LINPACK / HPL

LINPACK measures floating-point performance by solving a dense system of linear equations using LU factorization. Its parallel form, High Performance Linpack (HPL), is the benchmark behind the TOP500 list of the world's fastest supercomputers, published twice a year. Because dense linear algebra is highly regular and compute-bound, LINPACK is the canonical way to report a machine's achievable FLOPS, and reaching a high HPL score has become a milestone for new HPC systems.

What It Measures

The benchmark reports performance in FLOPS (floating-point operations per second), commonly GFLOPS, TFLOPS, or PFLOPS, and at the leading edge ExaFLOPS. Two figures matter: Rmax, the maximum measured performance on the benchmark, and Rpeak, the theoretical peak from clock and execution-unit counts. Their ratio, the efficiency, shows how much of the hardware's potential the system actually realizes on this ideal workload, which exposes how well the interconnect, memory, and software stack are balanced.

Methodology

HPL solves a problem of size N, where N is chosen large enough to fill most of available memory, maximizing the ratio of compute to communication and thus the achievable efficiency. The matrix is distributed across nodes; the run uses tuned BLAS libraries (such as optimized GEMM, or GPU equivalents) and MPI for communication. Tuning the block size, process grid, and BLAS implementation strongly affects results, and getting a top score is itself an engineering effort. Reported runs disclose N, node count, interconnect, accelerators, and libraries. For TOP500, results undergo validation of the solution's residual to confirm the answer is correct, not merely fast.

How to Interpret Results

Higher Rmax is better. Read efficiency (Rmax/Rpeak) to judge how well-balanced and well-tuned a system is; high-end clusters often reach 60 to 90 percent on HPL, while GPU-heavy systems can sit lower if data movement dominates. A low ratio signals interconnect bottlenecks, untuned libraries, or memory limits. Use LINPACK to compare HPC systems and accelerators on dense floating-point capability, to validate that a new cluster is performing near its design point, and to track acceptance testing during deployment.

Limitations

LINPACK is famously favorable: dense LU factorization has a near-ideal compute-to-communication ratio and rarely resembles real scientific codes, which are often sparse, memory-bound, or communication-heavy. High HPL scores can mask poor performance on irregular workloads, which is exactly why the HPCG benchmark was created as a more representative complement that stresses memory and communication. Treat LINPACK as a peak-capability and ranking metric, not a predictor of application throughput, and read it alongside HPCG and application benchmarks.