First Linux Benchmarks Of AMD FX-8150 Bulldozer

[ChrisXY]

they thank you and i know this: "Well even with Integer and Floating point calculations you can still get even better performance with fine tuned hand written optimizations. "

i just want to write some basic stuff about compilers.
the people are so brain-death they think the compiler fix all there problem magic magic...

the difference between compiler in Integer+Floating vs hand written Assembler is not so big than Compiler vs SIMD units....

SIMD instructions set blow the compiler away...

But what well optimized compilers do nowadays is kind of magic.
I remember running a knapsack programming assignment written in ADA and compiled with gcc-ada on some server with Opteron 275 and on my netbook with an Atom N270. The server had Ubuntu with gcc 4.4 and my netbook Archlinux with gcc 4.5 I believe. The netbook ran it faster...

Now I'm a bit tired and I'm not very deep into compilers, but can compilers really not "abuse" those instructions and for example everytime the program needs to execute something like a=b*c+d (how often that may happens) run it via the FMA instruction?

[PsynoKhi0]

You say it needs improvements in the kernel, well who exactly stopped AMD from pushing this 1 year ago when they had their first Bulldozer samples ready, could test and all of that. That's how you do things, not wait till it's out, then say, oh but it will work better with these.

Fear of giving away too much info too early? Last minute die changes? I'm actually curious.
I do think BD has had a fairly catastrophic release, though people paint all black way too fast.

[deanjo]

If AMD were to start doing their own thing with instructions, then AMD's x86 would quickly turn into a fringe server architecture that only runs operating systems specially compiled for it with GCC, like any other number of CPUs from IBM, Sun, etc... So following Intel's lead is still their best option, and things won't change until regulators grow a pair and decide to break up the Intel monopoly racket.

Can't say I agree there since quite the opposite happened when AMD did introduce x86-64 extensions which where quite different then anything intel was offering at the time.

[curaga]

I disagree. Intel regularly uses it's monopoly to force it's technology on everybody, and the constant influx of SSE instructions are what keep the x86 monopoly going. AMD could come up with their own instructions, but how well do you think they would do with incompatible instructions and significantly less marketshare? Furthermore, there's only so many instructions that are actually useful, Intel could sue them for IP infringement if anything was deemed to similar.

If the new instructions were actually useful, they would get used. See Via Padlock, being supported in nearly every relevant sw (you can even mine bitcoins on it, as fast as on an Intel quad ).

[deanjo]

Is this a new forum gimmick? Deanjo's golden stars?

Did you want one too? lol

[pingufunkybeat]

Is this a new forum gimmick? Deanjo's golden stars?

[deanjo]

Yes sure. but i think its like fighting again brain-death zombies they never understand the True nature.

Compilers only work for Integer and Floating point calculations but not for SIMD SSE,SSE2,SSE3,SSE4.2,AVX,FMA3,FMA4,XOP calculations.

Because these SIMD operations are not logical and not causality but a Compiler only abstract the Logical and Causality from a Logical program language construct to an 01 Computer code for the Integer or Floating point unit.

In a world of 100% compiler makes the work there is no SIMD(SSE-unit)

Well even with Integer and Floating point calculations you can still get even better performance with fine tuned hand written optimizations. But I will still give you a gold star. ;D

[nepwk]

What you say has some basis, the compilers can be optimized to favor Intel, at least on Windows. On Linux though gcc is open source, no one stops AMD from contributing. Also being beaten by your old CPU in some benchmarks while having 110% more transistors is unforgivable. This is not Intel or the compilers, you're losing to your own older generation.

I disagree. Intel regularly uses it's monopoly to force it's technology on everybody, and the constant influx of SSE instructions are what keep the x86 monopoly going. AMD could come up with their own instructions, but how well do you think they would do with incompatible instructions and significantly less marketshare? Furthermore, there's only so many instructions that are actually useful, Intel could sue them for IP infringement if anything was deemed to similar.

If AMD were to start doing their own thing with instructions, then AMD's x86 would quickly turn into a fringe server architecture that only runs operating systems specially compiled for it with GCC, like any other number of CPUs from IBM, Sun, etc... So following Intel's lead is still their best option, and things won't change until regulators grow a pair and decide to break up the Intel monopoly racket.

[del_diablo]

Linus Torvalds stopped AMD.

Nah, he accept anything that is clean code. Its as simple as that. So far nothing has stopped AMD from "contributing", or to be more spesific "upstreaming such a thing".

[deanjo]

and you don't get the point that the instruction set for hand optimization code isn't a compiler problem the GCC can't magic do the job for you. you have to rewrite the software to benefit from AVX and FMA4 and SSE4.2

Wow, I'm impressed Q. That's a point that I have been trying to beat into some peoples heads for a long time.