cuOpt

[Stuart Rogers]

cuOpt executes metaheuristics on a GPU in order to accelerate the solution of MILPs. It also has an implementation of PDLP, for solivng LPs, that runs on a GPU. These may be useful for OR-Tools. For example, the GPU metaheuristics might be useful for generating feasible solutions for CP-SAT.

https://github.com/NVIDIA/cuopt

This GTC 2025 talk explains how cuOpt solves LPs (via PDLP) and MILPs (via heuristics, PDLP, and simple branch and cut): https://www.nvidia.com/en-us/on-demand/session/gtc25-s72290/

This GTC 2025 talk discusses how SimpleRose (https://simplerose.com/) combined their Rose solver with cuOpt (to generate upper bounds via heuristics) to solve MILPs 60x faster: https://www.nvidia.com/en-us/on-demand/session/gtc25-s72603/

[Stuart Rogers]

cuOpt requires an Nvidia GPU. Here are examples using cuOpt: https://github.com/NVIDIA/cuopt-examples

[Laurent Perron]

Cuda PDLP is interesting. 

Cuda routing is very limited in what it can solve, mostly pure academic problems where it crunches moves in parallel.

I do not know about cu-mip yet. I do not know how it performs against feasibility jump and violation ls. 

The Nvidia limitation is a serious one. For instance, none of my workstations, laptop sport a Nvidia GPU. 

Laurent Perron | Operations Research | lpe...@google.com | (33) 1 42 68 53 00

--
You received this message because you are subscribed to the Google Groups "or-tools-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to or-tools-discu...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/or-tools-discuss/0beadfa3-3c32-4e56-8076-61f7c734dd71n%40googlegroups.com.