Sigh, we hate it when the rest of the industry can't stay in character. Look, people, we're the soap opera, right? So when an allegedly technically astute German computer magazine starts publicly lambasting Apple for calling the Power Mac G4 a "supercomputer" because its testing using the industry standard SPEC benchmarks reveals that, for floating point operations, a 1 GHz G4 processor is a total dog on quaaludes compared to even a relatively ancient 1 GHz Pentium III, our role is to bask in the drama, wring our hands, and milk it for all it's worth and then some. Meanwhile, it's up to the "real" technical press to notice and report that said German magazine's methodology had more holes in it than a colander made out of Swiss cheese. In this manner, you, the viewer, would get both an entertaining dose of hysterical melodrama and a serious technical grounding that would let you enjoy our hysterics while realizing you'd have no reason to run screaming into traffic. See? It's all about demarcation.

Unfortunately, as faithful viewer Timothy Ritchey pointed out, it looks like everyone else out there forgot their lines. Even The Register, normally way better about this sort of thing, surprised us by running a headline of "Benchmarks demolish Apple speed boasts"-- sheesh. So now, much as we hate to do it, we're forced to backtrack and mention a couple of obvious reasons why c't's benchmarking results aren't really a reason to start drafting a suicide note.

First and foremost, c't itself admitted that the G4 should have mopped the floor with "the x86 FPU with its antiquated stack structure and eight registers only"-- so why, when the G4 was shown to be half as fast as the Pentium III, did the magazine just say "gee, we guess the G4's no supercomputer" and then saunter away, hands in pockets, whistling a jolly tune? Doesn't anyone think it's strange that they failed to mention that the SPEC2000 test, as compiled, utterly ignores the G4's Velocity Engine registers, which are what gives that chip its supercomputer-class, greater-than-gigaflop floating point performance? What c't did is tantamount to forcing you to write with your toes and then telling you that your handwriting sucks.

What's more, while the industry just loves SPEC benchmarks, faithful viewer Mark Davis reminds us that they've always been biased towards Intel processors, in part because the SPEC code just floods the chip with a constant stream of perfect instructions and let it work at peak efficiency, which is nothing like how real software is processed. As you may recall from Jon Rubinstein's "Megahertz Myth" spiel, Intel's recent chips take a speed hit from the recurring need to clear and refill those extra-long pipelines due to incorrect predictive branching-- it's that whole "pipeline tax" thing. With the SPEC test, there are no data dependency bubbles, and therefore no pipeline tax, so Intel's chips perform better than they would in actual battle conditions.

Apple itself obliquely refers to this problem on its G4 page: "Another aspect of speculative operation worth noting is that it is possible to create (for testing purposes) a contrived set of instructions that can make the processor guess correctly much more often than it would under real-world conditions. Thus a 'benchmark' with no relation to actual performance can be crafted to cleverly avoid the bubble problem and thus indicate unrealistically high performance." No one's mentioned by name, but all signs point to SPEC-- which was never meant to test real-world performance.

Good enough, people? So don't panic, and just remember that it's totally ludicrous to think that a 1 GHz G4 would perform only half as well as a 1 GHz Pentium III at real-world tasks. Chip-level benchmarks like SPEC mean nothing when it comes to getting your work done; for cryin' out Pete's sake, c't even disabled one of the G4 processors in that dual rig for the sake of measuring the performance of a single chip, but you're not going to turn one off to run Photoshop, right? Right. Now, back to the drama; we just hope that the other sites can remember their roles next time, because we're supposed to be rattling on about black turtlenecks and Reality Distortion Fields, consarn it, not analyzing benchmark data. Seriously, if we ever have to say "incorrect predictive branching" in an episode again, we're gonna have to bust some heads.