The Graph 500 measures performance against three kernels: search, optimization (single-source shortest path), and edge-oriented. Results on these kernels are informative for application performance in business areas as diverse as cybersecurity, medical informatics, social networks, data enrichment, and symbolic networks such as the human brain.
“Regardless of the domain, the overarching goal of HPC is to overcome global problems by harnessing and leveraging the power of data,” said Murphy. “It is exciting to see the Graph 500 list evolve as we continue to push forward on large-scale data analytics and architectural challenges we face in developing memory and interconnects for these powerful machines.”
Built by Fujitsu and operated by Japanese research institute RIKEN, The K-Computer has dominated the Graph 500 since July 2015.
Developed by a small cadre of well-known supercomputing experts, which includes Georgia Tech School of Computational Science and Engineering Chair David Bader, the Graph 500 list is updated every six months during major supercomputing conferences. It is recognized as a leading indicator of development and investment into high-performance computing (HPC) worldwide, and often reveals trends regarding which technologies are popular in the machines.
The latest list was presented by the Graph 500 executive committee, which includes Richard Murphy, director of Micron’s Advanced Computing Solutions Pathfinding and cofounder of the Graph 500; Peter Kogge, Ted H. McCourtney Professor in the Department of Computer Science and Engineering at Notre Dame; Andrew Lumsdaine, University of Washington and Pacific Northwest National Laboratory distinguished faculty, and Georgia Tech’s Bader.
- K computer – RIKEN Advanced Institute for Computational Science (AICS) (82944 nodes, 663552 cores)
- Sunway TaihuLight – National Supercomputing Center in Wuxi (40768 nodes, 10599680 cores)
- DOE/NNSA/LLNL Sequoia – Lawrence Livermore National Laboratory (98304 nodes, 1572864 cores)
- DOE/SC/Argonne National Laboratory Mira – Argonne National Laboratory (49152 nodes, 786432 cores)
- JUQUEEN – Forschungszentrum Juelich (FZJ) (16384 nodes, 262144 cores)
- ALCF Mira – 8192 partition – DOE/ALCF (8192 nodes, 131072 cores)
- Fermi – CINECA (8192 nodes, 131072 cores)
- Tianhe-2 (MilkyWay-2) – Changsha, China (8192 nodes, 196608 cores)
- ALCF Mira – 4096 partition – DOE/ALCF (4096 nodes, 65536 cores)
- Turing – CNRS/IDRIS-GENCI (4096 nodes, 65536 cores)
Complementing the Graph 500 is the Top 500 list, which is also updated every six months during high-profile HPC conferences. The Top 500 evaluates machines based on how they solve a linear system of equations.
In this new age of big data, we need to measure not just how quickly computers can chew on sets of numbers, but rather how quickly computers can build knowledge from massive-scale data sets,” said Bader. “That’s the difference between Top 500 and Graph 500.”
According to Lumsdaine, “The top machines in the Graph 500 are different than those in the Top 500, so we know we are measuring these machines along different dimensions. More and varied information is always important when evaluating machines at the scale and cost of these top machines.”
The Graph 500 is directed by a steering committee of more than 30 international HPC experts from academia, industry, and national laboratories.