NVIDIA cuOpt and GPU Friendly Optimization Algorithms
Last week, NVIDIA announced a major development for a specific group of people. They are open sourcing their solver “cuOpt”, very soon. This is especially good news because, with such a major example we can build optimization algorithms on GPUs with increased speed and ease.
However, introduction of GPUs is not an instant cure for most of the existing optimization problems. GPUs have their own immense advantages starting with parallelism but they are also more restrictive compared to CPUs. Therefore, we will require new approaches to current algorithms.
Many optimization (or “decision intelligence”) software companies already started to implement GPU utilization to their arsenal. I strongly believe that other firms will also follow suit.
Research will be directed to GPU friendly algorithms and probably some algortihms will be embedded into applications and even mobile devices. As a byproduct, operations research in the future will mostly fall in computer science category than applied mathematics.
Perhaps, not all problems will be solvable by GPU friendly algorithms but as once von Neumann stated “If you have an application that satisfies the axioms, well use it. If it does not, then don’t” (source).
You can already use an implementation of PDLP (cuPDLP-C) with HiGHS in their recent update. For the rest, let’s wait a little and meanwhile you can try their website.
