NVIDIA released mPDLP (Multi-GPU Primal-Dual hybrid gradient for Linear Programming), a distributed solver within cuOpt that partitions large linear programming problems across multiple GPUs connected via NVLink (Nvidia blog). The solver achieves nearly 10x faster solve times on the widely-studied zib03 benchmark—a 104-million-nonzero problem that has challenged researchers since 2008—compared to single-GPU performance from a year prior (Nvidia blog). The system also delivers up to 6x lower peak memory usage per GPU for problems with up to 2.1 billion nonzeros, addressing the memory-capacity bottleneck that previously forced impractical solve times (Nvidia blog).
For commerce practitioners managing supply chains and energy grids, mPDLP removes a critical constraint: historically, large planning problems could exceed the time window available for operational decisions. By distributing the Sparse Matrix-Vector Multiplication (SpMV) computation across GPUs and using min-cut partitioning to minimize cross-GPU communication, mPDLP enables teams to evaluate larger models and more uncertainty within practical planning horizons (Nvidia blog). Partners including Kinaxis and PSR have already adopted the approach for globally constrained supply planning and large-scale energy-system capacity expansion (Nvidia blog).