HPCG — High Performance Conjugate Gradient
Introduced in 2013 as a more realistic complement to HPL, HPCG uses sparse iterative solver patterns matching real scientific applications. While HPL measures best-case peak compute, HPCG reveals how well a system performs on the memory-bandwidth-bound workloads that dominate real HPC science.
What HPCG Measures
HPCG solves a sparse 3D partial differential equation using a preconditioned conjugate gradient method. This access pattern — irregular, sparse, low arithmetic intensity — mirrors real scientific codes far better than HPL dense matrix operations.
The 3–10% Rule
Systems typically achieve only 3–10% of their HPL performance on HPCG. A system scoring 1 EFLOPS on HPL may score only 60–100 PFLOPS on HPCG. This gap directly reveals the memory subsystem quality and how well the hardware handles sparse, irregular workloads.
Memory Bandwidth Dependence
HPCG is fundamentally memory-bandwidth bound. Systems with high-bandwidth HBM memory (GPU clusters) have a much better HPCG/HPL ratio than CPU-only systems with DDR memory. H100 HBM3 (3.35 TB/s) vs DDR5 (90 GB/s) explains why GPU clusters dominate HPCG rankings.
Procurement Value
Always request HPCG results alongside HPL. A vendor quoting only HPL is hiding their system performance on real workloads. The HPCG/HPL ratio is a reliable predictor of how the system will perform for computational fluid dynamics, molecular dynamics, and finite element codes.
HPCG500 Ranking
A separate HPCG500 ranking is published alongside TOP500. Systems are ranked by GFLOPS achieved on HPCG. The ordering often differs significantly from TOP500 — systems optimized for sparse, irregular workloads rise while HPL-optimized systems fall.
Scientific Relevance
HPCG patterns match: finite element methods, molecular dynamics, computational fluid dynamics, seismic simulation, and quantum chemistry codes. If your HPC workload involves sparse matrices or irregular memory access, HPCG is a far better procurement predictor than HPL.
Running HPCG
HPCG is freely available from hpcg-benchmark.org. Problem size should be large enough to fill available cache (minimum 24 iterations, problem size fitting in DRAM). Runtime is typically 1–2 hours for a valid submission — much faster than a full HPL run.
Combined with HPL
Use HPL for peak performance comparison and TOP500 ranking. Use HPCG for realistic workload prediction. Combine both with STREAM memory bandwidth tests and your actual application benchmarks for a complete procurement picture.