GraphBLAS (en.wikipedia.org)

GraphBLAS, the open API that frames graph algorithms as linear-algebra primitives on sparse matrices, reached version 2.1.0 (released Dec 22, 2023). The specification—formalized in a C API but implemented across C++, CUDA, Python, MATLAB and Julia bindings—standardizes opaque sparse matrix/vector objects, semiring-based operations (e.g., max-plus, min-plus, Galois field), masked assignments and a non-blocking execution mode for lazy/asynchronous evaluation. Born from the BLAS tradition and under development since 2013, GraphBLAS provides a portable, high-performance foundation for graph analytics and has proven useful beyond classical graph problems, attracting interest from machine learning and bioinformatics communities. Technically, GraphBLAS maps graphs to adjacency or incidence matrices and implements graph primitives (BFS, traversals, transforms) via sparse matrix ops such as matrix-vector and matrix-matrix multiply over custom semirings. Features like masked assignment and descriptor flags enable succinct, efficient expressions (e.g., BFS via GrB_vxm with masks), while semirings let you encode boolean, tropical, or Galois-field computations uniformly. The standard’s reference implementations (including SuiteSparse:GraphBLAS) and bindings power high-performance systems such as RedisGraph and have shown large speedups over dense BLAS for sparse neural-network forward propagation, making GraphBLAS a key enabler for scalable, sparse-first AI/ML and graph database workloads.