Performance with AMD

[abhimodak]

Hi

I wondering about performance of exe built with Intel Fortran on AMD processors. For example, LinPack test (that involves 1000x1000 matrix with factorization and solves) showed almost factor 2 reduction when used with AMD opteron 6274 when compared against Xeon 5687. I guess that for such a test, the clock-speed would matter. Operton6274 is 2.2GHz whereas Xeon5687 is 3.6 GHz.

However, looking at the results by polyhedron benchmarks (http://www.polyhedron.com/pb05-win32-f90bench_SBhtml), I see that an AMD processor with 3.2GHz is again a factor 2 slower when compared against an i-5 with 3.3GHz.

This is quite an open ended querryand I am sure there has been some discussion in past. I couldn't find any threads on this forum. Can someone please enlighten?

Abhi

[TimP]

Not only between one manufacturer and another, there is a trade-off between number of hardware threads supported and performance per thread. For example, if you use only 1 thread on a CPU which supports 32 threads, you don't expect the performance you get on a possibly more expensive CPU supporting fewer threads.
There are many widely quoted benchmarks in various areas of analysis. Several of the SPECfp benchmarks are normally carried out with Fortran, with ifort well represented.

[abhimodak]

Many thanks Tim. I have checked SPEC website and there is a lot of information; will take time to understand. If possible, can you point me to something along the lines of the polyhedron benchmarks which will make it easier to understand?

Abhi

[Steven_L_Intel1]

The Polyhedron site is intended to compare Fortran compilers, holding all else the same. I think the AMD and Intel CPUs used are of different generations, so simply comparing clock speeds is not appropriate. Also, performance can be highly dependent on the type of workload and, of course, the compiler used.

Intel's goal is that you should get the best performance with the combination of Intel processors and Intel compilers. If you choose non-Intel processors, our goal is to deliver as good or better performance on that processor than any competing compiler. We don't always meet this on every program, but as the Polyhedron results show we do reasonably well against the goal. For more information on compiler choices, see our Optimization Notice.

Media sites such as Anandtech and Tom's Hardware frequently run comparisons across a range of current and earlier CPUs from Intel and AMD. They tend to focus on games, but have other tests that may be more relevant to you.