This question is related to the other Graph500 GTEPs question.

Thinking about the growth of computing power requires operationalizing "compute". Our question based on the well-known OpenAI paper pointing to a 3.5-month doubling time in compute uses Floating-point Operations (FLOPs) as a measure. However, some have argued that for certain purposes another measure -- TEPs -- makes more sense. In particular, the brain seems more bottlenecked by its ability to transmit information over long distances rather than its ability to perform basic arithmetic operations. This is important if treat the computational capacity of the brain as a key input into our forecasts.

Traversed Edges Per Second (TEPs) is a measure of the computational power and communication ability of a computer. An edge is the connection between two nodes in a graph, and to traverse the edge is to communicate between those two nodes.

A Breadth-First Search (BFS) is a well known graph search technique that starts with a single source node, then, in phases, finds and labels its neighbors, then the neighbors of its neighbors, etc.

The BFS benchmark runs a Breadth-First Search through a graph over 1 petabyte in size.

As of Nov 2018 the current top performing machine produced 38621.4 GTEPs.

This question will be resolved using the Graph500 June 2019 report, based on the BFS benchmark.

Data

Nov 2018: http://graph500.org/?page_id=433 Specification of the benchmark: http://graph500.org/?page_id=12#sec-9_2