NVIDIA GTC2025 Highlights: NVIDIA Open Source cuOpt Embarks on a New Era of Decision Optimization

Global enterprises are making critical decisions every minute and second. The logistics company decides on the scheduling plan for freight trucks, retail companies consider the optimal configuration of shelves, and airlines urgently change their routes after the storm. These are not simple route choices, but rather high-risk puzzles involving millions of variables. Once they go wrong, they not only cause economic losses but also lead to customer churn. However, this situation is changing.
NVIDIA announced the open source of its AI decision optimization engine NVIDIA cuOpt, allowing developers to obtain powerful software for free and achieve real-time optimization on a large scale.
Optimizing ecosystem leaders COPT, FICO's Xpress team, HiGHS, IBM, and SimpleRose are integrating or evaluating cuOpt to accelerate decision-making across various industries.
Gurobi Optimization is evaluating and testing cuOpt solvers to improve first-order algorithms and achieve higher levels of performance.
NVIDIA is collaborating with the COIN-OR Foundation to open source cuOpt, widely regarded as the oldest and largest operations research software library.
At the same time, a team of researchers from Arizona State University, Cornell University, Princeton University, Pavia University, and the Zuse Institute in Berlin are exploring its capabilities to develop next-generation solvers that can solve complex optimization problems at lightning fast speeds.
With this technology, airlines can adjust flight plans during flight to prevent chain reaction of delay; The power grid can achieve real-time load balancing and avoid the risk of power outages; Financial institutions can rely on real-time risk analysis to optimize their investment portfolio management.
Optimize and speed up, upgrade decision-making
The most well-known AI applications are all related to prediction, whether it's weather forecasting or generating the next word in a sentence. However, prediction is only half of the challenge. The true value lies in the real-time processing of information.
This is exactly where cuOpt comes in handy.
CuOpt dynamically evaluates billions of variables (inventory levels, factory output, transportation delays, fuel costs, risk factors, and regulations) and provides the best measures in near real-time.
With AI agents and simulation technologies driven by large language models taking on more and more decision-making tasks, the demand for real-time optimization has reached an unprecedented height. With the support of NVIDIA GPU, cuOpt can speed up these computing tasks by several orders of magnitude.
Unlike traditional optimization methods that traverse the solution space sequentially or in finite parallel, cuOpt utilizes GPU acceleration to simultaneously evaluate millions of possibilities, thereby finding the optimal solution for a specific instance at an exponential speed.
CuOpt does not replace existing technology, but rather enhances its implementation. By working in conjunction with traditional solvers, cuOpt can quickly determine high-quality solutions and help CPU based models more efficiently eliminate invalid paths.
Why is optimization so difficult?
How can cuOpt achieve better solutions?
Every decision, including truck scheduling, scheduling work shifts, and balancing grid loads, is a complex optimization problem with exponential potential answers.
From this perspective, taking hospital scheduling as an example: arranging next month's shifts for 100 nurses, the number of possible combinations may even exceed the number of atoms in the observable universe.
Traditional solvers typically use sequential or finite parallel search methods, like using a flashlight to explore a huge maze, where only one path can be explored at a time. And cuOpt has changed this mode. It can intelligently evaluate millions of possibilities and achieve exponential improvement in optimization speed.
In the past, tasks such as personnel scheduling, logistics path planning, and supply chain optimization often required hours or even days of computation time.
NVIDIA cuOpt has changed this situation, with specific performance data as follows:
Linear programming acceleration: In large-scale benchmark tests, the average solving speed is 70 times faster than CPU based PDLP solvers, with acceleration ratios ranging from 10 times to 3000 times.
Mixed integer programming (MIP): In the case of SimpleRose, the speed of solving mixed integer programming is increased by 60 times.
Vehicle route planning: As shown in Lyric's case, the speed of dynamic route planning has increased by 240 times, empowering service cost optimization and near real-time route adjustment.
In the past, it took hours or days to make a decision, but now it only takes a few seconds.
Optimize and create a better world
Better optimization can not only improve the operational efficiency of enterprises, but also make the world more sustainable, resilient, and fair.
Intelligent decision-making can reduce resource waste. The power grid can optimize power allocation, reduce power outages, and achieve efficient grid connection of renewable energy sources such as wind and solar energy. The supply chain achieves dual optimization of cost and emissions while minimizing excess inventory through dynamic adjustment.
In underserved areas, hospitals can allocate beds, doctors, and medication in real-time to accelerate the process of patients receiving life-saving treatment. Humanitarian aid organizations responding to disasters can quickly re plan and determine the best way to distribute food, water, and medicine, reducing delays at critical moments. The public transportation system can dynamically adjust according to real-time demand, easing congestion and shortening travel time for millions of passengers.
CuOpt not only relies on the powerful hardware, but also on its intelligent search strategy. It does not exhaust all possible solutions, but instead explores a vast search space intelligently, focusing on key constraints to find the optimal solution more quickly. By using GPU acceleration, cuOpt can evaluate multiple solutions in parallel, achieving efficient and real-time optimization processing.
Industry Support - The New Era of Decision Intelligence
Leading companies in optimization fields such as FICO, Gurobi Optimization, IBM, and SimpleRose are exploring the advantages of GPU acceleration or evaluating the possibility of integrating cuOpt into their workflows, as well as its potential in industrial planning, supply chain management, and scheduling.
Wise decisions, powerful systems, outstanding results
CuOpt redefines large-scale optimization.
As mentioned above, for enterprises, this means that AI optimization can reset schedules, plan fleet routes, and reallocate resources in real-time, thereby reducing costs and improving flexibility.
For developers, it provides a high-performance AI toolkit that solves decision problems 3000 times faster than traditional CPU solvers when dealing with complex optimization challenges, such as network data routing (optimizing video, voice, and network traffic to reduce congestion and improve efficiency) or power distribution (balancing supply and demand of the power grid, minimizing losses, and ensuring stable transmission).
For researchers, this is an open experimental field that takes AI decision-making to new heights.
CuOpt will be released in open source later this year for free use by developers, researchers, and businesses.
The application of cuOpt
For enterprise production deployment, cuOpt provides support as part of the NVIDIA AI Enterprise software platform and can also be deployed as an NVIDIA NIM microservice, making it easy to integrate, scale, and deploy across cloud, on premises, and edge environments.
With its open-source version, developers will be able to easily access, modify, and integrate cuOpt source code into their own solutions.