Other meanings of LINPACK benchmarks
High-performance computing
LINPACK benchmarks are standardized computer performance benchmarks derived from the LINPACK linear algebra library. They measure how quickly a computer solves dense systems of simultaneous linear equations, usually by reporting floating-point operations per second (FLOPS). The family ranges from fixed-size historical tests to scalable distributed-memory implementations such as HPL, whose results have been used to rank the world’s fastest supercomputers.1
LINPACK benchmarks measure the numerical throughput of a computer while solving dense linear equations. The underlying computation is typically an LU factorization with partial pivoting, followed by triangular solves; the operation count is known, so elapsed time can be converted into FLOPS.1
The benchmark grew out of LINPACK, a software library and accompanying users’ guide developed for numerical linear algebra. Early versions included fixed-size tests, notably a 100-by-100 problem, while later versions allowed larger or user-selected problem sizes. The distinction matters: a fixed problem tests a particular workload, whereas a scalable test can expose the capacity of a large parallel machine.
LINPACK performance is therefore a property of a hardware-software configuration, not simply of a processor model. Compiler choices, numerical libraries, memory placement, communication networks, and problem size all influence the reported result.
Modern supercomputer rankings generally use High Performance LINPACK (HPL), a portable implementation designed for distributed-memory systems. HPL distributes a dense matrix across many nodes, factors it using blocked algorithms, and coordinates the work with message passing, commonly through MPI.2
The TOP500 project uses the achieved result, called Rmax, to order systems, while also publishing Rpeak, an estimate based on processor and accelerator capabilities. The gap between the two values indicates how effectively the complete system sustains the benchmark workload; it is not itself a universal efficiency rating.
HPL’s adjustable matrix size lets system operators select a problem that fits available memory. This flexibility makes the test practical for machines ranging from research clusters to the largest national supercomputers, but it also means results require context about configuration and scale.3
LINPACK benchmarks are strong indicators of dense floating-point performance, but they are not general-purpose measures of computer speed. The workload emphasizes arithmetic intensity and regular matrix operations; it says less about databases, operating-system responsiveness, sparse matrices, irregular graph algorithms, file systems, or many ordinary applications.
Large systems can also obtain different results from the same nominal hardware through tuning. Choices involving block size, process mapping, look-ahead, accelerator kernels, compiler optimization, and network topology affect both computation and communication. A high ranking consequently demonstrates an optimized capability for this workload, not necessarily superior performance in every scientific or commercial task.
For that reason, researchers often pair LINPACK results with application benchmarks and other measures of memory bandwidth, communication, energy use, or resilience. The benchmark remains useful precisely because its mathematical problem is reproducible and its principal operation count is well defined, even when its scope is deliberately narrow.1
The benchmark family contains more than the single number commonly quoted in rankings. Historical LINPACK tests used fixed matrix dimensions, while parallel variants introduced larger, adjustable problems; HPL is a later implementation rather than the name of the original library itself.2
A subtle feature is that matrix size is constrained by memory and distribution strategy. Increasing the problem size can improve the ratio of arithmetic work to communication, but it can also increase execution time and expose failures in memory capacity or network coordination. Reported precision, software version, processor count, and accelerator use therefore matter when comparing results.
LINPACK also helped establish a shared vocabulary for supercomputing performance: gigaflops, teraflops, and petaflops describe rates of floating-point work, while Rmax records what the system actually sustained. These terms describe scale, not automatically scientific usefulness or energy efficiency.3
LINPACK benchmark results should be compared only when problem definition, precision, hardware configuration, software, and reporting conventions are understood.
Help improve the encyclopedia. Reports go straight to the site manager.