GCC 4.7 AMD Bulldozer Compiler Tuning

[phoronix]

Phoronix: GCC 4.7 AMD Bulldozer Compiler Tuning

As the latest AMD Bulldozer Linux benchmarks, here are updated figures on compiler tuning for the FX-8150 processor when using GCC 4.7.1.

http://www.phoronix.com/vr.php?view=17911

[clavko]

Oh, come on. There can't be so much difference just with compiler flags.
That would mean Gentoo is faster than Calculate Linux, and we damn
well know that it ain't so. Stop this madness once and for all, Michael!

[uid313]

The real question is, how does this compare against Intel Core i5 and Core i7?
How does it compare against Ivy Bridge?

And does it have good graphics device drivers on pair with Intel?

Yeah, we all know that AMD have better graphics hardware than Intel.
But if the drivers suck, then its useless.

[hoohoo]

Oh, come on. There can't be so much difference just with compiler flags.
That would mean Gentoo is faster than Calculate Linux, and we damn
well know that it ain't so. Stop this madness once and for all, Michael!

I have measured cases where the march compiler flag made a significant difference (10 or 20% difference) to the speed of resulting binary. They can make a real difference.

BD being a weird new arch needed from compiler makers and OS makers to update their code, just like every new arch does.

When BD initially got to market, the performance numbers were strange and strongly suggested that there were things that could be done with code optimization.

That is why you see articles like this one periodically: new versions of compilers should (hopefully) produce better code for BD (and other CPUs to whatever degree thee compiler makers have worked on the CPU).

[mmstick]

Oh, come on. There can't be so much difference just with compiler flags.
That would mean Gentoo is faster than Calculate Linux, and we damn
well know that it ain't so. Stop this madness once and for all, Michael!

You obviously don't know much about how compilers work...

[Dukenukemx]

So what does Ubuntu 12.10 use?

[mmstick]

What compiler flags the kernel was compiled with doesn't really impact the programs you run that much (in fact they are already defined and you cannot change them much). All things considered, the kernel runs well on high end processors as is, but I'm sure a nice compiler optimization might improve your boot times slightly. What really matters is processor intense applications need to be compiled well so that what takes a lot of time, will take less time. In some cases, as in those that make heavy use of AVX, XOP, FMA4 instructions, you can see a double improvement in processing speed if you compiled it with bdver1.

[bridgman]

And does it have good graphics device drivers on pair with Intel?

It's a discrete CPU... doesn't need graphics drivers.

[clavko]

@hoohoo, @mmstick
Sorry guys, it seems that my irony was too deep yesterday
I'm a Gentoo user myself, and am aware of the advantages
that compiler flags bring, and these are not negligible.

I was refering to benchmarks that were published on Phoronix
pitting Sabayon or Calculate Linux against Ubuntu/Arch. Every
time it was stressed that these distro's were 'Gentoo based'.
But both of these distro's come with precompiled binaries and
you can be sure that there are no specific cpu optimizations.
Therefore, performance like, they can't be like Gentoo.

Perhaps, now that "we know" that cpu flags give performance
it is time for a real Gentoo benchmark

[mlau]

I'm curious: some of the Benchmarks shown use external libraries.
Were those rebuilt with the changed cflags as well? And glibc/gcc itself (for the fortran libs for instance)?