect relationships between the different lines on the graph for example there was one line for the human population another line for quality of life another line for the amount of pollution in the environment you see where we're going with this the main prediction was that the human population would continue to grow steadily until around the turn of the century when it begins to Plateau over the same time period the overall quality of life for people their access to food shelter income and such would steadily decrease and then beginning in the 80s the amount of pollution would begin to rise sharply and grow exponentially at a certain point the graph shows that both population and quality of life hit a steep Decline and by around the year 2050 we end up back at the population level that we started with in 1900 essentially meaning the end of modern civilization as we know it if you are curious the model predicted the year 2020 would be the major turning point where everything starts to really fall apart take from that what you will obviously this is a fairly rudimentary study and there's nothing here that a human scientist couldn't have worked out and charted on their own but if we start to imagine how many researchers and how many working hours it would have taken to produce these same graphs all of that is replaced by one simple machine you feed it the data and the computer produces a complex forecast in a minimal amount of time now all we have to do is Imagine That Baseline example scaled up to 1 million ion times more processing power or 10 million times how many more variables and data sets could you include in a modern forecasting model the more calculations the computer can make the more accurate its prediction will be let's Jump Ahead to the 1990s and the Human Genome Project sequencing a genome is determining the exact genetic code of an organism and being the naturally curious animals that we are humans wanted to know our own specific code what makes us what we are this is obviously a very difficult question to answer because the genome contains the entire blueprint on how to construct a human being yet it all fits inside the nucleus of a single cell it's not really something that is fathomable on a human scale but the easy answer is that the genome is a code and all we need to do is understand how it works to do that researchers used supercomputers in the year 2000 IBM dedicated the world's most powerful supercomputer exclusively to the Human Genome Project at the time it was reported that the upgrade in computing power would shrink the weight time for advanced calculations down from 447 years to 1 month to give you an idea of what they were working with there are over 35,000 individual genes that make up a full strand of DNA and if we're trying to identify a particular genetic effect that would make a person more susceptible to a disease like breast cancer then that often doesn't come down to just one defective Gene It's a combination of genetic factors so researchers would need to identify all of the trillions of unique combinations and then figure out which one was associated with the highest likelihood of developing breast cancer the machine that IBM had created for this job was capable of Performing over 7 trillion calculations per second which was already 600 times faster than the computer IBM released in 1997 and that was the first ever to defeat a world chess champion but even these machines from 20 years ago are nothing but toys compared to Modern supercomputer clusters in the year 2000 they had just broken into the terlop scale 7 trillion calculations per second is seven teraflops the Xbox series X gaming machine from 2021 comes with 12 teraflops of computing power so the machine that helped decode the human genome wouldn't even meet the standards of a literal toy in today's market and coming back to the present day we've now reached the ex scale which is 18 zeros behind the number that's six orders of magnitude beyond the teraflop scale 12 zeros behind the number this is a term that people like Elon Musk throw around often orders of magnitude but what does that mean let's say on one side you have a go-kart and it has 10 horsepower that would be a very fast little machine and on the other side you have a Tesla Model S plaid with 1,000 horsepower that's a very big and fast machine what is the difference in power between the two if we scale up the go-kart by one order of magnitude we'd be at 100 horsepower if we scale that up by a second order of magnitude we reach 1,000 horsepower so the difference between a fast go-kart and the fastest Tesla is just two orders of magnitude if you scaled up the go-kart by six orders of magnitude you'd be at the equivalent of a small rocket engine so what do we use these rocket powered computers to do for us in the modern day well there are a bunch of things including the last one we'll mention which might blow your mind but one of the more interesting use cases to me at least is in astrophysics we know that the Universe operates on a much longer time scale than a human being our meager 100-year existence is insignificant to deep time so we will never be able to watch a Galaxy form or a star die or a black hole consume an entire solar system worth of matter but we know that these things happen and we even have a pretty good idea of how they happen and why they happen we can visualize the entire process through the computer models essentially cheating time itself in a more down toe example we use supercomputers to predict the weather and we're not just talking about if it's going to rain tomorrow when you see a computer model of a hurricane that shows the strength speed and direction of the storm over several days that is a supercomputer at work we're all very accustomed now to seeing long-term climate models these simulations that predict how global weather patterns ocean temperatures and sea levels will change over time they all tend to agree that some really bad stuff is not very far away now a lot of people take comfort in the idea that this is all a big conspiracy it's totally fake and the planet is going to be just fine and it's hard to blame them when the reality is that we simply took all of the atmospheric and climate data that has been collected all around the world for the past 100 years and then fed it into a cold calculating machine then we asked it to tell us what comes next we do the same thing with medical science when scientists are modeling the spread of a certain respiratory virus that you may have heard of they use supercomputers to do that the same with sequencing the Genome of said virus which can now be done almost instantly which in turn leads to the very fast development of new vws to try and combat the c-word all powered by supercomputers it's taken a very long time to write this script I don't want to trigger any YouTube algorithms which not by coincidence are also powered by supercomputers the amount of computing infrastructure behind YouTube is gargantuan but it's also very necessary imagine having to manually filter through every single video posted to YouTube in real time it could not be done you need that algorithm working to identify what content you actually want to watch hopefully this video is one of the good recommendations so thumbs up for the algorithm anyway that's all really interesting but the biggest use case for supercomputers right now is going to blow your mind a bit if you don't already know this nuclear weapon testing yeah because the last real world nuclear weapon test conducted in the United States was back in 1992 the Chinese stopped in 1996 India and Pakistan stopped testing in '98 North Korea is the only known state that is still conducting live nule nuclear bomb tests as recently as 2017 there are a few other alleged incidents from China and Russia in the past decade that could have potentially been nuclear explosions but we can't say for sure either way that's not stopped the development of new nuclear weapon technology it just means that the testing has moved from The Real World to the simulated world if a nuke explodes in The Matrix does it make a sound so if you've ever wondered what world governments particularly the United States are doing with super computer technology mostly they are watching mass casualty events and potential doomsday scenarios play out over and over again okay let's bring it back to Elon Musk we Nam dropped Elon at the start of the video for a reason and now we are actually going in for the payoff Tesla is developing their very own supercomputer and because this is a musk related project it of course has an edgy name that is trying way too hard to be cool but Dojo is actually something different from all all of the above supercomputers that we have been talking about now we are really bringing AI into the picture just like modeling the climate of a virus or the end of the world an AI model needs to take a gigantic amount of data and translate it into a prediction what comes next the output from chat GPT is just a series of words that are placed one after the other based on the model's best prediction for what should come next when you finish this video YouTube is going to make several predictions for what you might want to watch next the difference between traditional supercomputer models and AI models is that AI needs to be trained this is why Tesla calls their computer Dojo because what they're building is not a superc computer in the traditional sense it is specifically an AI training cluster so what's the difference the level of precision necessary for high performance computing and creating scientific models of viruses and DNA on the atomic level is obviously very high so a traditional supercomputer like Frontier will operate at 64 bits but training an AI model that is essentially trying to predict and replicate human behavior does not require a high degree of precision we're just not that complicated so AI training can be done at 16bit or even 8 bit levels that means a cluster like Dojo can accomplish more calculations more flops with less power because each calculation is easier to solve an easy to understand example of this would be the number Pi if you ask any competent high school math student to calculate the area of a circle they'd probably do 3.14 * the radius squared and they would get a correct answer if you ask a more advanced student they might calculate 3.1416 * radius squ and they would get a virtually identical but slightly more precise correct answer and for an even higher level of precision you could calculate 3.14159 and so on and so forth high performance Computing is calculating pi to the 50th decimal place AI training is using 3.14 and calling it good enough so when Tesla advertises that Dojo can reach the ex scale of compute they're also including a little asterisk that specifies that it's happening at 16bit de ision and not the industry standard 64-bit so yes this is going to give Elon Musk a significant leg up in his mission to make cars drive themselves and unleash an army of autonomous humanoid robots into the world but it's not going to make him an all- knowing Super Genius either and that's perfectly fine that's not calling Dojo a bad computer it's probably going to be very good at what it is specifically designed to do we're just trying to give you some realistic perspective on a very complicated subject and we hope that you found that valuable today if you enjoyed this video and want to see more content like this please consider supporting us directly on patreon and hit the like button if content like this starts to perform well for us then we can continue to make a lot more high quality videos for you in the future

Channel: The Tesla Space

Views: 102,103

Keywords: supercomputers, supercomputers explained, dojo supercomputer, tesla dojo explained, frontier supercomputer, super computer, supercomputer, fastest computer in the world, how supercomputers work, how supercomputers are used, supercomputers documentary

Id: wazb0ellIwY

Length: 21min 17sec (1277 seconds)

Published: Mon Jan 01 2024