Single or double precision?

[rase]

I am considering to change the floating point precision in my Fortran projects from single (KIND=4) to double (KIND=8) generally. I remember to have read somewhere that IA32 CPUs do all floating calculations with double precision anyway, so the overall computing time should be the same, independent of single or double precision. Is that assumption true? Has anybody done systematic timing tests concerning single and double precision?

Greetings, Wolf

[Steven_L_Intel1]

It is true that by default the actual computations are done in double precision, but there is extra memory traffic because of the larger data. I would not expect to see a big hit on performanc. For better performance, use one of the "Generate code for ..." (/Qx) options appropriate for your processor so that you use the SSE instructions. These are faster than the old x87 code.