Last week's discussion of Apple's deteriorating competitive position with respect to the high end desktop got misunderstood by a lot of readers who thought I was comparing the G5 to the new "woodcrest" Xeons. I wasn't, but that kind of thing is always fun, so lets start by looking at the test configurations used by Eric Bangeman at arstechnica in a comparison between the new and old Mac Pro machines:
- Mac Pro
- 1GB PC2-5300 FB-DDR2
- NVIDIA GeForce 7300GT (256MB)
- 2 x 2.66GHz Xeon 5100 CPUs (dual-core)
- Power Macintosh G5
- 1GB PC3200 DDR SDRAM
- ATI Radeon 9800 Pro (256MB)
- 2 x 2.5GHz PowerPC 970FX CPUs (single-core)
Note that the arstechnica work includes a Core Duo iMac and reports on some other Mactel gear too, but these are out of the game from a performance perspective and therefore uninteresting here.
Now look closely at these two configurations and note the most important thing: that's two G5 cores at a cumulative 5.0 Ghz being compared to four "Woodcrest" cores at a cumulative 10.64Ghz. And yes, the new machine won on most tests -but generally by margins of 25% or less.
Specifically by 25% on ripping and 17% on Quicktime encoding - but it got drubbed by an average of 30% on Arstechnica's extremely simple Rosetta code tests, and established a high water mark of only 23% on the XBench CPU test component - terrible for a machine with more than twice the CPU cycles available.
Overall it won easily on the seven test XBench 1.2 series - scoring better than 4.5 times faster on the threading test and more than four times faster on the user interface tests to finish with 131.77 points overall to the G5's 71.1 - an 85% margin. Unfortunately these particular tests are highly suspect, heavily configuration dependent, and subject to measurement error. See, for example, this rather quick and dirty discussion - and see also this comparison and its supporting documents for some quite different XBench results reported by macintouch.com.
The four Xeons achieved their best credible relative advantage (23%) on the XBench CPU test with a score of 139.23 - or about one point per 76.4 Mhz. In contrast the dual G5 only reached 113.2 - but that's one point per 44.17mhz. In other words, pop a Xeon out of that box to equalise the number of cores at work, and you can expect the G5 to win by a resounding 63%!
Oh, and lest we forget the basic point of the original comparison: Microsoft is reportedly paying IBM $105 per unit for a three core, 3.2Ghz, G5 derived CPU - about half the $206 the big Asian assemblers are thought to be paying for the base 2.66Ghz "Woodcrest" Xeons they're putting in the new Macs.
So if we play "lets pretend" for a moment, we could imagine a dual core G5 based 2006 Mac Pro that came with a decent graphics controller and modern memory to retain Apple's 2004 competitive edge on the basis of both cost and performance dominance over its Wintel competitors. And that's the point: Apple's edge is gone: now it's just different bruises, same bananas.