HPC & AI Systems

Explainer

HPC & AI Systems explained

High-performance computing (HPC) means using very large clusters of processors working together to solve problems that one machine cannot handle in a reasonable time. These clusters are called supercomputers. They run weather models, physics simulations and, increasingly, AI training.

A supercomputer is built from thousands of nodes, each with CPUs and usually GPUs, joined by a fast network and a shared storage system. Programs are split into many pieces that run at once and swap results, often using a standard called MPI (message passing interface).

How it differs from the usual approach

A normal server runs many unrelated jobs. An HPC system is tuned to run one huge job across all its nodes, so the speed of the network and the memory matters as much as the processors. Cooling and power delivery are also designed at room scale, often with liquid cooling. The same design ideas now appear in AI clusters, which blurs the line between HPC and AI infrastructure.

Where it is used

• Weather forecasting and climate modelling.
• Nuclear stockpile and materials simulation at national laboratories.
• Drug discovery and protein and molecular simulation.
• Engineering simulation such as aircraft aerodynamics and crash testing.
• Training large AI models on clusters of thousands of accelerators.

What to watch

• The TOP500 list ranks supercomputers twice a year, in June and November. El Capitan, at Lawrence Livermore National Laboratory in the US, took first place in November 2024 at roughly 1.7 exaflops on the LINPACK benchmark. It has since been joined at the top by newer exascale systems, so check the current list.
• Frontier, at Oak Ridge National Laboratory, was the first system to pass one exaflop on that benchmark, in 2022.
• Power use is a central limit. Large new systems are planned around the electricity and cooling available at the site.
• The gap between scientific supercomputers and commercial AI clusters is narrowing, since both rely on the same accelerators and networks.

Frequently asked questions