High Performance Conjugate Gradients (HPCG)

High Performance Conjugate Gradients (HPCG)
Name	High Performance Conjugate Gradients
Acronym	HPCG
Introduced	2013
Purpose	Supercomputer benchmarking
Developers	Consortium of national laboratories and vendors

High Performance Conjugate Gradients (HPCG) High Performance Conjugate Gradients (HPCG) is a supercomputer benchmark designed to complement Top500 rankings by measuring the performance of modern high-performance computing systems on sparse linear algebra workloads. It emphasizes memory access, interconnect latency, and algorithmic characteristics typical of real-world applications such as computational fluid dynamics and structural analysis, providing an alternative perspective to LINPACK-based metrics used by the TOP500 project. HPCG aims to reflect application-relevant performance for systems from national laboratories to commercial data centers.

Overview

HPCG evaluates performance using a conjugate gradient-like solver on a sparse symmetric positive-definite system and combines operations resembling those found in production applications from institutions such as Argonne National Laboratory, Lawrence Berkeley National Laboratory, and Oak Ridge National Laboratory. The benchmark executes a sequence of kernels including sparse matrix-vector multiply, multigrid preconditioning, and vector operations, stressing subsystems designed by vendors like Intel Corporation, NVIDIA, AMD, IBM, and interconnects from Mellanox Technologies and Cray Inc.. Results are reported in terms of GFLOP/s and are intended to complement metrics produced by projects such as High Performance LINPACK and communities represented by organizations like IEEE and ACM.

History and Development

HPCG emerged from discussions at workshops attended by representatives of Department of Energy, European Union research initiatives, and national laboratories including Los Alamos National Laboratory and Sandia National Laboratories. Its development involved contributors from commercial vendors including Microsoft Corporation, Google, and Oracle Corporation alongside academic groups at Massachusetts Institute of Technology, Stanford University, University of California, Berkeley, and University of Illinois Urbana-Champaign. The benchmark was announced to address criticisms raised by leaders such as those at SEAWOLF and evaluations from analysts at Gartner, Inc. and think tanks like RAND Corporation. Early design choices were influenced by algorithmic research linked to figures such as Gene H. Golub and Michael L. Overton and computational projects like Blue Gene and Titan.

Benchmark Design and Methodology

HPCG’s methodology uses a geometric multigrid preconditioner applied to a 3D finite-difference discretization, orchestrated by a conjugate gradient kernel, reflecting implementations found in codes developed at Princeton University, California Institute of Technology, and University of Cambridge. The benchmark specifies data layouts, halo exchange patterns, and tolerance criteria similar to libraries such as PETSc, Trilinos, and Hypre. It stresses hardware characteristics pioneered by platforms including Fujitsu, Hewlett Packard Enterprise, and Huawei by exercising memory hierarchy, PCIe links, and network topologies like InfiniBand, Omni-Path, and Ethernet. Validation and verification procedures mirror best practices from agencies like National Institute of Standards and Technology and use test harnesses similar to those in SPEC suites.

Implementation and Software

Reference implementations of HPCG are provided in C++ and use MPI for parallelism with optional threading via OpenMP and acceleration through CUDA for NVIDIA GPUs or HIP for AMD GPUs. Community-contributed ports exist for software stacks maintained by Canonical Ltd., Red Hat, and package managers from Anaconda, Inc., while containerized deployments leverage platforms from Docker, Inc. and orchestration from Kubernetes. Integrations with performance tools from Intel VTune, NVIDIA Nsight, and profiling suites from Cray and HPC Toolkit support tuning efforts. Source repositories and issue tracking have been hosted by organizations such as GitHub, Inc. and collaborative projects coordinated at XSEDE.

Performance Results and Comparison

HPCG results are published alongside LINPACK numbers for systems at facilities like Argonne Leadership Computing Facility, Oak Ridge Leadership Computing Facility, National Energy Research Scientific Computing Center, and cloud offerings from Amazon Web Services and Google Cloud Platform. Typical GFLOP/s measured by HPCG are orders of magnitude lower than LINPACK for the same system due to memory- and communication-bound kernels; comparisons have been analyzed in studies by researchers at ETH Zurich, Technical University of Munich, and Rutherford Appleton Laboratory. Vendor submissions often include performance tuning credits from teams at NVIDIA Research, Intel Labs, and AMD Research, and results feed into procurement discussions in ministries such as Ministry of Science and Technology (China) and agencies like European Research Council.

Impact and Adoption

HPCG influenced procurement and system design decisions at institutions including Lawrence Livermore National Laboratory, Japan’s RIKEN, and national supercomputing centers in Germany, France, and China. It prompted vendors to optimize memory subsystems and interconnects, motivating architecture changes similar to those seen in successive generations of Cray XK7, IBM Summit, and Fugaku-class systems. The benchmark has been adopted in academic publications in journals such as Journal of Computational Physics, SIAM Journal on Scientific Computing, and presented at conferences like SC Conference, ISC High Performance, and Supercomputing Conference.

Limitations and Criticisms

Critics from academic groups at University of Oxford, Imperial College London, and École Polytechnique Fédérale de Lausanne note that HPCG, while more representative than LINPACK for certain applications, still simplifies real workloads and may not capture domain-specific behaviors from codes developed at NASA, European Space Agency, or commercial firms like Siemens and General Electric. Observers from think tanks such as Center for Strategic and International Studies and commentators in outlets like Nature and Science argue that any single benchmark can bias architecture choices; similar debates occurred during the adoption of SPEC benchmarks and the transition from GFLOPS-centric metrics. Ongoing work by consortia including OpenMP ARB and standardization bodies like ISO seeks to address reproducibility, portability, and fairness in benchmarking practice.

Category:Supercomputer benchmarks