Computer configuration for solving large scale MIPs

Here are some recommendations made by Tobias Achterberg on IBM's CPLEX forum on choosing computer hardware for solving large scale MIPs, which is found in AMPL's mailinglist:

Even though the best MIP codes are parallel, they have some critical sequential parts.  Thus single-core computation speed is important, and there is likely to be little useful speedup past 16 or 32 cores.

Large amounts of memory are essential, especially when running in parallel with many cores.  It is worth getting a machine that allows for the largest amount of memory that can be obtained at reasonable cost.

Often the bottleneck is memory bandwidth, so also you want to have the fastest memory you can get.  Note that for current processor chips your computer needs to be equipped with the correct number of RAM DIMMS to get full speed.  This number varies by processor, and must be obtained from information provided by the computer or processor manufacturer.

A reasonable estimate of MIP performance on different chips is given by the SPEC benchmarks at http://www.spec.org/cgi-bin/osgresults?conf=cpu2006.  Do a "Simple Request" for a processor of your choice, look into the "SPECfp2006" tables, and check out the "Base Run Time" of the "450.soplex" test.  The lower the number, the better.

As of March 2011, Tobias's recommended configuration was a dual-processor X5677 (4 cores) or X5680 (6 cores) with six 4GB RAM DIMMS -- at least if your budget permits.  Of course computer technology is constantly evolving and something better will come along soon, if it hasn't already.