[Eug Wanker]

IBM XL Fortran Advanced Edition V8.1 for Mac OS X

Full compliance with Fortran 77, Fortran 90, and Fortran 95 standards
Partial draft Fortran 2000 standard support
IBM and industry extensions
Industry-leading optimization technology
Exploitation of G5 architecture
Symbolic debugging support

XL C/C++ Advanced Edition V6.0 for Mac OS X

Generation of highly optimized code which exploits the IBM PowerPC architecture within the Power Mac G5
C and C++ programming language standards conformance
Binary compatibility and coexistence with gcc V3.3
Support for the Velocity Engine through the AltiVec programming interface
GCC command line compatibility
Documentation that supports the Mac OS X help architecture.


No details on the pricing though.

[Krypton]

Will this have the side effect of faster code for G3 and G4 machines as well?

[Turias]

Originally posted by Eug Wanker:
No details on the pricing though.

How much do similar compilers usually go for?

[Turias]

Originally posted by Krypton:
Will this have the side effect of faster code for G3 and G4 machines as well?

No, this is for the G5 only.

[Eug Wanker]

Originally posted by Krypton:
Will this have the side effect of faster code for G3 and G4 machines as well?

Well, I'm not a programmer, but the programmer types at Ars say it gives a big speedup for certain types of code on G4s too, but obviously it's best on the G5. Dunno about the G3, but I'm guessing probably yes (if you use the most generic settings) since these compilers do not auto-vectorize code.

Original posted by Turias:
No, this is for the G5 only.

See above. There are settings that optimize only for the G5 though.

How much do similar compilers usually go for?

Well, I'm not sure, but Intel gives its compilers away for free for educational use, and charges I believe hundreds to thousands for commercial use.

[ratlater]

If a developer used this IBM compiler to make the program G5 optimized, would they then need to use a different complier and create a different version for non-G5 machines? Or would one compile with the IBM compiler make the program G5 optimized and still work with G3s and G4s?

thanks,
-matt

[Eug Wanker]

Check out page 21 and 22 of this PDF to see IBM's xlc/xlf performance vs. gcc/g77. (GCC is the default compiler in Apple's Xcode right now.)

Personally, I think this announcement is foreshadowing a killer Power Mac announcement.

[larkost]

From the information on their site you can tune for either G4 or G5 (no mention of the G3, so I would have to guess no).

To me the most important note in the announcement is that it does compile Obj-C. That was not in the preview, but is explicitly stated on IBM's site:

Objective-C
XL C/C++ Advanced Edition compiler supports Objective-C.

Objective-C is an object-oriented programming language based on standard C and is popular in the Mac OS X programming environment. The Objective-C programming is a basis for writing to Cocoa frameworks.

The other note (nothing new) is that this is intended to work within XCode, and is GCC 3.3 compatible (mostly).

[dashiel]

Originally posted by Eug Wanker:
Check out page 21 and 22 of this PDF to see IBM's xlc/xlf performance vs. gcc/g77. (GCC is the default compiler in Apple's Xcode right now.)

Personally, I think this announcement is foreshadowing a killer Power Mac announcement.

wow this could be huge, perhaps i'm misreading this but, those graphs show xlc02-05, the official release from IBM states it's version 6.

does that mean that we could see even more performance enhancements from those graphs?

this could really bury the whole "controversy" over the "fastest personal computer on the planet". as i recall most windows sites complained that the tests apple used, both computers were using GCC and not intel's compilers for the intel chips. apparently if they had used the intel compilers, the intel machines outperformed the G5, but not drastically. if i'm reading these graphs correctly, these compilers might erase that difference completely, if not put the G5 categorically ahead.

of course i know f all about compilers

[geekwagon]

Originally posted by Turias:
How much do similar compilers usually go for?

Sun charges $1000 for their Workshop Compilers. I haven't had to buy IBMs compilers yet but I am guessing it would be similar.. That said they might make it cheaper on OS X because it is more of a client platform..

[Eug Wanker]

Originally posted by Eug Wanker:
these compilers do not auto-vectorize code

Hmmm... What does this mean?

The -qaltivec option which supports the Power Mac G5 Velocity Engine through the AltiVec programming interface. This interface provides for highly parallel operations for both integer and float point data.

Is this just support for Altivec using standard programming methods (since IBM's other chips don't support Altivec)?

Original posted by dashiel:
wow this could be huge, perhaps i'm misreading this but, those graphs show xlc02-05, the official release from IBM states it's version 6.

Well a few things:

1) Those are the results for the SPEC tests, on different hardware and a different OS. While still *nix based, it's not quite the same thing as a G5 with Mac OS X.

2) -05 is the level of optimization. -05 does super duper optmization, but may break the program. -03 is safer, but the speed of the final program is not as fast.

3) Some programs will get a huge speedup, while other will not get much speedup at all.

Originally posted by larkost:
To me the most important note in the announcement is that it does compile Obj-C.

Hey cool, if really true.

[mudmonkey]

Originally posted by geekwagon:
Sun charges $1000 for their Workshop Compilers. I haven't had to buy IBMs compilers yet but I am guessing it would be similar.. That said they might make it cheaper on OS X because it is more of a client platform..

XL Fortran for AIX costs slightly over $2,100.

Let us all hope to all hopes that Apple helped fund the compiler development and will include it in the developer tools...

EDIT: Just looked up the price from IBM for XL Fortran on AIX (no Mac pricing available): $2,921.00

[Eug Wanker]

Originally posted by mudmonkey:
Just looked up the price from IBM for XL Fortran on AIX (no Mac pricing available): $2,921.00

Hmmm... Not for your average programmer. Mind you, AIX isn't your average OS either. Hopefully it'll be cheaper for the Mac OS X types.

What about the cost of XL compilers on AIX for educational users? Just wondering.

EDIT:

It's $499 for the single user licence, but no word on educational pricing yet.

[Deestar]

Because of it speed gain on G4 and G5 computers do you guys think that apple will compile 10.4 with these IBM compilers and make it a G4 and G5 only upgrade?

[Eug Wanker]

Originally posted by Deestar:
Because of it speed gain on G4 and G5 computers do you guys think that apple will compile 10.4 with these IBM compilers and make it a G4 and G5 only upgrade?

1) AFAIK, compiled code will run on the G3 (depending on the settings).

2) Considering the G3 iBooks are still out there for sale, Apple would never build X.4 to exclude machines without Altivec.

[Charles Bouldin]

I've been using the beta version of xlf 8.1 since last August, and it is >>>excellent<<<. Solid, bug-free, highly optimizing. On two processor runs of computational fluid dynamics codes the 2.0 ghz G5 with xlf is almost twice as fast as a 2.8 ghz Pentium 4, ie, the G5 machine is as fast as a hypothetical 5 ghz Pentium!

I am very impressed with xlf. Can't comment on xlc, or compatibility with objective C, but it sure seems to me that Apple should recompile the entire OS with xlc.

In combination with the G5, xlc and xlf make for a very serious scientific workstation. I suspect that eventually things like Photoshop will have their core code built with xlc and will run much faster.

Edit: I forgot to add that, if the $499 price is real (source for that???), this is a bargain. I expect the edu price will be even better. If this integrates nicely with xcode, I think a lot of developers will want it.

[Deestar]

Originally posted by Charles Bouldin:
I've been using the beta version of xlf 8.1 since last August, and it is >>>excellent<<<. Solid, bug-free, highly optimizing. On two processor runs of computational fluid dynamics codes the 2.0 ghz G5 with xlf is almost twice as fast as a 2.8 ghz Pentium 4, ie, the G5 machine is as fast as a hypothetical 5 ghz Pentium!

I am very impressed with xlf. Can't comment on xlc, or compatibility with objective C, but it sure seems to me that Apple should recompile the entire OS with xlc.

In combination with the G5, xlc and xlf make for a very serious scientific workstation. I suspect that eventually things like Photoshop will have their core code built with xlc and will run much faster.

Bloody Hell, pretty impressive!

[Eug Wanker]

Originally posted by Charles Bouldin:
I've been using the beta version of xlf 8.1 since last August, and it is >>>excellent<<<. Solid, bug-free, highly optimizing. On two processor runs of computational fluid dynamics codes the 2.0 ghz G5 with xlf is almost twice as fast as a 2.8 ghz Pentium 4, ie, the G5 machine is as fast as a hypothetical 5 ghz Pentium!

I am very impressed with xlf. Can't comment on xlc, or compatibility with objective C, but it sure seems to me that Apple should recompile the entire OS with xlc.

In combination with the G5, xlc and xlf make for a very serious scientific workstation. I suspect that eventually things like Photoshop will have their core code built with xlc and will run much faster.

Edit: I forgot to add that, if the $499 price is real (source for that???), this is a bargain. I expect the edu price will be even better. If this integrates nicely with xcode, I think a lot of developers will want it.

What were your CFD speeds without xlf?

It seems that 1.4:1 speed ratio for G54 speed keeps popping up. That means if Apple is able to release a dual Power Mac any time soon that's anywhere north of 2.4 GHz, it WILL be the fastest desktop computer on earth for many fp-intensive apps. It would be roughly twice as fast as a non-Prescott Pentium 4 3.4 GHz (if such a beast existed) for CFD.

As for the $499 price, somebody at Ars bugged a bunch IBM sales guys and that's what they told him. We'll know for sure on Friday since that's when everything goes online.

[awrc]

Originally posted by Eug Wanker:
Hmmm... What does this mean?

The -qaltivec option which supports the Power Mac G5 Velocity Engine through the AltiVec programming interface. This interface provides for highly parallel operations for both integer and float point data.

Is this just support for Altivec using standard programming methods (since IBM's other chips don't support Altivec)?

I'd interpreted it as being a flag to enable access to the API for the AltiVec extensions. Leave the flag out and you can't use 'em - safer to have the default be "no G5 specific optimizations" than the other way around.

So if you've got a litle loop performing some simple math on an array of 1024 pieces of data...

If the compiler was auto-vectorizing it'd spot the parallelism during analysis and automatically do something like strip-mining it to get break it up into AltiVec style chunks. Auto-vectorizing compilers sometimes also do clever stuff like extract parallelism from places you'd never expect to find it. Basically anything doing that level of analysis and restructuring of code can come up with strange and unusual mutations of your code that you'd never find yourself.

My favorite (possibly anecdotal) compiler optimization story is of an IBM test compiler that did *very* thorough analysis of dataflow and variable use. The test program did all sorts of heavy math, but didn't do anything with the result except print it to the screen. Since the result was always the same, they ended up removing the print statement. This resulted in the compiler optimizing the program out of existence completely.

Hmm, I have a "favorite compiler optimization story". I need a life.

With -altivec you can write your loop using AltiVec specific calls to handle the data in parallel chunks. Never really looked at the internal architecture of AltiVec in its various guises, so I don't know the numbers. You get all the benefits of the AltiVec unit for those cases where you can spot the parallelism and exploit it.

Without the -altivec flag it's one element at a time, unless the compiler's capable of taking advantage of software pipelining, or similar. Again, not being acquainted with the G5 architecture, I can't say whether this is the case or not, as all my architectural background is a decade old and refers to compiler optimizations aimed at supercomputers.

Strangely, everything I learned then has become useful again in the past couple of years, as the features you found in supercomputers then are beginning to turn up in desktop computers now.

Well a few things:

1) Those are the results for the SPEC tests, on different hardware and a different OS. While still *nix based, it's not quite the same thing as a G5 with Mac OS X.

True, but SPEC is a decent way of at least comparing oranges with oranges, even if your primary interest is apples. Hmm, bad pun. The two used to be very different fruit, but nowadays, when smple GUI operations exploit 3D cards, maybe it's not so different any more.

2) -05 is the level of optimization. -05 does super duper optmization, but may break the program. -03 is safer, but the speed of the final program is not as fast.

The cases where high levels of optimization break code are fairly few and far between and quite often result from bad programming practices/too-clever coding. The cases where high levels of optimization cause the size of generated code to explode due to berserker loop-unrolling and where the resulting code is impossible to debug are much more common.

3) Some programs will get a huge speedup, while other will not get much speedup at all.

Hey cool, if really true.

That's about the size of it. No matter how much parallelism you can take advantage of, you're going to be bounded by that part of your code that can't be parallelized. This part can be pretty big.

Al

[dashiel]

Originally posted by Eug Wanker:


1) Those are the results for the SPEC tests, on different hardware and a different OS. While still *nix based, it's not quite the same thing as a G5 with Mac OS X.

2) -05 is the level of optimization. -05 does super duper optmization, but may break the program. -03 is safer, but the speed of the final program is not as fast.

3) Some programs will get a huge speedup, while other will not get much speedup at all.

thanks for the clarification.

[i_wolf]

I have been using the beta for some time now... and I can honestly say that i have never yet broken a binary by using the highest level of optimizing in XLC.
That was with a beta build as well. I would imagine that the latest release is rock solid. IBM have a good reputation for solidly made compilers.
From what i gather pretty much most apps should get a big speed up while some if tweaked will get a MASSIVE speed up!
Most developers I have spoken with agree that the situations where XLC wasn't that much faster than GCC was probably down to its early beta build. These situations to reiterate were few and far between. IMO, XLC and XLF also benefits G5's current 'non spectacular' integer performance greatly. Those who claim that G5's integer performance is only "so - so" compared with opteron and P4 should recompile with this compiler. In some cases it produces Integer code that is 60 to 80% faster than what GCC achieves. Most cases it sees about 40 to 50 integer boost from my tests.

I would imagine that this latest build would prove much better in the field. Also there were not that many cases where XLC yielded little benefits over GCC. Most of the time it made gigantic performance differences.
Another thing thats pretty amazing is that IBM have yet to implement autovectorising in their compiler. This feature alone could offer even bigger performance leaps again.
I wonder if Apple will offer their members in the ADC a special price on this compiler?!?
Whats even more interesting is that IBM specifically made a point of referencing that this compiler set is for OS X and G5. The future looks very good. Never got this type oof support from Motorola!

[Eug Wanker]

G5 1.8 SPEC results from way back, with gcc/g77:

int: 937
fp: 1051

Anyone want to wager how high it will go with xlc/xlf?

[staypufd]

From the product page...


XL C/C++ Advanced Edition compiler supports Objective-C.

Objective-C is an object-oriented programming language based on standard C and is popular in the Mac OS X programming environment. The Objective-C programming is a basis for writing to Cocoa frameworks.

[theolein]

This is fantastic. I can see the next generation of both OS and applications being compiled with this to make the Mac really shine.

[[APi]TheMan]

Originally posted by theolein:
This is fantastic. I can see the next generation of both OS and applications being compiled with this to make the Mac really shine.

Yeah.

People should release software like this every day. Too bad it's 11:31 PM... This just made my day. Hah!

[i_wolf]

well there is nothing stopping anyone here from downloading the beta version and going and compiling some of the open source apps that are already out there.

[mudmonkey]

Originally posted by i_wolf:
well there is nothing stopping anyone here from downloading the beta version and going and compiling some of the open source apps that are already out there.

Well, the betas came out before Panther, and prebinding libraries generated from XL is broken in Panther.

There are other issues with the beta which, undoubtedly, are fixed in the release.

[Eug Wanker]

Hmmm... $499 may not be quite right.

$499 just for xlc. Dunno about xlf.

[Eug Wanker]

Originally posted by i_wolf:
Another thing thats pretty amazing is that IBM have yet to implement autovectorising in their compiler. This feature alone could offer even bigger performance leaps again.

Dunno what progress they've made with the XL compilers, but IBM is at least working on auto-vectorization for gcc. That message was from 6 months ago though.

[i_wolf]

excellent stuff!!! However I hope that IBM also work with Apple to improve GCC. Currently it really doesn't begin to get the chip to work anywhere near its potential. I hope they also add this autovectorisation to XL range of compilers also... i would imagine this will happen though. It would be the icing on the cake. These compilers coupled with the G5 and OSX really make this a kick ass Unix platform. Nigh on untouchable performance wise for hte price and GHz.
Happy days.

EDIT: Forgot to say thanks for the link there EUG

[Eug Wanker]

Originally posted by Eug Wanker:
Hmmm... $499 may not be quite right.

$499 just for xlc. Dunno about xlf.

Posted elsewhere on the net:

Fortran compiler, one year of tech support and upgrades $999
C/C++ compiler, one year of tech support, and upgrades $499.

[Charles Bouldin]

Originally posted by i_wolf:
I have been using the beta for some time now... and I can honestly say that i have never yet broken a binary by using the highest level of optimizing in XLC.

I've seen things break with -O5 on XLF. I've tested two large codes with XLF, and one of them just runs wrong with -O5. It's a subtle thing, and at first I thought it had to do with using FMULADD (or whatever it is now called), ie, assembly instructions that do things like
b =a*x + c as one instruction. This is often subtly different than using separate instructions. However, that isn't it. Punchline is that the most aggressive optimization does break things sometimes.

That said, xlf rocks like no compiler we've ever seen on the Mac! I'd put xlf up with VAX, SGI, and Cray fortran....one of the best I've ever used, and far, far better than any Fortran available on the Mac before.

BTW, I also used the Portland Group 64 bit fortran on Opteron and the Intel 8.0 compiler on Itanium. These things break code much more than xlf when you turn up the optimizer. In fact, I haven't yet made meaningful comparisons with xlf/G5 because I can't get the damn things to compile our code! Mind you, I am speaking of a fortran code that compiles flawlessly with 7-8 other fortran compilers, including the 32 bit versions of Intel and PG. The 64 bit versions just don't work right.

[Charles Bouldin]

I consider this definitive information, because I called IBM's office in Canada and was told: They do "tiered" pricing. For me, the US .gov pricing is

xlf $609
xlc $304

Educational pricing

xlf $400
xlc $200

Everyone, run, don't walk to swamp IBM with orders. They should be rewarded for (1) a public beta that let everyone test the product, (2) this incredibly reasonable pricing.

[Eug Wanker]

Originally posted by Charles Bouldin:
I consider this definitive information, because I called IBM's office in Canada and was told: They do "tiered" pricing. For me, the US .gov pricing is

xlf $609
xlc $304

Educational pricing

xlf $400
xlc $200

Everyone, run, don't walk to swamp IBM with orders. They should be rewarded for (1) a public beta that let everyone test the product, (2) this incredibly reasonable pricing.

Nice to see those educational prices, but:

1) I am assuming those are US bux, correct?
2) I'm curious as to why you'd call IBM Canada to get US.gov pricing, in US$, esp. when you're based in Maryland.

BTW, it's been posted elsewhere that the full price is:

xlf $999
xlc $499

EDIT: Some additional info on the educational prices, posted elsewhere:

After a bit of hunting I got a hold of someone at IBM who gave me the
academic pricing for XLF and XLC/C++ compilers.

The pricing is as follows:
XLC/C++ $200 plus $25 for media
XLF $400 plus $25

This is a onetime fee. The license is a single user license.

After the first year, if you want upgrades support etc.. it costs:
XLC/C++ $39.90 / year
XLF $80 / year

[Charles Bouldin]

I'm curious as to why you'd call IBM Canada to get US.gov pricing, in US$, esp. when you're based in Maryland.

No mystery to it....I just called the phone number on the IBM web page. They're a multi-national corporation, and the compiler team is based in Ottawa.

[Eug Wanker]

Originally posted by Charles Bouldin:
No mystery to it....I just called the phone number on the IBM web page. They're a multi-national corporation, and the compiler team is based in Ottawa.

w00t! Go Canada!

[Eug Wanker]

Absoft is now selling IBM's xlf compiler, bundled with Absoft's tools, for the same price that IBM sells the compiler alone.

See here.

Commercial - $999
Academic - $599
Government - $399

The Absoft debugger itself is $300 if you buy it separately.

[ryaxnb]

People still use Fortran? Fortran stinks. Use the top 2 excellent replacements:
* REALBasic. Simple Fortran projects can easily be replaced with REALBasic programs.
* C++. C++ is a fast and efficient programming tool.

[Richard Edgar]

People still use Fortran? Fortran stinks

Indeed they do. And do you have anything to back up your rhetoric?

[Mr Scruff]

Originally posted by ryaxnb:
People still use Fortran? Fortran stinks. Use the top 2 excellent replacements:
* REALBasic. Simple Fortran projects can easily be replaced with REALBasic programs.

Please tell me you're joking.

[Eug Wanker]

So some crazy Germans have gone ahead and done SPEC2000 testing on a Power Mac G5 2.0 dual, comparing GCC/Absoft/NAGware vs XLF/XLC. No CodeWarrior tests though.

Code:
Compiler       SPECint2000   SPECfp2000

gcc 3.3            738
xlc                794

gcc/Absoft 8.2                  699
gcc/NAGware 4.2                 785
xlc/xlf                         977

xlc speedup for int code: 6-15%
xlc speedup for int code: 8% (vs. fastest tested non-xlc compiler)

xlf speedup for fp code: 22-41%
xlf speedup for fp code: 24% (vs. fastest tested non-xlc compiler)

[Catfish_Man]

Originally posted by Eug Wanker:
So some crazy Germans have gone ahead and done SPEC2000 testing on a Power Mac G5 2.0 dual, comparing GCC/Absoft/NAGware vs XLF/XLC. No CodeWarrior tests though.

Code:
Compiler       SPECint2000   SPECfp2000

gcc 3.3            738
xlc                794

gcc/Absoft 8.2                  699
gcc/NAGware 4.2                 785
xlc/xlf                         977

xlc speedup for int code: 6-15%
xlc speedup for int code: 8% (vs. fastest tested non-xlc compiler)

xlf speedup for fp code: 22-41%
xlf speedup for fp code: 24% (vs. fastest tested non-xlc compiler)

Woahhh... those SPEC scores SUCK. Apparently SPEC really dislikes the G5 and/or PPC compilers (for comparison, 3GHz P4s get about 1500/1500, and I don't think you get anyone arguing that the P4 has the same IPC as the G5).

[Eug Wanker]

Originally posted by Catfish_Man:
Woahhh... those SPEC scores SUCK. Apparently SPEC really dislikes the G5 and/or PPC compilers (for comparison, 3GHz P4s get about 1500/1500, and I don't think you get anyone arguing that the P4 has the same IPC as the G5).

BTW, those German scores are here. Yeah, they aren't the greatest, but note the G4 1 GHz scores (gcc 2.95)... SPECint: 307, SPECfp: 188 Motorola itself claims a G4 7455 1.25 GHz scores 500 int, and doesn't publish SPECfp scores at all. My uneducated guess is that a G5 2.0 is 2X-3X as fast as a G4 1.25 GHz in fp.

Note that c't didn't deactivate the other CPU or use special libraries. Both of those things might have increased the scores somewhat. It's interesting to note that Apple gets 800 int and 840 fp with gcc/NAGware, and IBM gets 828 int and 1036 fp with xlc/xlf, but with a slower 1.8 GHz G5.

A standard 3 GHz P4 doesn't get 1500/1500 though. Intel itself says with icc 8 it gets 1226 int and with ifc 7 it gets 1213 fp. Still, it'd take about a G5 2.5 to match those fp scores, and even a faster CPU to match those int scores, unless IBM can speed up the compilers. (For instance, Intel's compilers do autovectorization, but IBM's do not.) The one caveat about Intel numbers though (or so people have said) is that although Intel's compilers are super fast, they sometimes do strange things to real-life code. YMMV. (I'm not a programmer so I don't know.) People tend to use Microsoft's compilers etc. Scores with gcc or Microsofts compilers are much lower than Intel's.

BTW, I've graphed the c't G5 2.0 compiler info.

[Catfish_Man]

Originally posted by Eug Wanker:
BTW, those German scores are here. Yeah, they aren't the greatest, but note the G4 1 GHz scores (gcc 2.95)... SPECint: 307, SPECfp: 188 Motorola itself claims a G4 7455 1.25 GHz scores 500 int, and doesn't publish SPECfp scores at all. My uneducated guess is that a G5 2.0 is 2X-3X as fast as a G4 1.25 GHz in fp.

Note that c't didn't deactivate the other CPU or use special libraries. Both of those things might have increased the scores somewhat. It's interesting to note that Apple gets 800 int and 840 fp with gcc/NAGware, and IBM gets 828 int and 1036 fp with xlc/xlf.

A standard 3 GHz P4 doesn't get 1500/1500 though. Intel itself says with icc 8 it gets 1226 int and with ifc 7 it gets 1213 fp. Still, it'd take about a G5 2.5 to match those fp scores, and even a faster CPU to match those int scores, unless IBM can speed up the compilers. (For instance, Intel's compilers do autovectorization, but IBM's do not.) The one caveat about Intel numbers though (or so people have said) is that although Intel's compilers are super fast, they sometimes do strange things to real-life code. YMMV. (I'm not a programmer so I don't know.) People tend to use Microsoft's compilers etc. Scores with gcc or Microsofts compilers are much lower than Intel's.

BTW, I've graphed the c't G5 2.0 compiler info.

Ah, I must have been thinking of the top end P4EE. That matches up a bit better with what I'd expect. I've heard the same things about Intel's compiler (usually from GCC people explaining why GCC sacrifices speed for correctness).

[i_wolf]

well these numbers are impressive in my opinion. Remember the G5 is a new architecture from the ground up over previous PPC processors. It will take a while for compilers to improve etc... Even looking at the XL range which are very young I would imagine that these results would improve significantly over the next year as IBM fine tune their compiler more. Finally, SpecINT and SpecFPU are nice tests and all... but unfortunately not necessarily representative of real world performance in any way shape or form. For example Opteron scores are fairly similar to Pentium 4 scores with ICC 8. Most developers would agree that raw FPU and raw Integer performance of an Opteron in real world situations is significantly more powerful than Pentium 4 performance.
Another caveat to consider is that AFAIK the P4's FPU performance is significantly aided by autovectorised SSE2 code. The G5's is not. Now this in itself is problematic since SSE2 does not meet IEEE fpu specifications and can break code in real world situations where precision is lost.
Anyway its all fun and games.
Regards
i_wolf

[Angus_D]

Originally posted by ryaxnb:
People still use Fortran? Fortran stinks. Use the top 2 excellent replacements:
* REALBasic. Simple Fortran projects can easily be replaced with REALBasic programs.

HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA HAHAHA.

*wipes tears from his eyes*