GraphBLAS

GraphBLAS is a collection of algorithms based on graph operations employing linear algebra over semirings. The sparse matrix and vector operations defined in GraphBLAS are beneficial for developing various graph algorithms. This library provides access to built-in types and operators, which allows users to create new types and operators for specific applications without needing for a recompilation of the GraphBLAS library.

In his talk, Tim Davis from Texas A&M University introduces semirings as a novel way to redefine matrix multiplication, moving beyond conventional operations. This talk dives into performance improvements achieved by using GraphBLAS, showcasing significant speedups compared to traditional methods, such as a BreadthFirst Search (BFS) in contrast to a BFS in GraphBLAS. With applications in a wide variety of areas, such as drug interaction analysis, understanding and detecting fraud, GraphBLAS has had an impact on improving numerical computations of sparse matrices. In this talk, Tim discusses various methods and operators within GraphBLAS, its interfaces with MATLAB^®, Octave, and Python^®, and the impact of performance improvements across different benchmark operations. He also covers the use of GraphBLAS in deep neural networks and future plans to incorporate kernel fusion and enhance performance on CUDA-enabled GPUs in collaboration with NVIDIA^®.