AMD vs Intel

[vlado]

You may have heard that AMD filed an antitrust case against Intel this summer; one of the allegations is that the Intel compiler purposefully generates code that runs slower on AMD processors than on Intel machines. Well, it is true - just made a few tests this week-end... Code compiled with recent versions of the Intel C++ compiler runs about 10% slower on AMD machines. Now guess which compiler we are using to build V-Ray

Luckily, our development machines are AMDs, and since newer compiler versions seemed to generate slower code, we are still using an older version that is much less sensitive to the processor type. Still, it was quite a shock when I found out about this...

Best regards,
Vlado

[nvanherpt]

yikes

that's good to hear you're on to it vlado, considering how many amd x2's have recently been purchased around here.

[percydaman]

interesting... nothing really surprises me anymore. I haven't purchased an Intel chip since i bought a PIII 500 quite awhile ago.

[Da_elf]

yeah. i honestly can say im not supprised. especially with intel going mac now im sure they want to try to block out competition.

[Morbid Angel]

dope, this is bullshit. Is there somekind of amd compiler :P??
These competition wars are not going to bring any one any good. Same goes to windows vista or whatever its called. The thing is supposed to be so anally packed with antipiracy and friendly features that you practicly have no control over your computer...

[GeneCrucean]

Wait, how can you prove this?

It's a completely different chip. It's not like your testing two P4's that are both the exact same speed. I'm not saying that the compiler does or doesn't compile slower code but before jumping the gun. Proof?

[Morbid Angel]

i guess proof would be you take 5 amd's and 5 intels run same code on all of them and if all amd's show 10% drop then you know its true.

[dbuchhofer]

he's talking comparitively between COMPILER versions.. meaning that its gotten slower with the new compiler, not necessarily in comparison to how it runs on intelium pentixeon 4's..


Unless i'm misreading

[GeneCrucean]

I'm confused, it's the compiler speed were talking about? Like the amound of time it takes to compile a build?

[cpnichols]

I think it is the speed of the program after it has been compiled with the compiler.

[GeneCrucean]

he's talking comparitively between COMPILER versions.. meaning that its gotten slower with the new compiler, not necessarily in comparison to how it runs on intelium pentixeon 4's..


Unless i'm misreading

Well that's what I thought but then are you saying that the builds have gotten slower on the same computer when compared to older builds? Or are you saying that you compiled for the same machine with an old compiler and the new compiler and the new one is slower? Have you tried this on a P4 or Xeon?

Sorry guys, I'm just trying to make sure things are clear here.

[vlado]

For some proofs by others, see these links:

http://www.intelliot.com/blog/archives/2005/07/12/
http://yro.slashdot.org/comments.pl?...3&cid=13042922
http://www.swallowtail.org/naughty-intel.html

Basically, some routines (e.g. copying of a block of memory) have several versions. The program selects a version based on the processor on which it is running. Very roughly, the resulting code looks like this:

if (intel_processor) use version A;
else use version B;

It so happens that "version B" is significantly slower than "version A".

That means that even if you use the same program, it will behave differently on Intel and non-Intel processors - so, for example, any benchmarks performed with this program will not be valid - you are no longer testing the same sequence of instructions.

As far as V-Ray is concerned, we are still using version 7.1 of the Intel compiler. I have tried to compile V-Ray with recent versions (8.1 and 9.0); the resulting code was always a little slower when rendering, compared to the 7.1 version - which is why we never saw a reason to upgrade. Incidentally, the machines on which we have tested this have always been with AMD processors.

This weekend I examined the code generated by versions 7.1 and 8.1 of the compiler. With general settings, version 7.1 generates code that virtually does not depend on whether the processor is genuine Intel or not. However, version 8.1 generates code that is much more dependent on the processor type. Comparing them head to head, the 8.1 code was a few percent slower than the 7.1 code on my dual Opteron machine.

I then applied the patch described in this page to vray60.dll, vrender60.dlr and libmmd.dll:
http://www.swallowtail.org/naughty-intel.html
This patch effectively removes the check for the manufacturer of the processor (Intel or non-Intel), but leaves other checks on (e.g. SSE or SSE2 support).

The result was a speed-up of 10% compared to the 7.1 code, on the very same Opteron machine.

Best regards,
Vlado

[RErender]

The result was a speed-up of 10% compared to the 7.1 code, on the very same Opteron machine.

Does it stay the same speed on the intels or does it get slower?

[GeneCrucean]

Thanks for the info Vlado. I appriciate it.

[Da_elf]

yeah. so a compile with the old one and then a compile with the new one. test both on both types of chips. if the intel remains the same bu the amd drops by 10% then we know the new one favors intel for sure