How Supercomputers Really Work: The Machine That Lost #1

Описание: In June 2026 the world's fastest supercomputer lost its crown. We take El Capitan apart: what an exaflop is, the AMD MI300A chip that puts CPU, GPU and memory in one package, a node with four of them, 87 cabinets and the HPE Slingshot network, how software slices a problem across 11,520 nodes (MPI), why Amdahl's law caps the speedup, storage and checkpoints, 29.6 megawatts of power and fanless liquid cooling, mixed-precision AI benchmarks, the new number one LineShine from China, and where India's AIRAWAT and National Supercomputing Mission stand. Numbers come from the TOP500 list, Lawrence Livermore National Laboratory, HPE, C-DAC and India's PIB; our own arithmetic is labelled.

Chapters
0:00 Cold open
0:30 Exaflops and Linpack
0:59 Meet El Capitan
1:24 The MI300A chip
1:52 CPU vs GPU
2:19 One node
2:39 Cabinets
3:00 The network
3:22 Slicing the problem (MPI)
3:46 Amdahl's law
4:12 Memory and storage
4:35 Checkpoints
5:05 Power and cooling
5:22 Liquid cooling
5:43 Efficiency
6:03 What it computes
6:27 Mixed precision
6:48 LineShine takes #1
7:34 India: AIRAWAT
7:59 National Supercomputing Mission
8:24 The big picture

What you'll see
TOP500 June 2026 (67th list): LineShine 2.198 exaflops, El Capitan 1.809, Frontier 1.353, Aurora 1.012, JUPITER Booster 1.000.
El Capitan (LLNL): 11,520 nodes, 46,080 AMD MI300A APUs, about 1.1 million CPU cores, 4 APUs and 512 GB per node, Slingshot-11, 401 PB Lustre storage at 2.6+ TB/s, 29.6 MW during Linpack (34.8 MW peak), 100% fanless direct liquid cooling, 60.94 gigaflops per watt.
MI300A: 24 Zen 4 cores, 146 billion transistors, 128 GB HBM3 (LLNL / AMD).
Our arithmetic: Amdahl's law with 99.9% parallel work on 46,080 chips gives about 978x; 1.809 exaflops vs 8.2 billion people doing one calculation a second is about 7 years; AIRAWAT's 8.5 petaflops is roughly 200 times less than El Capitan.
India: AIRAWAT at C-DAC Pune (8.5 PF sustained; ranked 75th in May 2023 and 188th in November 2025), and the National Supercomputing Mission (38 systems, 44 PF combined; PARAM Rudra built on Rudra servers made in India).

Sources
TOP500, LineShine Debuts at No. 1 (23 June 2026) - https://top500.org/news/lineshine-deb...
LLNL HPC, El Capitan system page - https://hpc.llnl.gov/hardware/compute...
LLNL, El Capitan Getting Started (March 2026) - https://hpc.llnl.gov/sites/default/fi...
LLNL, El Capitan reigns supreme across three major benchmarks - https://www.llnl.gov/article/53006/el...
HPE, El Capitan press release (Nov 2024) - https://www.hpe.com/us/en/newsroom/pr...
C-DAC AIRAWAT - https://airawat.cdac.in/airawat/resou...
PIB, PARAM Rudra at IIT Bombay (National Supercomputing Mission) - https://www.pib.gov.in/PressReleasePa...
Wikipedia, Supercomputing in India (AIRAWAT ranking, Nov 2025) - https://en.wikipedia.org/wiki/Superco...

All animation is original (procedural 2D diagrams made for Megaworks); chip and cabinet drawings are schematic, not photographs. Music: original. Narration is AI-voiced (Chatterbox-Turbo TTS, synthetic voice). Not affiliated with LLNL, HPE, AMD, TOP500 or any lab.

How It Works #7. Tell us what to take apart next in the comments.

Megaworks, mega works, Mega Works

#HowItWorks #Supercomputer #AI #Engineering #Technology #Megaworks