Understanding High‑Performance Computing (HPC): Market Size, Technologies, Metrics, and Core Components

Over the past 15 years, High‑Performance Computing (HPC) has become one of the fastest‑growing IT markets, often outpacing sectors such as online gaming and tablets. Forecasts predict the global HPC server market will reach $148 billion by 2021, with the entire HPC ecosystem exceeding $300 billion.

HPC refers to computing systems that use many processors within a single machine or a cluster of machines to solve large problems by dividing them into many smaller, parallel tasks. These tasks run concurrently on different nodes, and their results are combined to produce the final solution, dramatically reducing execution time.

Based on the relationship between parallel tasks, HPC can be classified into two main categories:

High‑Throughput Computing (HTC) : workloads consist of many independent tasks with little or no inter‑task communication, often used for massive data searches and Internet‑scale computing. HTC aligns with the SIMD (Single Instruction/Multiple Data) model.

Distributed Computing : workloads are split into parallel tasks that require tight coupling and extensive data exchange, fitting the MIMD (Multiple Instruction/Multiple Data) model.

Typical HPC systems comprise four parts: compute, storage, network, and cluster management software. The dominant processors are x86‑based CPUs running Linux, assembled in blade configurations, with interconnects such as InfiniBand (IB) and 10 GbE.

HPC nodes are categorized as:

MPI (thin) nodes – dual‑socket servers.

Fat (large‑memory) nodes – multi‑socket servers with abundant RAM.

GPU‑accelerated nodes – equipped with graphics processors for massive parallel floating‑point performance.

GPU vendors include NVIDIA (graphics and compute cards like K2000, K4000, K20X/K40M/K80), Intel Xeon Phi (compute cards such as 5110P, 3210P, 7120P, 31S1P), and AMD (combined graphics/compute cards like W5000, W9100, S7000, S9000, S10000).

Performance metrics are expressed in FLOPS: MFLOPS (10⁶), GFLOPS (10⁹), TFLOPS (10¹²), PFLOPS (10¹⁵), and EFLOPS (10¹⁸). CPU performance can be estimated by the formula:

Node Performance = CPU Frequency × Core Count × Number of CPUs per Node × Instructions per Cycle

. Latency of memory and disk accesses is also a critical metric.

The most widely used benchmark for measuring HPC floating‑point performance is the LINPACK benchmark, which solves dense linear systems using Gaussian elimination. LINPACK results form the basis of the TOP500 list and are complemented by other benchmarks such as TPC‑C, IOmeter, and STREAM.

Memory in HPC clusters includes three DIMM types:

UDIMM – unbuffered, low‑cost, high‑speed but less stable for large workloads.

RDIMM – registered, stable, scalable, and more expensive.

LRDIMM – load‑reduced, offering higher speeds and lower power consumption at a higher cost.

Non‑volatile DIMMs (NVDIMM) replace battery‑backed modules with super‑capacitors and flash storage, providing longer data retention without the environmental concerns of batteries.

InfiniBand is the primary high‑performance interconnect in HPC, offering low latency, high bandwidth, and RDMA support. Major vendors are Mellanox, Intel, and QLogic, with product families supporting FDR, QDR, and EDR speeds. Host Channel Adapters (HCA) and Target Channel Adapters (TCA) serve as the endpoints for IB connections.

Parallel file systems are essential for TOP500 systems; examples include Lustre, GPFS, Hadoop, NFS, and others. These distributed file systems present a unified namespace to users while spreading data across many nodes, simplifying data management and improving scalability.

For further reading, the article links to analyses of AI‑driven HPC solutions, data copy management technologies, and detailed market evolution reports.