Computers used to take up a whole room and were only as powerful as a basic calculator!

Since then, computers have  gotten a lot more impressive. And I’m not talking about the one  you’re watching this video on, or even the ones powering groundbreaking  experiments at your local university.

Although those are very cool,  and they’re doing a great job. I’m talking about the world’s  most powerful supercomputers. These are massive machines doing  trillions of calculations every second to process data about everything  from the tiniest cells in our bodies to distant galaxies.

If your laptop is a minivan,  these things are like Ferraris. And lucky for us, there’s a Top500  list that keeps track of them. I’m not a supercomputer, so I  can’t tell you about all of them.

But I can tell you about the top 5. Including what makes them so powerful, where they get that power  from, and what on Earth … or in space … they’re doing with it. [intro music] Let’s start with the fifth most  powerful computer in the world: the HPC6 from the Italian  energy company Eni, in Rome. This computer was just switched on in 2024, and it has a whopping 477.9 petaflops of power!

No, that’s not the name  for how it lands in a pool. In fact, keep this computer far  away from large bodies of water. A petaflop is a unit to  measure a computer’s power, which is basically how fast  it can carry out calculations.

For computers, just like sports cars, this  is the name of the game. Speed = power. Flops is an acronym which stands for “floating point  operations per second.” It still sounds like it might  have something to do with water … or maybe beach sandals.

But anyways, here’s what it’s really about. Computers work by breaking  down a task into basic math,   adding and multiplying  numbers, trillions of times. Each addition or multiplication is a “flop.” The “peta” part refers to the fact that these  computers do huge amounts of “flops” every second, as in something like 10 to the 15th power!   a standard desktop computer  is generally in the billions of flops.

But the HPC6 does more than 400 quadrillion  additions or multiplications every second! It needs that much power because  Eni is an energy company. A lot of HPC6’s computing power  goes toward energy research like developing better batteries.

As cool as that research is, the most interesting thing about HPC6 is that it’s an example of how high-performance  computing can be done sustainably. Supercomputers like HPC6 have a pretty  hefty energy impact for two main reasons: firstly, they need energy to run, and two, they need energy to power  the systems that cool them down.   But HPC6 is located at Eni’s Green  Data Center that efficiently controls   the temperature of computers,  using direct liquid cooling. That means they’re running a solution that’s weirdly similar to antifreeze  close to the computer chip.

This antifreeze solution pulls  heat from the computer directly, which is more efficient than letting  it just spread out into the air. Then, the facility can even redirect some of the   heat from the computer to other parts of  the building when the weather is cold. So even though it’s one of the most  powerful computers in the world, HPC6 has an astonishingly low energy impact.

It uses one third or less power than  most of the other Top 5 supercomputers! Speaking of, at number four on the list, we have Eagle, with a power of 561.2 PFlops. And the cool thing about this one is that it’s not just used for specialized  tech and research applications.

It’s a supercomputer you can  access from your own couch. That’s because Eagle is part of  Microsoft’s Azure cloud computing services, which provide computing power to a  large number of individual users. “Cloud” computers aren’t based in one location. Instead, their servers, storage,  and software are in the cloud, or accessed using the internet.

This is useful because cloud computers can split up a lot of power into smaller pieces that can be sent to customers as needed. See, huge computers like  Eagle have tons of “cores”, distinct computing units that can  be divided out to different tasks. Eagle has more than 2 million of them  in data centers around the world, and each core can do calculations on its own.

If a user needs more than one, they need  to be able to communicate efficiently. And since we’re in “computer time”, on the  scale of trillions of calculations per second, this means fast. Eagle does this by using a  networking method called InfiniBand that connects servers and individual computers.

While many networks become slower as they try to process more and more data at the same time, Inifiniband sends  packets of data one at a time, allowing data to move through the  network quickly and efficiently. This communication makes a wide  variety of applications possible, from training machine learning  models to setting up a huge database. So Eagle’s power comes not just from  its ability to do calculations quickly but from its ability to communicate  those calculations quickly with itself.

The next three computers on the list  take power to a whole new level. They’re all “exascale” computers,  meaning they can reach 10^18 Flops per second. And at 1,012 PFlops, the world’s  third most powerful computer is Aurora at Argonne National  Lab in Lemont, Illinois. his computer is twice as powerful as Eagle.

One reason for its performance  is its standout GPU. A computer’s GPU and CPU … the Graphics Processing Unit  and Central Processing Unit … act as the computer’s brain. They’re the things doing all those  calculations and making the computer compute.

But while the CPU processes a wide  variety of tasks one after the other, the GPU is more specialized. It’s particularly good at doing a  lot of the same task in parallel, or at the same time. And as it happens, this is useful  for a lot of scientific applications, which are Aurora’s bread and butter.

Argonne and other labs use Aurora as  a resource for computational research, which can leverage supercomputer  power in a few different ways. First, Aurora can be used to analyze data. We have tons of scientific data out there, ranging from telescope images to genome sequences, and it takes lots of computing  power to make that data usable or get any information from it.

Second, Aurora can be used to generate data. For example, simulating a system we don’t have data for at the required scale, like the constant tiny motions  of molecules in our body, and how they interact with one another. And third, Aurora can be used  to make predictions from data.

Training machine learning models  on existing data can find patterns that let us use what we know to make  educated guesses about what we don’t. All three of these applications  involve doing the same thing over and over and over again  on slightly different inputs, making them perfect tasks for a good GPU. And since the hardware is up to the task, researchers using Aurora are building tools that make the best use of the  computer’s power and speed.

For example, one of these tools is a  machine learning model called MProt. Scientists can use it to study  proteins, important molecules that allow cells and organs to function. Being able to change proteins or make new ones can be helpful for things like  improving agriculture, treating disease, and even breaking down plastic in the environment.

Of course, running machine learning models can take a lot of time and  a lot of computing power. Luckily, Aurora has plenty of the latter, and MProt is designed to  run tasks at the same time. This makes it faster and easier for  scientists to get the predictions they need.

We’re getting close to the  fastest computer on Earth. But before I tell you about  the last two, a quick ad. Thanks to Saily for supporting this SciShow video!

Saily is an eSIM app that helps you stay connected with a local SIM card in over 200 places  around the world from a single installation. That’s the power of eSIMs. Just download  an app, pick a plan, and install.

It’ll activate instantly when  you land in that cool new spot. You can choose between a global or regional plan depending on where your travels will take you. And all Saily eSIM plans are  compatible for iOS or Android devices.

If it doesn’t work on your  phone, you get a full refund. And even if it does, you have  a 30-day money-back guarantee. And there’s chat support available 24/7  for help at any point along that process.

Even SciShow’s Staff Writer went on a weekend trip that would have been a lot smoother with Saily. She took a Spanish rideshare  from Madrid to Salamanca. But at the end of the trip, she couldn’t find  the meetup spot for the car back to Madrid!

The driver said they were across town. So she had to pull up a map  fast with no time to find wifi. After spending more money than  she would have if she had Saily, she made it home.

You don’t have to be like Emma. You can download the Saily app  and use code SCI at checkout for an exclusive 15% discount  on Saily eSIM data plans. We’re getting close to the  fastest computer on Earth.

Aurora’s power is just beat  out by the number two computer. Frontier, at Oak Ridge National Lab in Tennessee, clocks in at 1,353 PFlops. Like Aurora, it’s used to  accelerate scientific research.   But it beats Aurora in power  because a computer’s power is about more than just the computer itself.

On top of the energy, space, and cooling  systems it takes to run a supercomputer, its power also relies on software  that uses that hardware effectively. An inefficient application slows the  computer’s performance way down.   To figure out the best software  implementations for Frontier, Oak Ridge National Lab launched the Center  for Accelerated Application Readiness, or CAAR, project. This project studied a variety of  scientific software across different fields, from biology to physics, to  optimize the use of Frontier.

One example is CHOLLA, a software for simulating the behavior  of gases inside galaxies. This type of simulation can  help scientists understand how different types of galaxies formed over time, and predict the fates of stars  and supernovae within them. CHOLLA is optimized for  parallel computing on GPUs, which is already pretty good  for a scientific supercomputer.

But to really take advantage of  Frontier’s hardware and organization, scientists had to make a few adjustments. Originally, CHOLLA split its  calculations between the GPU and CPU. This meant that most of the data  stayed on the CPU side of things while the program ran, and the program needed to spend extra  time sending data back and forth.

But with exascale computers like Frontier, the GPUs are usually the star of the show, and it’s more efficient to not  make them share the limelight. So in a new version of CHOLLA, researchers moved data storage and multiple  calculations over to the GPU, which was quite a task. It involved some heavy  editing of the original code, and writing a whole new software package!

But their hard work paid off,  because the new version of CHOLLA could run on Frontier 8 times faster than before. Thanks to the optimizations, the code now works with Frontier’s  hardware to get lightning-fast performance, helping scientists simulate galaxies  at a rate that’s out of this world. And that brings us to the world’s  top most powerful supercomputer: the aptly-named El Capitan,  with 1,742 PFlops of power El Capitan is also a supercomputer used  for scientific research at national labs.

But it edges out Frontier and Aurora  because it’s just built different. See, its cores use an architecture that combines the powers of  the CPU and the GPU into one – the APU, or Accelerated Processing Unit. Combining them makes it easier  to send calculations to each GPU, and makes parallel computations a lot faster.

This boost in speed opens doors  for even heftier applications,   including one known as “Cognitive  simulation”, or “CogSim”. This application involves three  parts: analyzing experimental data, running simulations, and  training machine learning models. That’s all the stuff Aurora can do.

The difference with El  Capitan is that it’s working to make machine learning predictions more accurate by doing all of those things together. Scientists at El Capitan’s home  lab in Livermore, California, use simulation, machine learning,  and experiment in a cycle. Experimental data and simulated  data help train models, which can be used to predict experimental  results and improve simulation.

Those predictions then go back  into running new simulations and designing new experiments, and the cycle continues. With this iterative process, each new step doesn’t just provide more insight  into the system being studied, it also slowly improves the  methods we use to study it. They call this process cognitive simulation because it uses machine  learning to more systematically, or “intelligently”, improve simulations.

And implementing a m ore efficient  way to blend together simulation and experiment using machine learning  doesn’t just inform new experimental designs; it can also automate them. One group used CogSim during data collection on  how lasers interact with high-energy plasmas, basically swimming pools of dense,  super hot electrons that resemble stars. The idea was to help scientists  understand the extreme environments that facilitate things like nuclear fusion.

Using CogSim models sped up experiments by automatically adjusting the lasers and iterating on each previous experiment, which helped scientists build better  models out of their data in real time. As you might expect, looping through  these steps is a long and slow process. But El Capitan’s APU gives it a power boost  that lets scientists speed up the loop, iterating through the pieces  of scientific discovery faster than any previous computers could.

These five sports cars of the computing world are making huge strides in  technology and scientific research, allowing us to learn more things  more efficiently than ever before. They still take up entire rooms, but we’ve come a long way  from the basic calculator. And the Top500 list is  constantly changing!

Next year, a new supercomputer might dethrone El Capitan. The bigger they are, the harder they Pflop. [ outro ]