cuPDLP+: A Further Enhanced GPU-Based First-Order Solver for Linear Programming (2507.14051v1)

Abstract: We introduce cuPDLP+, a further enhanced GPU-based first-order solver for linear programming. Building on the predecessor cuPDLP, cuPDLP+ incorporates recent algorithmic advances, including the restarted Halpern PDHG method with reflection, a novel restart criterion, and a PID-controlled primal weight update. These innovations are carefully tailored for GPU architectures and deliver substantial empirical gains. On a comprehensive benchmark of MIPLIB LP relaxations, cuPDLP+ achieves 2x - 4x speedup over cuPDLP, with particularly strong improvements in high-accuracy and presolve-enabled settings.