HPL — High Performance LINPACK
HPL solves a dense system of linear equations using LU factorization with partial pivoting. It has been used to rank the TOP500 supercomputers since 1993 and remains the universal reference for comparing HPC system performance — despite being an optimistic measure of peak capability rather than real-world performance.
What HPL Measures
Sustained floating-point performance (TFLOPS or EFLOPS) on dense matrix operations. Solves Ax=b using LU factorization — a best-case workload for modern CPUs and GPUs, heavily optimized by vendors using BLAS libraries like MKL, OpenBLAS, or cuBLAS.
Key Tuning Parameters
N (problem size) — fill ~85% of total RAM for maximum efficiency. NB (block size) — typically 192–384, must be tuned per architecture. P×Q (process grid) — must equal total MPI ranks; keep P/Q ratio close to 1.
Efficiency Targets
HPL efficiency = sustained / theoretical peak FLOPS. World-class systems achieve 60–90%. Frontier (1.206 EFLOPS, 60.3% efficiency), Aurora (1.012 EFLOPS, 50.6%). Below 50% suggests tuning opportunities or hardware issues requiring investigation.
TOP500 Rankings
Published biannually (June and November) since 1993. The world reference for tracking HPC progress. Frontier (USA, 1.206 EFLOPS) is the current #1. All systems require a verified HPL run submitted to the TOP500 committee for inclusion.
Limitations of HPL
HPL measures best-case peak compute — not representative of most real workloads. Scientific applications that are memory-bandwidth bound or sparse achieve only a small fraction of HPL performance. Always request HPCG alongside HPL for a complete picture.
BLAS Libraries
HPL performance depends critically on the BLAS implementation: Intel MKL for Xeon systems, OpenBLAS (open-source), NVIDIA cuBLAS for GPU-accelerated HPL. Vendor-optimized BLAS routinely delivers 95%+ of theoretical peak on matrix operations.
Acceptance Testing Use
HPL is the standard component of HPC acceptance testing. Vendor contracts typically specify a minimum HPL efficiency (e.g. ≥70% of peak) that must be demonstrated before payment. A failed HPL test triggers vendor remediation of hardware or software configuration.