I've been studying computer science for some time, and I always assumed I had a solid grasp of how computers work. They process data, execute programs, and occasionally freeze when I overload them with too many Chrome tabs.
Simple enough, right?
Then I started reading about supercomputers, and suddenly everything I thought I knew felt tiny. The scale, the speed, and the complexity were on a completely different level.
It all began with a project for my Parallel Computing class. The assignment seemed simple: pick two supercomputers from the TOP500 list and compare them. I went to the site and started scrolling. The first name that caught my eye was El Capitan. Next to it was the word "exaflops," which I didn't really understand, but it sounded impressive.
For the second one, I chose Fugaku, mostly because the name sounded cool. Later I learned it's actually named after Mount Fuji.
What I expected to be a simple research task quickly evolved into a deep dive into a world where computers don't just run code-they help simulate the universe.
Discovering El Capitan
Why the Name Fits
El Capitan is the fastest supercomputer on Earth. Its name comes from Yosemite’s famous granite monolith, which feels right because this machine is basically a mountain of computing power.
- Built by Hewlett Packard Enterprise (HPE)
- Powered by AMD MI300A APUs
- Lives at Lawrence Livermore National Laboratory, California
Mind-Blowing Performance
- 1.742 exaFLOPS which equals 1.742 quintillion calculations per second
- If every person on Earth (8 billion people) performed one calculation per second non-stop, it would take us about 7 years to match what El Capitan does in one second.
The real magic is not just speed. El Capitan does not rely on one giant processor. It has millions of tiny compute units working together at the same time.
Inside the Chips
At the heart are AMD’s MI300A APUs (Accelerated Processing Units):
- 24 CPU cores for general-purpose computing
- CDNA 3 GPU cores for heavy math operations like AI or simulations
- 128 GB of high-bandwidth memory shared between CPU and GPU
Think of it as having two chefs sharing the same kitchen instead of passing ingredients back and forth between separate kitchens.
Parallel Computing at Extreme Scale
This is where things get wild. All 44,544 APUs are connected through HPE’s Slingshot-11 network. I spent a long time trying to understand it; the technical papers are dense, but basically it lets all the nodes communicate with extremely low latency.
Parallel computing works like this: a massive physics simulation is broken into thousands of chunks. Each APU handles one chunk, but they constantly share information, synchronize results, and coordinate next steps.
Keeping Cool
All that processing power generates heat. El Capitan uses direct liquid cooling, with chilled water flowing along each chip. It is quieter and more efficient than air cooling and allows the system to run continuously at around 30 megawatts, about the power usage of a small town.
Real-World Impact
El Capitan isn’t just about speed. It’s used to tackle important real-world problems:
- nuclear weapons stockpile stewardship: It helps scientists keep nuclear weapons safe using computer simulations, without testing them.
- Designing new materials
- Modeling climate systems
- Studying exploding stars
If your laptop is a calculator, El Capitan is a universe simulator.
Exploring Fugaku
If El Capitan is a mountain of compute power, Fugaku is its twin across the Pacific. Its name comes from Mount Fuji, and it lives up to it.
A Different Approach
- Developed by Fujitsu and RIKEN, Japan
- CPU-only design, no GPUs
- Uses A64FX processors with Arm’s Scalable Vector Extension (SVE)
Each A64FX chip:
- 48 compute cores plus 4 assistant cores" (52 total cores per A64FX)
- Direct connection to HBM2 memory for high bandwidth
- Feeds data fast enough so the CPU cores never wait
Extreme Parallelism
When I first saw “158,976 nodes,” I thought it was a typo. That’s 158 thousand separate computers, each with 48 cores, all connected through the Tofu-D interconnect.
All these nodes work together like one giant computer. They split up big problems, like climate simulations, and constantly share data so every node knows what its neighbors are doing.
Real-World Applications
Fugaku handles massive problems like:
- Climate modeling
- Earthquake simulations
- Drug discovery
- Pandemic research-this one really stood out. During COVID-19, Fugaku simulated how respiratory droplets spread in places like trains, offices, and restaurants. These simulations helped determine better ventilation and social distancing guidelines.
Efficient Power Use
- Water-cooled design keeps it stable at 30 MW
- Once topped the Green500 list for energy efficiency
Fugaku shows that supercomputers don't need GPUs to achieve incredible performance-innovative CPU design and networking can be equally powerful.
Two Supercomputers, Two Approaches
Both El Capitan and Fugaku demonstrate the power of parallel computing. They split huge problems into smaller tasks and run them at the same time, but each takes a different approach:
| Feature | El Capitan | Fugaku |
|---|---|---|
| Manufacturer | HPE/AMD | Fujitsu/Arm |
| CPU/GPU | Hybrid MI300A APUs | CPU-only A64FX |
| Nodes | 11,039,616 cores (CPU+GPU) | ~7.6 million cores |
| Memory | Shared HBM3 on chip | HBM2 per CPU |
| Interconnect | Slingshot 11 | Tofu-D |
| Peak performance | 1.742 exaFLOPS | 442 petaFLOPS |
| Cooling | Direct liquid | Water-cooled racks |
Both demonstrate how parallel processing can solve problems that would take conventional computers decades or centuries to complete.
Final Thoughts
I started this project thinking, “Pick two supercomputers, compare specs, done.” Three all-nighters later, I’d watched installation timelapses, read way too many AMD whitepapers, and actually got excited about interconnect topologies.
Here’s what stuck: these machines aren’t just “fast computers.” They represent completely different ways of thinking about computation. El Capitan’s unified memory architecture solves problems that plagued GPU computing for years. Fugaku shows that sometimes elegance beats brute force.
Parallel computing isn’t just running tasks at the same time. It’s about orchestrating millions of processors to tackle massive problems while constantly syncing and sharing data. A task that might take nearly an hour on a single computer can finish in just a few minutes when split across multiple processors. These principles aren’t limited to supercomputers. Modern AI training, weather simulations, video games, and many software applications all rely on distributing work efficiently. The key shift in thinking is asking not just “How do I make this faster?” but “What can run simultaneously?” This mindset enables work on challenges big enough to require machines like El Capitan and Fugaku.
And I’ll never complain about my laptop being slow again, El Capitan does more in one second than my computer could in multiple human lifetimes.
Honestly? That’s the coolest thing I’ve learned all semester.
Top comments (0)