ExaModels is currently one of the fastest optimization modelers, and can evaluate (sparse) derivatives in parallel on various GPUs

Nice initiative! Just a naive question regarding doi.org/10.1137/23M1.... You don't have to assume the strong Legendre-Clebsch condition to establish the convergence of your algorithm? What's the behavior of the interior-point method when the problem has singular arcs?

Nw! We are using double precision, we need that to get an accurate convergence in IPM. It turns out the modern GPUs are much performant with double precision than the previous generations.

In our experience we can achieve 1e-8 on medium instances, 1e-6 on large-scale ones. The condensed strategy induces an additional ill-conditioning that renders the method slightly less accurate than HSL solvers.

Deadline is November 29th. The link to submit a contribution is: roadef2025.org/soumettre-un...

For CPU versus GPU, our main message is that you don't have to use an expensive GPU: I can get reasonable speed-up on the GPU I have on my laptop.

Good question. We did some benchmarks a while ago, we found out it's non-trivial to estimate the energy usage on the CPU. As such we just made a benchmark comparing our new method (HyKKT) versus our previous one (Biegler). Solving one OPF problem amounts to at most 500J using the new method.