MERBIT: New GPU method speeds up sparse matrix workloads by 27%

Sparse Matrix-Vector Multiplication (SpMV) is critical for iterative tasks like graph analytics and sparse solvers, but real-world graphs with irregular sparsity patterns make GPU acceleration hard. A new paper from Qi Zhang and colleagues introduces MERBIT, a method designed specifically for repeated SpMV on irregular, graph-like matrices (e.g., PageRank). MERBIT combines two proven ideas: global-level merge-path partitioning to balance work over nonzeros and row boundaries, and local-level encoding of each segment using a compact bit-field descriptor. This dual approach improves workload balance and enables coalesced memory access for both loading the matrix and writing outputs.

On 50 large irregular datasets, MERBIT achieved average speedups of 1.27x (single precision) and 1.25x (double precision) over NVIDIA's cuSPARSE COO format, beating both academic baselines like Ginkgo and other published methods. The paper incorporates three additional optimizations to further boost performance. While MERBIT is currently a research prototype, its consistent gains across diverse real-world graphs suggest it could become a standard building block for high-performance graph analytics and iterative solvers on GPUs.