Hi Marcus,
I'm not sure I'm the Paul to which this was addressed, but I'm happy to chip in (pun deliberate). No, I'm not aware of CPLEX-specific hardware benchmarks. To the best of my knowledge, CPLEX does not exploit GPUs, although a fairly recent announcement suggests that may be coming someday (likely not soon). If I were shopping for a computer with CPLEX performance on large MIPs in mind, I would look at the usual three suspects: CPU speed; number of CPU cores; and RAM. I think CPU speed would be top of mind. Between RAM and number of cores, it's a bit less clear. I think parallelization yields decreasing returns as you crank up the core count. At the same time, I would assume that at some point extra RAM would stop paying for itself. Also, you can somewhat compensate for RAM by getting a fast, decent size SSD internal disk (where CPLEX would swap out chunks of the tree).
There's also an interaction term in the "model". To exploit more cores, you may need more RAM, because parallel threads get copies of chunks of the tree. So RAM usage goes up with thread count. Unfortunately, I have no idea of what an ideal ratio of RAM to cores would be.
Paul
------------------------------
Paul Rubin
Professor Emeritus
Michigan State University
------------------------------