UK What high performance computing means
High performance computing (HPC) is the practice of splitting a large computing job across many processors so that they work on it at the same time. Instead of one server taking weeks to run a simulation, hundreds or thousands of processor cores and GPUs share the work and finish in hours.
An HPC system is usually a cluster: many servers, called nodes, linked by a low-latency network and managed as one machine. The largest clusters are called supercomputers. Performance is measured in floating-point operations per second (FLOPS). The TOP500 list, published twice a year, ranks the world’s fastest systems using the HPL (LINPACK) benchmark, and in 2022 Frontier at Oak Ridge National Laboratory became the first system on the list to pass one exaflop, a billion billion operations per second.
HPC used to mean scientific simulation on CPUs. Today most new systems are built around GPUs, and the same clusters train AI models as well as running physics.
How HPC works
An HPC job is divided into many smaller tasks that run in parallel and exchange results as they go. A weather model, for example, splits the atmosphere into a grid and gives each node a block of it, with neighbouring nodes swapping boundary values at every time step. Software built on MPI (Message Passing Interface) handles that exchange. Every cluster combines the same building blocks:
-
//
Compute nodes. Servers with high-core-count CPUs such as AMD EPYC or Intel Xeon, and plenty of memory.
-
//
Accelerators. GPUs that run the parallel maths, from data centre cards in PCIe slots to multi-GPU boards linked by NVLink.
-
//
Interconnect. A fast, low-latency network, typically InfiniBand or Ethernet with RDMA, so nodes can share data without waiting on each other.
-
//
Parallel storage. File systems such as Lustre or IBM Storage Scale that let every node read and write at once.
-
//
Scheduler and software. A job scheduler such as Slurm queues work and assigns nodes, alongside MPI, CUDA and the application itself.
What an HPC cluster looks like
A small cluster might be a single rack: a head node that users log in to, a few dozen compute nodes, a storage server and a network switch. A national supercomputer fills a hall with thousands of nodes and needs its own liquid cooling and power plant. Many organisations now rent HPC capacity from cloud providers by the hour, and keep an on-premises cluster for steady workloads or data that cannot leave site.
| System | Typical scale | Processors | Network | Typical use |
|---|---|---|---|---|
| Workstation | One computer | 1 CPU, 1–2 GPUs | Standard Ethernet | Design, small simulations, model development |
| GPU server | One rack server | 1–2 CPUs, up to 8 GPUs | 10GbE to 100GbE | AI training and inference, rendering |
| HPC cluster | Tens to hundreds of nodes | CPUs and GPUs in every node | InfiniBand or RDMA Ethernet | Engineering simulation, research, AI training |
| Supercomputer | Thousands of nodes | Tens of thousands of GPUs or CPUs | Custom or high-end InfiniBand | Climate, physics, national research |
The boundaries are loose: a single server with eight GPUs can outperform an older cluster, and what counts as HPC depends on whether the work is spread across many processors in parallel.
What HPC is used for
HPC handles any problem where the maths is too big for one machine.
-
//
Weather and climate. National forecasting centres run global atmosphere and ocean models several times a day.
-
//
Engineering simulation. Computational fluid dynamics and finite element analysis test aircraft, vehicles, turbines and structures before anything is built.
-
//
Life sciences. Genome sequencing, protein structure prediction and drug screening.
-
//
Energy. Seismic imaging for oil, gas and carbon storage, and modelling of power grids and fusion experiments.
-
//
AI training. Large language and vision models are trained on GPU clusters built the same way as HPC systems.
HPC at the edge
Some HPC work cannot wait for data to reach a data centre. Radar and sonar processing, signals intelligence, airborne mapping and autonomous vehicle test fleets generate data faster than any link can carry it, so the computing goes into the vehicle, ship or field site. That calls for HPC-class hardware in rugged, often short-depth chassis built for vibration, wide temperatures and limited power.
Our edge AI and rugged HPC servers include AMD EPYC systems with dual NVIDIA RTX PRO 6000 Blackwell GPUs, and our rack mount servers range up to dual Intel Xeon 6 GPU servers that take H200 NVL cards. For processing power figures, see our explainer on TFLOPS and how it differs from TOPS.
Quick answers on high performance computing
High performance computing: common questions
What does HPC stand for?
High performance computing. In computing it means using many processors in parallel to run jobs too large for one machine.
What is an HPC cluster?
A group of servers that work as one system. Compute nodes, storage and a head node are joined by a fast network and managed by a job scheduler, so one job can run across all of them at once.
Is HPC the same as a supercomputer?
A supercomputer is the top end of HPC. All supercomputers are HPC systems, but most HPC runs on smaller clusters and GPU servers in universities, engineering firms and the cloud.
What is HPC in cloud computing?
HPC capacity rented on demand. Cloud providers offer GPU instances, fast networking and parallel storage by the hour, which suits short projects and peaks without buying a cluster.
What is the difference between HPC and AI?
HPC is the infrastructure; AI is one workload. AI training runs on HPC-style GPU clusters, and HPC centres increasingly mix simulation with machine learning on the same hardware.
We build GPU servers and rugged HPC systems with the operating system, GPU drivers and CUDA stack installed and validated as standard. Tell us the workload, the site and the power and cooling available, and we will specify the processors, GPUs and networking to match.