High-Performance Computing
Aggregating large amounts of computing power to solve problems too big for a single machine.
Definition
High-performance computing (HPC) uses clusters of many interconnected processors, often with accelerators, to solve large-scale computational problems by running work in parallel. It is the tool of choice for simulation, modeling, and large-scale data analysis.
The field has shifted from chasing raw clock speed to managing parallelism and data movement, since the energy cost of moving data now often exceeds that of the arithmetic itself. Modern HPC design is as much about the memory hierarchy and interconnect as about the processors.
The field is defined today less by peak arithmetic speed than by the cost of moving data, since fetching operands from distant memory or another node dwarfs the cost of the arithmetic itself. Consequently, the memory hierarchy, the interconnect, and the storage system shape performance as much as the processors. Designing for HPC means designing for data locality and communication, a mindset different from ordinary programming.
Building blocks
- Compute nodes with multi-core CPUs and often GPUs.
- A fast interconnect linking nodes with low latency.
- Parallel file systems for high-throughput storage.
- A scheduler that allocates jobs across the cluster.
Why it matters
Many scientific and engineering problems are simply infeasible on one machine. HPC makes high-resolution simulation and large parameter studies practical, and its performance is limited as much by communication and memory bandwidth as by raw arithmetic speed.
Fusion connection
Kronos runs its plasma physics and engineering simulations for Hyperion on HPC resources, where parallelism and scalability determine what resolution and how many design cases are achievable.