Citation¶
If you use cuPDLPx in published work, please cite the computational paper. If your work also relies on the restarted Halpern PDHG method or its convergence theory, please cite the theoretical paper as well.
Computational paper¶
Algorithmic enhancements, GPU implementation, and numerical results are described in this paper:
@article{lu2025cupdlpx,
title = {{cuPDLPx}: A Further Enhanced GPU-Based First-Order Solver
for Linear Programming},
author = {Lu, Haihao and Peng, Zedong and Yang, Jinwen},
journal = {arXiv preprint arXiv:2507.14051},
year = {2025},
url = {https://arxiv.org/abs/2507.14051}
}
Theoretical paper¶
Restarted Halpern PDHG and its reflected variant are developed in this paper:
@article{lu2024restarted,
title = {Restarted Halpern PDHG for Linear Programming},
author = {Lu, Haihao and Yang, Jinwen},
journal = {arXiv preprint arXiv:2407.16144},
year = {2024},
url = {https://arxiv.org/abs/2407.16144}
}
Background¶
Papers¶
- David Applegate et al., Practical Large-Scale Linear Programming Using Primal-Dual Hybrid Gradient, NeurIPS 2021.
- David Applegate et al., PDLP: A Practical First-Order Method for Large-Scale Linear Programming, 2025.
- Haihao Lu and Jinwen Yang, cuPDLP.jl: A GPU Implementation of Restarted Primal-Dual Hybrid Gradient for Linear Programming in Julia, 2023.
- Haihao Lu et al., cuPDLP-C: A Strengthened Implementation of cuPDLP for Linear Programming by C language, 2023.
- Haihao Lu and Jinwen Yang, An Overview of GPU-based First-Order Methods for Linear Programming and Extensions, 2025.
- Daniel Cederberg and Stephen Boyd, Presolving for GPU-Accelerated First-Order LP Solvers, 2026.
Blogs¶
- Cara Touretzky, Robert Luce, and David Torres Sanchez, Using GPUs to Solve LPs: What's in It for Me?, Gurobi blog, 2025.
- Cara Touretzky and Robert Luce, Introducing Gurobi's First GPU-Accelerated Solver: Test a Beta Version of Gurobi's PDHG Implementation on NVIDIA's GPU Hardware, Gurobi blog, 2025.
- Imre Pólik, GPU Acceleration of the Hybrid Gradient Algorithm in FICO Xpress, FICO blog, 2025.
- Nicolas Blin, Accelerate Large Linear Programming Problems with NVIDIA cuOpt, NVIDIA Technical Blog, 2024.
- Artelys, Artelys Knitro 15.0: New Tools for Your Large-Scale Models, Artelys news, 2025.
- Scaling up linear programming with PDLP, a Google Research article by Haihao Lu and David Applegate.
- Mathematical background for PDLP, the OR-Tools reference for PDLP formulations, residuals, rescaling, and infeasibility certificates.
Related projects¶
Open-source solvers¶
- cuPDLP.jl, the earlier Julia GPU solver.
- cuPDLP-C, the C implementation of cuPDLP.
- PDQP.jl, a Julia first-order solver for convex quadratic programming on CPUs and NVIDIA GPUs.
- PDHCG, a GPU-accelerated first-order solver for convex quadratic and conic quadratic programming.
- HPR-LP-C, a C GPU solver for LP based on the Halpern Peaceman–Rachford method.
- cuOpt, NVIDIA's open-source GPU-accelerated optimization library for LP, MIP, and vehicle routing.
- HiGHS, an open-source solver for LP, MIP, and QP that includes a PDLP-based first-order LP solver.
- D-PDLP, a distributed LP solver built on cuPDLPx for execution across multiple GPUs.
- CoolPDLP.jl, a Julia implementation of PDLP and its variants with support for CPUs, multiple GPU architectures, and batched solves.
Commercial solvers (implementing PDLP)¶
Benchmarks¶
- MIPLIB 2017, a source of LP relaxation benchmark instances.
- Mittelmann benchmarks, independent benchmarks and test sets for optimization software.
Acknowledgements¶
The development of cuPDLPx is partially supported by AFOSR Grant No. FA9550-24-1-0051, ONR Grant No. N000142412735, and the NVIDIA Academic Grant Program.