MacEdition Logo
 

So Sayeth Soup: The things I tell you will not be wrong.

By SoupIsGood Food, (soup@macedition.com), June 29, 2003

It took them far too long to come around, but once they did, IBM really delivered the goods. The G5 is a world class workstation disguised as a mild-mannered desktop PC.

Feedback Farm

Have something to say about this article? Let us know below and your post might be the Post of the Month! Please read our Official Rules and Sponsor List.

Forums

Want to dig even deeper? Post to the new MacEdition Forums (beta)!

Well, at $2000 for the low end, not that mild mannered. But to configure an equivalent Dell to match Apple’s stripped-down model feature-for-feature will set you back just about $2200, and still won’t give you Bluetooth, Firewire 800 or 802.11g. Not too shabby, and the faint of pocketbook can save two hundred bucks on their G5 by opting out of the SuperDrive.

At the high end of computing, there is nothing that can compare to the G5. It’s got twenty-thousand-dollar Sun and SGI workstations outgunned at a fraction of the price. This is a full-bore, 64-bit Unix workstation, and it makes mincemeat out of anything with two processor cores or less.

You may have heard different. You may have heard that Apple fudged the numbers, cheated on the benchmarks, that the Pentium is faster after all, and that Steve Jobs lied to us. I’m here to tell you it’s an ugly urban legend, repeated by an all-too-predictable “geek chorus” of fools and liars, in a blind panic over the fact that Apple ships faster boxes than Alienware, repeated by those who have an agenda against Apple and an investment in slow, legacy x86 processors.

The thought of a Mac with more horsepower than an Alpha workstation was just too much to bear for one sad little troll, so he deliberately misrepresented the VeriTest white paper, accused Apple of cheating, and managed to get this nasty little propaganda job posted to the front page of Slashdot, computer geek central, where he managed to find many sympathetic ears. We can forgive “spl”, as he’s not a professional journalist, and savaging him for being too stupid to understand the white paper and too lazy to ask anyone with a clue is like kicking a puppy for peeing on the rug; but now it seems that Mark Hachman at ExtremeTech all but plagiarized that troll-cum-article by “spl.” He pretends that his bold discovery was all his own, independently verified by “spl”...like the rest of the universe doesn’t read Slashdot and wasn’t aware of the issue the day before his odious little hatchet job saw the light of day. Who does he think he’s fooling? Since he’s leeched onto spl’s idiocy, and magnified it in a widely read online publication, Mark gets to bear the brunt of my wrath for being too stupid to understand the white paper, and too lazy to ask anyone with a clue. Pretty piss-poor behavior for a professional journalist. He’s obviously from the Jayson Blair school of reporting, only without the writing ability.

First off, turning off hyperthreading increases Pentium performance on SPEC benchamarks, especially the SPECrate, according to Dell. Mark also gripes that Apple turned off software pre-fetch for the G5, maximizing the number of streams available to the processor for a single-threaded application. I wonder if he realizes that hyperthreading is a method of software pre-fetch built into the Pentium 4, and that turning it off maximizes the number of pre-fetch streams for a single threaded task? Evidently not. So, in essence, he’s complaining that Apple turned off the Pentium 4’s hyperthreading to jimmy the score, and then turned off the G5’s hyperthreading equivalent to make it extra unfair! Gosh, he’s bright.

There’s also some confusion about an extra library that Apple compiled against. Well, according to Apple, lstmalloc is going to be in the version of MacOS X shipping with the G5, and part of the performance tuning suite of developer tools Apple is planning for Panther. It’s not a one-off spoiler used to skew the benchmark, and probably required for the G5 to even run on the ugly hack pre-release version of MacOS X (rumored to be known as “Smeagol” in house.) You’d think a professional journalist would be able to pick up the phone and ask Apple for confirmation or clarification...

The contest was against real-world systems, running real-world software and operating systems...basically, only the windowing and printing systems on each machine were disabled. As such, the choice of the GCC compiler was a fair one. Yes, they could have gotten better performance from the Intel compiler, which is designed specifically to skew performance benchmarks; but then Apple would get to use IBM’s compiler, which does the same.

According to this (pdf) article, IBM was able to produce SPEC scores better than those published by Intel for Itanium and the P4, on single CPU PPC970 systems running at a lower clock rate than the new G5s. They probably used much of the same methodology Dell used for their “official” scores: stripped down, custom tweaked OS and compilers.

The faster-clocked 970 chips in the G5 would simply have destroyed the Intel chips if VeriTest played the compilers against each other (I assume the G5, at least initially, is overclocked like the G4 systems – it would account for the discrepancies in the heat dissipation between IBM’s 970 specs and Apple’s G5 specs). What’s more, he seems to be upset that Apple’s able to provide you with performance tuning tools that help you choose the optimal compiler flags...there’s nothing equivalent on the PC side, especially in the Unix/Linux arena. Not only is it fair play, as no changes to the benchmark code itself were made, and hand-tweaks were applied on the PC side, it’s a clear indication of why development on the Mac is a superior experience for those with a need for speed.

And finally, SSE2, the Pentium’s sad and sorry little answer to the VMX/Altivec/Velocity Engine SIMD unit on the PowerPC, was enabled for the test. In GCC 3.3, the SSE flag covers both SSE and SSE2, and turning it on increases the Integer scores dramatically.

Which is probably cheating.

Apple didn’t compile with Altivec turned on... that would require the altivec and force_cpusubtype_all flags to be set... presumably because it wouldn’t help their scores since SPEC isn’t coded with AltiVec extensions. This means that either SSE2 was engineered to artificially inflate SPECint scores, or that SPECint was engineered to let SSE2 artificially inflate Intel’s scores. Either way, it smells really, really fishy. But we’ll give ’em the benefit of the doubt, and the edge in raw integer, as (contrary to “spl’s”protestations otherwise) almost nothing is dependent on raw integer performance these days. It’s all floating point math and SIMD vector performance... two areas where the G5 is untouchable.

So not only were the SPEC benchmarks fair, they seem to have given a little bit of an edge to the Pentium. Not that it helped.

Still, the SPEC benchmarks are universally regarded as useless at best and misleading at worst by anyone seriously interested in comparing different computing platforms. They don’t measure real-world performance, nor are they a good indicator of total system performance. Intel always seems to have scores near the top, but you’d never know it comparing a Dell Dimension to an SGI Fuel doing any real work with real applications.

In real-world application bake-offs where VMX could be put into play, the performance of the G5 is breathtaking. Its emphasis on high-powered muscle car mayhem is industry leading; 1GHz frontside bus, PCI-X slots for incredible peripheral throughput, Serial ATA, which provides SCSI speed at an ATA price point, Firewire 800, USB 2.0, 802.11g, gigabit ethernet – all built-in rather than tacked on. This box is designed to move bits quicker than anything the market has ever seen – a total package, and a fine opportunity to thumb our noses at those who so richly deserve it. It’s a shame about the video card options, but I’ve complained about that before.

E-mail this story to a friend

Talkback on this story!

Cannot connect to the database.
Please contact the administrator.