In the world of optimization, speed is everything. Whether optimizing supply chains, scheduling transportation, or tackling complex energy market decisions, the ability to solve large-scale mathematical programming problems quickly can mean the difference between an efficient operation and costly delays.
Image courtesy of NVIDIA.
At AMPL, we’ve long been committed to empowering optimization professionals with the best tools available. That’s why we’ve been collaborating with NVIDIA to integrate and test NVIDIA cuOpt—NVIDIA cutting-edge, GPU-accelerated solver—into AMPL. And the results speaks for themselves:
In early benchmarking, cuOpt has delivered up to 50x speed improvements over traditional CPU-based solvers on real-world problem instances related to electricity grid simulations.
But this is just the beginning. With NVIDIA’s recent announcement that cuOpt will be open-source, we’re excited about what this means for the entire optimization community. Making this powerful technology widely accessible is a game-changer—not just for AMPL and NVIDIA, but for the researchers, businesses, and developers looking to push the boundaries of mathematical optimization.
One of the most frequent questions we get from customers is:
“Can I speed up my optimization models by running them on GPUs?”
Until now, the answer has always been no—most traditional solvers are built for CPUs, and simply adding a GPU wouldn’t provide meaningful performance gains. But with cuOpt, that’s finally changing.
For the first time, we’re seeing substantial performance improvements in particular for LP by leveraging GPU acceleration, ushering in a new era where high-performance optimization is no longer constrained by CPU limitations.
In this post we’ll dive deeper into how AMPL is integrating cuOpt, real-world use cases demonstrating these gains, and what this shift means for the future of mathematical optimization.
For decades, optimization has been constrained by CPU-based solvers, relying on multi-core processing to tackle increasingly complex problems. While this approach has been effective, it has also hit performance limits—particularlly in large-scale, real-time decision-making scenarios.
Enter the NVIDIA cuOpt, as GPU-accelerated solver designed to unlock massive parallel processing power to Linear Programming (LP) and Mixed_Integer Programming (MIP) problems. By leveraging GPUs, cuOpt offers a fundamental shift in how optimization problems are solved—breaking free from traditional CPU bottlenecks.
At AMPL, we’ve been putting cuOpt to the test across various real-world optimization challenges, including energy market simulations, where Locational Marginal Pricing (LMP) calculations must be solved thousands of times in rapid succession. With traditional CPU-based solvers, these calculations can take minutes per run,making it nearly impossible to process thousands of demand scenarios within a practical timeframe. The only workaround has been to distribute workloads across large computing clusters, an approach that comes with high infrastructure costs and added complexity.
cuOpt changes the equation. By running these models on GPUs, we’ve seen large-scale problems solved in under two seconds with a 50X speedup—compared to minutes using traditional methods. This level of speedup is a game-changer for industries where real-time decision-making is critical.
While convergence challenges remain in some cases, early results suggest that GPU-accelerated optimization could soon deliver both speed and reliability at an unprecedented scale—unlocking new possibilities for businesses and researchers alike.
With cuOpt now becoming open-source, we believe this is just the beginning of a new frontier in optimization.
Our collaboration with NVIDIA has gone beyond just benchmarking—we’ve worked together to make cuOpt truly accessible within AMPL. Our development teams have seamlessly integrated cuOpt into AMPL’s modeling environment, allowing users to harness GPU acceleration without modifying their existing optimization models.
As with any emerging technology, there are challenges. While cuOpt has delivered exceptional performance gains, we’ve also identified areas for refinement, particularly in convergence behavior for certain problem instances. That’s why NVIDIA’s decision to open-source cuOpt is so impactful — by making this technology widely available, NVIDIA is inviting the global optimization community to contribute, refine, and push the boundaries of what’s possible.
We are excited to see how this collaborative effort will accelerate innovation and unlock new frontiers in mathematical optimization.
One of the first real-world tests of cuOpt in AMPL focused on electricity price forecasting, a critical task for market operators and participants, including power traders. These stakeholders must solve thousands of Locational Marginal Pricing (LPM) calculations to predict grid pricing accurately. Traditionally, these problems—often formulated as large-scale Linear Programs (LPs)—require CPU-based solvers, and running thousands of demand scenarios can take hours or even days.
While in reality, these are Mixed-Integer Programming (MIP) problems, LP relaxations are commonly used as approximations to reduce solve times and make large-scale forecasting computationally feasible.
This challenge represents one of the most significant applications of AMPL. The ability to rapidly build large-scale, highly complex models—such as simulating the electric grid for an entire region like Europe—is essential. AMPL is up to 100 times faster than several other modeling tools in generating such optimization models, drastically reducing the setup and processing time for each scenario. This efficiency makes AMPL a preferred choice for real-time decision-making in energy production.
In electricity price forecasting, AMPL runs ahead of time, executing a model similar to the one that will later determine which power plants to dispatch and, ultimately, the energy price at any given moment. This capability is critical for managing entire energy grids, as seen in real-world applications with the New York Independent System Operator (NYISO) and the Electric Reliability Council of Texas (ERCOT), where AMPL plays a key role in optimizing grid operations and ensuring a stable electricity supply.
To evaluate cuOpt capabilities, we conducted a benchmarking study using a real-world European electric grid optimization model. Despite extensive tuning of CPU-based solvers, achieving runtimes below 90 seconds remained a challenge—even when using the barrier method without crossover, which guarantees polynomial runtime.
With cuOpt GPU acceleration, some instances were solved in under two seconds—an incredible 50x speedup in initial tests.
These results suggest that the computational bottlenecks of existing LP methods have been significantly reduced. When combined with AMPL’s industry-leading speed in building optimization models, this breakthrough paves the way for real-time electricity market simulations, making high-speed forecasting and decision-making far more practical.
While challenges remain in convergence stability, these results highlight cuOpt potential to make real-time market simulations a reality. As GPU solvers continue to evolve, industries that depend on fast, large-scale mathematical programming—like energy, logistics, and finance—stand to gain substantial benefits, unlocking a new era of high-performance optimization.
This marks the beginning of a broader initiative to unlock the full potential of GPU acceleration in mathematical optimization. By combining AMPL’s powerful modeling capabilities—including its exceptionally fast model build times—with NVIDIA expertise in high-performance computing, we’re shaping the next generation of solvers capable of handling larger, more complex problems in real-time.
For industries like energy, logistics, finance, and manufacturing, this evolution means that optimizations once limited by CPU constraints will soon be able to run faster, more frequently, and at a much greater scale. With cuOpt now open-source, we look forward to fostering innovation, driving new advancements, and expanding the boundaries of what’s possible in large-scale optimization.
The open-source release of cuOpt marks a major milestone in making GPU-based optimization more practical and accessible. By integrating cuOpt into AMPL, we are removing adoption barriers and enabling users to explore GPU acceleration without the need for significant workflow changes.
While GPUs have long been leveraged for machine learning and simulations, their use in mathematical optimization has been more limited—until now. This integration bridges that gap, proving a powerful new tool for handling large-scale problems more efficiently.
For AMPL users, this means an opportunity to experiment with GPU-accelerated optimization in scenarios where speed is a limiting factor. While CPU-based solvers remain essential for many tasks, cuOpt introduces a transformative approach that offers significant performance improvements in applications such as large-scale simulations and real-time decision-making.
As both the solver and GPU hardware continue to advance, we anticipate GPU acceleration will become an increasingly vital complement of high-performance mathematical optimization.
We’re just getting started, and we’d love for you to be part of this journey. If you’re as excited about cuOpt as we are, here’s how you can stay connected and get hands-on:
Whether you’re a researcher, developer, or industry leader, this is a game-changing moment for optimization. Let’s explore the possibilities together!
Filipe Brandão
Head of Development