Hacker News Reader

Typical algorithms solving combinatorial problems are usually heavily branching, so not so much vectorizable. For example for a branch-and-cut MIP solver, you could run the LP-relaxed problems on GPU, but most of the logic lives in branching and cutting heuristics that are much harder to parallelize.

A lot of OR stuff can be parallelized, but it isn't a panacea. Those techniques don't necessarily get you the best answer sometimes. My company does some massive optimization problems and although we have lots of cores for all the simulations, any one is pretty much single core + a lot of RAM. It doesn't work well on GPU like some heuristic optimization problems can.

Edit: Someone smarter than me says it better below :)

For starters, CP and SAT are based on integers and boolean variables. Secondly, local search and systematic search are just no way near parallelizable to the degree required to see gains running GPU/TPU.