Supercomputers: The fastest computers in the world

In short: Supercomputers connect tens of thousands of processors and GPUs into one machine that runs simulations and AI training at unprecedented speed.

In more detail: A supercomputer is not a single super-processor but a cluster of very many ordinary servers (nodes), connected fast and programmed together. Performance is measured in FLOPS (floating-point operations per second). Current top systems reach exaflops, i.e. over 10¹⁸ computing steps per second.

In Depth

Structure

• Nodes with several CPUs and usually many GPUs as accelerators.
• Fast network (InfiniBand, Slingshot) with very low latency.
• Parallel file system (Lustre, GPFS) that can serve many nodes at once.
• Cooling: mostly direct liquid cooling; power consumption is in the double-digit megawatt range.
• Operating system: almost always Linux. A job scheduler (Slurm) distributes the jobs.

What they are used for

Weather forecasts and climate models, flow and crash simulation, drug research, astrophysics, materials research, nuclear fusion and training large AI models. They are programmed with MPI (message passing between nodes), OpenMP (threads), CUDA and other GPU models.

Rankings

The TOP500 list appears twice a year and ranks with the LINPACK benchmark. The HPCG benchmark and the Green500 list (energy efficiency) complement it. In Germany large systems stand for example in Jülich (JUPITER, European exascale computer), Munich (LRZ) and Stuttgart (HLRS).

Putting it in perspective

A current smartphone reaches trillions of computing steps per second; the difference to a supercomputer is a factor of several million. Access is via applications by researchers. Cloud providers offer similar clusters for rent. For most tasks, however, an ordinary computer or server is enough.

See also: Linux