Graph500
Graph500 ranks supercomputers on data-intensive graph traversal, measuring traversed edges per second (TEPS) rather than FLOPs. It exposes memory and interconnect performance for irregular analytics workloads.
Graph500 is a benchmark and ranking that measures supercomputer performance on data-intensive graph analytics rather than floating-point math. It was created because many modern workloads — social networks, cybersecurity, bioinformatics, knowledge graphs — are dominated by irregular memory access and communication, which traditional FLOP-based benchmarks like LINPACK do not capture.
The core workload generates a large synthetic graph using a Kronecker (RMAT) generator that produces realistic power-law degree distributions, then runs graph kernels over it. The primary kernel is breadth-first search (BFS); a single-source shortest-path (SSSP) kernel is also defined. These operations have poor locality and heavy fine-grained communication, stressing the memory and network subsystems.
What It Measures
The headline metric is TEPS — traversed edges per second — typically reported as GTEPS or TTEPS at scale. It captures how fast a system can walk graph edges, which depends on memory bandwidth, latency, and interconnect rather than arithmetic throughput. The benchmark also defines problem sizes by a scale factor (the log2 of the vertex count), so larger scales test bigger, more demanding graphs.
Methodology
A system constructs the graph, then runs BFS from many randomly selected source vertices, timing each search and verifying correctness. The reported TEPS is derived from the median search time across these roots, divided into the number of edges traversed. Submitters report the graph scale, TEPS, and full configuration. A Green Graph500 companion list ranks energy efficiency in TEPS per watt.
How to Interpret Results
Higher TEPS is better, but always compare at the same scale factor, since larger graphs are harder per edge. A machine that ranks high on TOP500 but low on Graph500 has FLOPs its memory and network cannot feed for irregular workloads. Use Graph500 to predict performance of analytics, graph databases, and sparse computations. The Green Graph500 list matters where power constrains data-center scaling.
Limitations
The synthetic RMAT graph, while power-law, differs from many real graphs in clustering and structure, so results are indicative rather than exact predictors. The benchmark focuses on BFS and SSSP, not the full range of graph algorithms like PageRank or community detection. Like other rankings, it can be over-tuned, and it does not capture graph storage, updates, or end-to-end analytics pipelines.