HiGHS and NVIDIA cuOpt: Driving open-source innovation in optimization
For the past six years, HiGHS, the world’s leading open-source linear optimization software, has been developed at the University of Edinburgh under the leadership of Julian Hall and Ivet Galabova. Built with contributions from graduate students, HiGHS has consistently delivered state-of-the-art performance, as confirmed by industry-standard independent benchmarks.
With NVIDIA cuOpt becoming open source, a new era of accelerated optimization is emerging—one that combines powerful algorithms with the computational strength of GPUs. This move presents a game-changing opportunity for researchers, developers, and businesses seeking high-performance, cost-effective solutions.
Unlocking the Power of GPUs for Linear Optimization
Historically, constrained and discrete optimization has been CPU-centric, with GPUs playing a limited role due to the nature of traditional solvers. However, with the rise of first-order methods for linear programming and ubiquity of GPU computing, the landscape is shifting.
HiGHS has already introduced cuPDLP-C, a first-order linear programming solver, and in our next release, this solver will run natively on NVIDIA GPUs. Looking forward, HiGHS aims to enhance its first-order solver algorithmically, extend GPU acceleration to quadratic programming, and explore use of GPUs in interior point methods, particularly for Cholesky decomposition.
Through collaboration with NVIDIA, HiGHS is excited to push the boundaries of what’s possible in GPU-accelerated optimization.
cuOpt Open-Source: A Milestone for the Optimization Community
NVIDIA’s decision to open-source cuOpt is a significant event for the optimization ecosystem. By making cuOpt freely available, NVIDIA is democratizing access to cutting-edge solvers, removing barriers for stakeholders who may find commercial software cost-prohibitive or inconvenient.
For HiGHS and other solver developers, open access to cuOpt source code will provide an invaluable opportunity to:
- Incorporate NVIDIA cuOpt into their frameworks
- Leverage technical innovations for enhanced performance
- Foster collaboration to drive further enhancements in optimization
This initiative marks a significant step forward in the evolution of open-source solvers, enabling researchers and industry professionals to accelerate their work.
HiGHS and cuOpt: A Powerful Collaboration
The synergy between HiGHS and NVIDIA cuOpt is already yielding promising results. The cuOpt heuristic mixed-integer programming (MIP) solver, when paired with HiGHS, enhances performance beyond what either solver can achieve alone:
- Benchmark tests indicate that combining HiGHS with cuOpt enhances the efficiency of solving MIP problems. For instance, when running MIPLIB benchmarks with a 5-minute time limit, HiGHS alone achieves results with a 28% gap from optimality. However, when integrated with cuOpt running on a H100 GPU, the gap from optimality improves to 21%.
- HiGHS has integrated a feature that enables cuOpt to accelerate the HiGHS MIP solver.
- This feedback loop between HiGHS and cuOpt is driving continuous performance improvements.
By leveraging each other’s strengths, the two solvers are pushing the limits of open-source optimization—creating faster, more scalable, and more efficient solution technology.
The Future of Open-Source Optimization
The HiGHS team is excited about the future of GPU-accelerated linear optimization and is committed to collaborating with NVIDIA to advance solver performance for the global community.
With open-source NVIDIA cuOpt and the ongoing developments in HiGHS, the future of optimization is faster, more accessible, and more powerful than ever.
Together, we are unlocking new possibilities for large-scale optimization—bringing high-performance solutions to researchers, developers, and enterprises worldwide.
