G5 roadmap revealed
Apple's G5 chip family will reach 25GHz in less than ten years' time, according to an anonymous poster over at MacRumours.com. The same correspondent also posts what purports to be a roadmap for the G5.
The chip family's future goes something like this. IBM will follow the current G5, the PowerPC 970 with the 980, based on the Power 5. The 980, says the source, will be chip that delivers Steve Jobs' promise of a 3GHz Power Mac this time next year. The part will kick off at 2.6-3GHz and max out at 4.5-5GHz. By comparison, the 970 will peak at 2.8GHz.
The 980 will include Simultaneous Multi-threading (SMT), which the source claims provides a 30 per cent performance gain over Intel's implementation SMT, called HyperThreading. The 980 will also feature IBM's eLiza fault-tolerance technology. Greater parallelism will be provided by extra computational units: one more AltiVec, two more floating point units, and two more integer units.
IBM has said publicly that the Power 5 chip, codenamed 'Squadron', will be more of a mid-range chip than the Power 4, but it will scale across IBM's server product range from blades up to the top of its high-end machines. And Power 5 will replace Power 4. That suggest Power 5 will be more of a family of chips than the Power 4, with a dual-core unit aimed at multi-chip modules at the top, with progressively stripped down versions as you go down the line.
For example, earlier this year, IBM fellow and chief architect Ravi Arimilli said: "If Opteron takes off, we will be there, and we will OEM it. I think it's a pretty good story for 2003, but it won't be the story in 2004 when Power 5 arrives."
If that doesn't imply a low-end server/workstation version of Power 5, we don't known what does. And the 980 - if that will be its designation - is the most likely candidate for the low-end incarnation of the Power 5.
IBM's public pronouncements on Power 5 have noted support for its eLiza autonomous computing initiative, which is designed to equip servers with the software and hardware needed to keep running when elements of the system go awry. If the 980 - we'll call it that for now - is part of a broader Power 5 family, eLiza support is logical. The Power 5 is expected to have some chip-level error recovery circuitry, and some folk have speculated that means better branch prediction.
SMT, meanwhile, has long been on the Power 5 roadmap, but we'd take the performance claims with a grain of salt. SMT only improves certain applications, and they tend to be the kind run by servers. While IBM may want that in the 980, for better blade performance, it's questionable how valuable it will be for Mac users.
The MacRumours.com source doesn't mention Power 5's dynamic power dissipation, which causes the chip to automatically power up or down according to its workload, nor Fast Path, which implements certain software tasks, such as networking stacks and virtual memory, in hardware on the chip. As we wrote last year, Fast Path sounds awfully like a vector engine, and as we've seen with the 970, IBM has the will to implement AltiVec.
The dynamic power system may be what allows Apple to get the G5 into notebooks, but a 970 die-shrink to 90nm seems most likely. Neither course is expected for some time, though the anonymous source pegs January or February 2004 for the first G5 PowerBooks. Power 5 prototypes have been fabbed at 130nm (0.13 micron), but the source claims the 980 will be fabbed at 90nm.
Beyond that, he or she claims, comes the sequence-following 990, to be fabbed at 65nm in 2005/2006 and running from 6GHz initially and rising to 10GHz. The PowerPC 9900 will take the platform to 45nm in 2007/2008. It will run at 9-10GHz and ramp up over time to 20-25GHz in 2010-2011.
By which time, of course, engineers really will be rubbing up against the physical limits of silicon chip circuitry, with transistors so narrow there's barely enough electronics to support a current. But IBM has to get to 65nm first, something it expects to do in 2005/2006, according to what is has said about the goals of its joint development partnership with AMD. That's the same timeframe the Power 6 is expected to appear within. A Power 6-derived version of the 970 is logical, but MacRumours.com's source could equally just be putting two and two together.
Ditto most of what else he or she has claimed, a lot of which could simply be derived from what IBM has said about Power 5, and the logical assumption that it will take the 970 and its successors in the same direction.
The discussion of the 980's architecture likewise could simply be an exercise of the 'think of a number... now double it' variety. We'd suggest caution before accepting all this as gospel, as does MacRumours.com. ®