Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on August 10, 2014 8:30 am wrote:
> Aaron Spink (aaronspink.delete@this.notearthlink.net) on August 10, 2014 8:04 am wrote:
> > Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on August 10, 2014 5:00 am wrote:
> >
> > > Also there is no point in comparing with Intel SPEC results due to the Intel compiler breaking
> > > some of the SPEC benchmarks - the correct result is likely about 55. If you want to compare
> > > CPU performance using SPEC, you have to use the same compiler and settings.
> > >
> >
> > I'll take the voracity of a fully sumbitted SPEC score over some estimate any day of the week.
> > You are making so many assumptions about that *estimated* score and then trying to rake Intel
> > over the coals for fully accepted results. AKA it doesn't matter if you think the Intel compiler
> > is breaking things, its an fully accepted result by SPEC, your argument holds no water.
>
> Whether official or not, the Intel SPEC scores do not translate into real world
> performance - if you believe otherwise you're deluded. Using the Intel scores
> as a basis for anything besides marketing bragging rights is just insane.
>
> To get a reasonable estimate of real world performance you can run SPEC using GCC, and this is
> what most companies actually do - after all GCC is the main compiler when using Linux servers.
>
> Wilco

If I am not mistaken, when compared to its main server competitors, i.e. Power and SPARC, gcc-all-around comparisons are more favorable to Intel (x86, not Itanium) than vendor's compiler vs vendor's compiler.

Also, among those 3 vendors + AMD, there appears to be good correlation between official SPECInt_rate scores and SAP SD 2-tier scores. Back when TCP-C was not dead, there was not a bad correlation between it and SPECInt_rate as well, except for big-core-count Itanium vs big-core-count IBM Power.