NVIDIA cuOpt Solver Cracks Four Previously Unsolved Optimization Problems

NVIDIA's cuOpt optimization engine has found solutions for four previously unsolved problems in the MIPLIB benchmark set, according to a technical paper published by the company's research team. The GPU-accelerated solver achieved a 0.22 primal gap score—roughly 67% better than traditional methods—while finding more feasible solutions than leading open-source CPU alternatives.

The breakthrough matters for industries running complex logistics, scheduling, and financial optimization at scale. Mixed integer programming problems underpin everything from airline crew scheduling to supply chain routing, and faster solutions translate directly to operational cost savings.

What Changed Under the Hood

The cuOpt team rewrote the feasibility pump algorithm—a decades-old approach to finding workable solutions—to exploit GPU parallelism. Two key modifications drove the gains.

First, they swapped out the traditional simplex algorithm for PDLP (Primal-Dual hybrid gradient), discovering that lower precision projections still produced quality results. This allowed the solver to iterate faster on larger problem sets. Second, they rebuilt the domain propagation algorithm for GPU architecture, adding bulk rounding and dynamic variable ranking.

The results speak for themselves. Across benchmark tests, GPU Extended FP with Fix and Propagate found 220.67 feasible solutions on average versus 188.67 for standard Local-MIP—a 17% improvement. More importantly, the objective gap dropped to 0.22 compared to 0.46 for the baseline approach.

Enterprise Integration Play

NVIDIA positioned cuOpt within its broader enterprise AI stack. The company specifically mentioned integration with Palantir Ontology and NVIDIA Nemotron reasoning agents, suggesting a push toward continuous optimization pipelines rather than one-off problem solving.

This fits the pattern. cuOpt already handles vehicle routing and linear programming problems, with documented performance claims of up to 3,000x speedups over CPU solvers for certain workloads. The open-source release through the COIN-OR Foundation lowers adoption barriers for enterprises already running NVIDIA hardware.

Hardware Requirements and Availability

cuOpt requires A100 Tensor Core GPUs or newer, limiting deployment to organizations with recent NVIDIA infrastructure. The solver is available now on GitHub with example notebooks covering emergency management and logistics use cases.

For companies already invested in NVIDIA's ecosystem, the MIP heuristics add another reason to consolidate optimization workloads on GPU infrastructure. The four newly-solved MIPLIB problems—liu.mps, neos-3355120-tarago.mps, polygonpack4-7.mps, and bts4-cta.mps—serve as proof points for enterprises evaluating whether GPU-accelerated optimization delivers on its promises.