What exactly are the algorithms inside nvGraph? I can’t find that anywhere documented. There are several algorithms for single-source shortest path (Dijkstra, Bellman/Ford, …, variations for dense/sparse, …, with/without negative weights, built-in negative cycle detection, … asf.), and the documentation only says that they are “parallel” algorithms, without specifying what and how they are parallelized. SSSP is not an algorithm, it’s a problem class. Same for widest path.
And there seem to have been no updates ever since it was introduced in CUDA 8. It has been several years, and still no parallel Floyd/Warshall for all-pairs shortest path. Or any of the many others that would make sense to have in a “parallel” (however vague and algorithmically unspecified that is) version. Wouldn’t it make sense for NVidia to implement the most frequently-used graph algorithms as a core CUDA component, so it scales well into all sorts of applications, coming straight from libraries?