Introducing Gurobi’s First GPU-Accelerated Solver: Test an Alpha Version of Gurobi’s PDHG Implementation on NVIDIA’s GPU Hardware

By Dr. Cara Touretzky, Senior Technical Account Manager, and Dr. Robert Luce, Principal Developer 

We’re excited to announce the first publicly available version of the Gurobi Optimizer built specifically for use on NVIDIA’s GPU hardware.   

Gurobi is on the leading edge when it comes to researching new GPU-enabled algorithms, and our customers are always pushing the bounds of what’s possible with optimization. That’s why we’re releasing this alpha version of the Gurobi Optimizer, which allows Gurobi customers to test their large-scale LPs using the PDHG algorithm on GPUs and evaluate potential performance benefits. 

What to Know About Gurobi’s PDHG Implementation 

It’s important to note that this algorithm is still in the alpha stage, meaning it is intended for experimentation rather than production use, as it is still in the early stages of technical readiness. Performance tuning and algorithmic developments are still ongoing.  

This alpha version is based on the full Gurobi 12.0.3 solver, with the addition of the PDHG algorithm for LPs, which will run on GPU hardware.  We recommend testing with NVIDIA H100 or newer (NVIDIA drivers must be installed, but a full CUDA installation is not required). Currently, only armlinux64 and linux64 are supported. For more instructions on how to get started with the alpha version, please see the documentation here.  

It is also important to reiterate some of our early observations about the impact of GPUs in the optimization space: With these new options for algorithm and hardware selection, it is more important than ever to test your model and perform benchmarking to see which combination is the best fit.  

Our initial findings are that the PDHG algorithm on GPU will offer the most benefit to large-scale, “giant” LPs, beyond what we may have considered possible just a few years ago. Of course, we’re also exploring models that may not be “giant,” but have a level of complexity and heavy compute needs that will warrant using the new algorithms and GPU hardware.  

Start Testing Your Models 

The alpha version is now available to current Gurobi customers with commercial or academic licenses via the Download Center. If you would like to try Gurobi’s GPU-accelerated PDHG implementation but are not a current user (or if you are a current user but need additional capacity), you can receive a free evaluation license by filling out the form here.  

If you need assistance conducting your benchmarking, or if you would like us to help review your benchmark results or share your success stories, please let us know by opening a ticket with our Experts team. 

The Gurobi Team would like to thank NVIDIA for their continued support and Haihao Lu (MIT) for inspiring exchanges on PDLP.