Excellent article on "P4 optimizations"

Gerry

Here's a very interesting article I found today, which shows the potential of the p4, if software was compiled for it. It looks like this is becoming much easier, with an intel plugin for microsofts VC++ 7.

http://www.aceshardware.com/Spades/read.php?article_id=40000189

[This message has been edited by Gerry (edited 03-07-2001).]

sutty

That seams a bit one sided, Why didn't the try an athlon compiler? There are ones made by AMD. That MS one is a generic one isn't it?

---

Ciaran Sutcliffe
aka: sutty
[HIV]sutty
For a good time goto:
http://www.hotinternetvirgins.com

[This message has been edited by sutty (edited 03-07-2001).]

Kali

newer athlons support SSE now anyway... any optimization made to binaries will most likely increase speed on AMD chips as well.

Gerry

Kali, the major improvements came from sse2, which the athlon, and the palomino do not support. Sutty, I dunno where these athlon compilers are? The athlon would get less benefit from a specialized compiler anyway, since it is designed to perform well on code optimized for the p6. As for it being one sided, well I thought it was remarkable that a compiler turned the tables entirely.

Kali

Even so the article is only dealing with FPU performance anyway in which the athlon pretty much destroys the p4 (even with intel-sse2 optimized compiling).
Its only when loops within the fpu programs (Flops) code were vectorized that a huge amount of difference was seen in the P4s performance.. up to a 240% increase for certain tests, something which I doubt AMDs own compiler (plugins) would be able to pull from the Athlon, which I'd warrant is close to its peak in those tests.

ill just throw in the most important graph for people not bothered to read the while thing

Gerry

yeah, well the last graph was the point of the whole article. Of course flops is a synthetic benchmark, but it pinpoints precisely where the p4 is failing right now in real world apps. The argument that this synthetic performance won't transfer across to real world apps doesn't work with the p4. It already has the advantage of greater memory bandwidth, when it gets fast ddr ram (with much smaller latency), and a bigger l2 cache (possibly a bigger l1 cache as well), it will probably beat the athlon at the very things which it gets trounced at right now, i.e. 3d rendering.

Thats why I felt this article was so significant. The performance in the last graph was not achieved by hand coding the benchmark in assembly just for the p4, it was with a simple compiler plugin, for the most widely used compiler. This means developers should be able to get programs running MUCH quicker on the p4, conveniently and quickly. Of course intels compiler still has a long way to go with reliability, but the p4 doesn't look such a bad chip now, and the early adopters may end up with decent performance after all.

_CreeD_

P4's are catching up with Athlons, and surpassing them now at their highest clockspeeds. BUT to do it they have to run at much higher CPU speeds and more importantly cost a hell of a lot more.
Even with easier compiler techniques I don't see P4 optimisations really becoming mainstream (A few of the bigger titles, and probably the odd add on patch maybe), simply because the P4 is a bit of an interim CPU. It's likely the last real re-design of the x86 architecture. 64 bit CPUs will be out in a few months, granted aimed at the workstation market initially, but why will any developer/programmer bother getting to grips with netburst optimisation when getting in on the groundfloor with EPIC makes more sense in the long run?
If Intel had their old market share this wouldn't be an issue, as the P4 would be the mainstream home-performance CPU. But it's not.
Also the Athlon is finally going the MP route with the 760mp chipset. After the usual early adopter rip off prices, I'd imagine a dual Palomino system will be about the same as a high end single P4 setup, again positioning it attractively in the 3d workstation market.

Kali

just on the 64 bit chip point there.. I've actually been looking for recent (i.e. this year) information on Intels IA-64, was this actually scrapped? because it seems to be rare to find anything post-august 2000.
and thats the last time I had read about it, stating it wasn't ramping very well and performance was on a par with a low end pentium III.

[edit]
oops new 64 bit project is McKinley not the crappy Itanium/Merced/IA-64 project.. which had 32bit x86 performance somewhere between a 486 and a low-end Pentium, not as I had stated above... whoa... sucky.
[/edit]

[This message has been edited by Kali (edited 04-07-2001).]