How fast was that Mac again?

Lovely as the new Intel-based Macs are (and tempted as I am by the iMac), those performance figures – 2–5x the speed of current G5’s – always looked a little suspicious. In “The Mac performance shell game” Paul Thurrott links to the following from Infoworld:

Apple used multiprocessor benchmarks to skew the performance advantage that its Intel-based machines enjoy compared to single-core PowerPC G4 and G5. Apple used the industry-standard SPEC suite components SPECint2000 and SPECfp2000, but here’s the catch: Apple used SPECint_rate2000 and SPECfp_rate2000. Both tests spawn multiple parallel benchmark processes and are specifically intended for comparing multiprocessor systems. Single CPU, or single-core machines do positively lousy on SPEC*_rate2000 tests. That’s predictable and universally understood. Add a second CPU or a second core and, as you would expect, SPEC*_rate2000 performance on any multiprocessor-optimized test skyrockets compared to a single-processor box.

Apple uses SPEC*_rate2000 tests as a foundation for claims that Intel-based Macs outperform PowerPC G4 and G5 by a factor of 2 to 5. Well, yeah. A dual-core anything outperforms a single-core anything else by a factor of 2 to 5 in benchmark tests that make use of multiple threads or processes, tests crafted specifically for the purpose of stressing SMP-based systems. It’s murky marketing, and the sad part is that Apple didn’t have to resort to it to make Apple’s PowerPC-to-Intel switch look like a smart one. Mac users have no choice, and users also know more or less what to expect performance-wise.

As Paul goes on to point out:

What should have tipped everyone off is that the Core Duo chip is aimed at notebook computers, not desktops. In the case of the new iMac, especially, it’s hard to understand how a notebook chip could do so well against a relatively high-end desktop chip, dual core or no. I’d be happy if the new iMac was simply as good as the iMac G5, from a performance perspective.

And, of course, that’s running Intel-native code, not PowerPC emulation done in Rosetta. Personally, I suspect that we’ll find once the first real-world benchmarks come out that it’s actually slower when running current PowerPC applications, although not by much.