High-performance computing (HPC)

High-performance computing refers to clusters of powerful, tightly interconnected computers, often thousands of processor cores plus specialized accelerators like GPUs, working together on a single computation that would be impractical on one machine: climate models, particle-physics simulations, large-scale genomics analysis or training large machine learning models.

HPC clusters are typically shared, scheduled resources: researchers submit jobs to a queue managed by scheduling software, which allocates nodes and enforces time and resource limits, rather than running work interactively on a personal machine. HPC differs from ordinary cloud computing mainly in interconnect: HPC systems use very low-latency, high-bandwidth networking between nodes so a single tightly coupled simulation can span hundreds of machines. It is a specific, tightly coupled form of the broader category of distributed computing, and many HPC jobs use massively parallel processing to split work across nodes.

Access to HPC is often the limiting factor in computational science, and jobs are commonly wrapped in a reproducible scientific workflow so runs can be repeated or scaled. Common pitfalls are underestimating how much of a workload cannot be parallelized, which caps the speedup from adding more nodes, a limit formalized as Amdahl's law, and treating queue wait time as free; in practice it is often the largest share of a project's calendar time.