Beyond Linpack: How HPCG and Graph500 Redefine Supercomputer Rankings

TOP500 and Linpack Benchmark

The 2017 first‑half TOP500 list confirmed China’s Taihu Light and Tianhe‑2 retaining the top two positions, continuing a years‑long dominance in the Linpack (HPL) benchmark, which measures raw floating‑point performance using the HPL 2.0 specification.

Emergence of HPCG

Jack Dongarra, co‑creator of TOP500, argues that Linpack no longer reflects real‑world workloads; the High Performance Conjugate Gradients (HPCG) benchmark, which stresses memory, network latency, and more complex differential‑equation calculations, is now used to assess overall supercomputer capability.

In the 2017 HPCG half‑year ranking, China’s Tianhe‑2 achieved 580 TFLOPS (second), Japan’s K computer led with 602 TFLOPS, and Taihu Light placed third with 480 TFLOPS. Taihu Light’s HPCG/HPL efficiency was only 0.4 %, the lowest among the top ten.

Graph500 and Green Graph 500

Beyond compute‑centric benchmarks, Graph500 evaluates data‑intensive workloads using the GTEPS metric (billion edges traversed per second). The Green Graph 500 applies the same metric but ranks systems by performance per watt, mirroring the Green500’s energy‑efficiency focus.

In the 2017 Graph500 list, Japan’s K computer took first place, China’s Taihu Light was second, and Tianhe‑2 fell to eighth.

Application Benchmarks in HPC

Application benchmarks target specific scientific or engineering codes, providing more realistic performance insight. Notable examples include:

GTC‑P (Gyrokinetic Toroidal Code) : solves the Vlasov‑Poisson equation for ion motion in tokamaks using particle‑in‑cell methods.

Meraculous : a large‑scale parallel genome assembler that builds and traverses de Bruijn graphs of k‑mers.

MILC : a suite of lattice QCD codes for quantum chromodynamics simulations on MIMD machines.

MiniDFT : density‑functional theory application that constructs and diagonalises Fock matrices using FFT‑based plane‑wave transforms.

MiniPIC : solves the Boltzmann equation for electrostatic fields in arbitrary domains using unstructured hex/tet meshes and MPI‑based particle tracking.

PENNANT : advanced architecture research code handling 2‑D unstructured finite‑element meshes with MPI, OpenMP, or CUDA parallelism.

SNAP : a proxy application modelling neutral particle transport, derived from Los Alamos’ PARTISN code.

UMT : executes time‑dependent, energy‑dependent radiation transport on 3‑D unstructured grids across distributed memory systems.

Micro Benchmarks

Micro benchmarks are more generic and often appear in RFPs, focusing on pure tool performance:

Crossroads/N9 DGEMM : a multithreaded dense matrix‑multiply test measuring sustained FLOPS per node.

IOR : evaluates parallel POSIX and MPI‑IO throughput.

Mdtest : measures metadata operation performance on parallel file systems.

STREAM : assesses sustainable memory bandwidth; the Crossroads/N9 variant is an enhanced version of John D. McCalpin’s original test.

Overall Perspective

Collectively, TOP500, Green500, HPCG, Graph500 and Green Graph 500 provide complementary views of supercomputer capability—raw compute, energy efficiency, memory‑intensive workloads, and data‑centric performance. While Linpack remains a cornerstone, the ecosystem of benchmarks ensures a more holistic evaluation of modern high‑performance systems.