Reality Check
The Power5 is expected in its first computer (Armada, a 64-CPU SMP system) in the first half of 2004. As such quarter/half targets go, that generally means we'll see it around June, 2004. With the Power4, we are seeing about a 2-year delay between the high-end server variant and the desktop derivative. That would put the "980" (or whatever) in June, 2006. That's the "upper limit", as unlike the Power4, the Power5 is reportedly being designed with high-end, low-end, and desktop variants (not derivatives) in mind. Still, I'd be surprised to see it affordable and in a desktop Mac less than a year after its Armada debut (June 2005).
Also, performance details are sketchy. "4 times as fast" can mean a lot of things. First, the Power5 reportedly supports simultaneous multithreading, which is related to Intel's "hyperthreading", but non-marketing-infested and IBM claims its implementation is an order of magnitude more efficient. IBM claims one processor truly acts as two (a 100% speedup) instead of as 1.2 as in Intel's implementation. If running "deep machine code" at four times the current Power4, I suspect that the Power5 is getting half of this gain (2x faster) just via simultaneous multithreading on highly-tuned code. I for one don't fully believe the claims of 100% performance improvement with a good SMT implementation, and expect that that component of the "speedup" will be closer to 1.5-1.75x instead of 2x.
Second, in pure performance/core terms, the 970 is getting SPEC numbers on a single core that the Power4 requires two cores to get as is. A dual-core 970 would theoretically then be about 1.5-1.75x as fast as an existing Power4. But, that having been said, we won't be seeing (as far as I can tell) a dual-core 980 either; the 980 will be closer to equal to 1/1.5 (2/3 or 67%) of the performance of the Power5 with dual cores, aside from any other compromises that might be introduced along the way for cost and production quantity reasons.
Thus, in "real world" terms, I don't expect 970->980 performance increase to be more than around 2x to 3x on the high end (1.5x-2x on the low end), and that over the space of two (or possibly three) years. Roughly consistent with Moore's Law.
So, guys, don't get your panties in a tizzy. The Wintel world occasionally gets ahead of itself like this as well, drooling over Itanium-3 numbers that likely will not end up nearly as impressive when the processor is out. But they're more used to it than us. It's good to finally have a server processor line related to our Macs, and certainly will be providing more of a push for performance than Moto was ever able to pull together. But don't expect true 4x improvement over the 970 a year from now!
