4

I consider buying a new desktop computer and I work on academic MIP optimization problems with Gurobi.

I know that GPUs cannot be leveraged by MIP solvers, so my question rather concerns the choice of the CPU (and possible related configuration matters).

Do you have any recommendation for a CPU fitting in a standard desktop computer (=available at computer retail stores) which purpose is primarily to solve MIP problems?

mondano
  • 143
  • 5
  • I'm voting to reopen this question as I believe that it has now been refined in a manner that can be answered on Hardware Recommendations. – AStopher May 27 '16 at 09:13
  • Can you define what a standard computer is? What kind of specs are we talking about? You can [edit] those details into the question. – ArtOfCode Jun 03 '16 at 20:24

1 Answers1

2

I have to admit, I had to educate myself about Gurobi.

It looks like it is a software suite for linear optimization problems. It also seems, it is a regular one, i.e. with "simple" C++ and R sequential implementations.

For that, any CPU in your budget will do. You should just try - within your budget limits - get the best clock/cache/cores combination. An AVX512 capable CPU seems to make sense.

But let me offer an extended perspective on this.

If performance is your main concern, you should have a look at GPU computing. As for GPU accelerated solvers, see e.g. https://developer.nvidia.com/amgx, which gives promising speedups of a factor of 10 within your reach.

Maybe, spend only reasonable money on an average CPU. If you get the Gurobi vendor to support AmgX, you can buy also a reasonable CUDA-capable GPU and still have speedup by a factor of 10, which would not be possible by any economically feasible CPU.

See also e.g. http://vratis.com/blog/ OpenFOAM and other software does actually already support AmgX, so maybe it would not take too much effort to convince Gurobi to think about it.

Perlator
  • 319
  • 1
  • 7
  • According to Gurobi, GPUs are as per Gurobi 6.0 (a bit dated) not particularly well suited to solvers like Gurobi ad similar: "Currently, GPUs are not well-suited for solvers like Gurobi ◦ For LP, sparse linear algebra does not parallelize to hundreds of GPUs ◦ For MIP, each tree node requires very different calculations, but GPU SIMD computations are designed for identical calculations on different data , please see: http://www.gurobi.com/pdfs/webinar-parallel-and-distributed-optimization-english.pdf – rakke Jan 29 '18 at 08:35