Floating-point advantage?

mannyg

Distinguished
Oct 20, 2009
4
0
18,510
So I've been in the hunt for a low budget home engineering PC. I intend to write code in FORTRAN and run FEM/FEA applications. I suspect that floating-point computation performance will be more important to me than integer computation. So as best as I can tell, the Whetstone GFLOPS benchmark should be the best indicator for my needs in Tom's suite of benchmarks. What strikes me is AMD's relatively strong GFLOPS performance compared to Intel (e.g. Athlon II x4 620 benched 30.3 gflops vs the much more expensive I5's 33.0 performance -- a 10% difference for a 2 fold price diff). So am I correct in relying solely on the Whetstone benchmark for floating-point computation performance? The only other comparable benchmark I've seen is CPU World's Linpack benchmark. Does AMD still place stronger emphasis on FP computations than Intel?

Manny
 

MarkG

Distinguished
Oct 13, 2004
841
0
19,010
Intel assume that anyone who cares about floating point performance will be using SSE and not x87 instructions, so they've let x87 support rot while adding more and more variants of SSE; does your compiler generate SSE code or x87 code?
 

mannyg

Distinguished
Oct 20, 2009
4
0
18,510


So are FLOPS benchmarks such as Whetstone and Linpack optimized for SSE instructions? If so, wouldn't both AMD and Intel chips benefit from this in the benchmark tests?



The compilers used are g95 and gfortran. I believe that they both have an SSE optimization option. Never tried it since my current CPU is a Thunderbird.
 
A inexpensive AMD quad core or a cheap INTEL dual core cpu will be good enough so long they are overclocked. Remember both perform around 5~10% of each other on a per clock basis. Intel overclocked 3~4ghz dual core 2~3.5ghz Intel quad. AMD quad 2.8~3.4ghz+.
 

mannyg

Distinguished
Oct 20, 2009
4
0
18,510


Maybe based on overall benchmarks, but when looking at FP computations, an Athlon x4 is almost neck and neck with an I5--at least according to Whetstone's glfops benchmarks. When comparing Linpack 100x100 benchmarks, the x4-630 scored a 1643 MFLOPS vs 1093 MFLOPS for I7-920. Now there may be more to this. That's why I'm seeking enlightnment here.
 

mannyg

Distinguished
Oct 20, 2009
4
0
18,510


I don't quite follow. Are you talking about the optimization settings used in the compilation process or are you talking about the way the FORTRAN code is written (i.e. calling libraries specific to SSE instructions). I would think that the way the FORTRAN code is written would not really matter. I fact, I would want the code to be portable, so any functions called would need to meet FORTRAN (90 or 95) standards which I believe is independent of x86 and SSE, correct?

Also, keep in mind that I will be running pre-compiled engineering applications that may not be optimized for specific chips. So, general FP computation benchmarks matter.
 

MarkG

Distinguished
Oct 13, 2004
841
0
19,010


SSE should benefit both, but AMD have generally had better non-SSE performance than Intel since the Pentium-4; I'm not sure whether that still applies today, but I would expect so.