New G4 roadmaps promise Apple harvest
RapidIO, revved bus options for iMac2
Exclusive While Apple watchers look for first sightings of the G5 chip, Motorola is taking the G4 into fertile territory.
By this time next year three as yet unannounced versions of the G4 will dramatically boost internal bandwidth, support switched fabric interconnects, and will see the processor talk to memory at full bus speed, according to disclosures from Motorola sources.
One caveat: as far as we know Apple hasn't committed to adopt these new G4 variants. Now it's got a portable operating system it can choose something really weird and exotic if it desires; but common sense and recent precedent suggest that these two processors will form the mainstay of Apple’s iMac2 and low-end professional lines.
According to Motorola sources, a tweaked version of the Apollo 7450 G4, the 7470, will be ready for volume production shortly after the end of Q2, in time for a summer ramp. The 7470 will be manufactured on a 0.13 micron process, allowing for a smaller die size with room for 512K of L2 cache, and support up to 4MB of DDR-SDRAM L3 cache.
The 7470 supports a modified bus protocol, MPX+, which supports double data transfer and which should effectively run at 266Mhz according to sources.
The MPX+ bus retains MPX's 36-bit addressing lines, and is described as an interim measure. Don't expect dramatic leaps in SMP scaling - two will remain the sweet spot. The 7470 should scale to 1.5GHz. In parallel development, Motorola is priming a cut-down 7470, labelled the 7460, which doesn't support L3 cache.
But the third addition to the G4 family, the 7500 - slated for volume production by this time next year, should steal some headlines. Like the 7470, the 7500 will be built to a 0.13 micron process, and will feature a 14 stage pipeline (11 integer).
Having boasted one of the shortest (and most efficient) pipelines in the business, with the G4's original four-stage unit, Motorola has been increasing the pipeline depth, which allows it to deploy higher clock frequencies. There's a whole science to optimizing processors to minimize the branch jump table penalty. This shouldn't detain us for too long but a crude summary is that a deep pipeline allows the processor to second guess the subsequent instructions, at the cost of clearing the pipeline when it gets the wrong answer. If a processor has a higher clock frequency, it can clear the pipeline quicker when a mis-prediction is flagged.
More noteworthy is that the 7500 will be the first desktop processor to employ the RapidIO architecture to communicate with the system at 500Mhz.
RapidIO is a switched fabric interconnect which mirrors the parallel Infiniband initiative: the former is endorsed by the embedded industry, the latter by big iron system vendors, so the two don't really overlap. But the move to a standard switch architecture away has been a long time coming: high-end systems use their own proprietary switches (in mainframe terminology, a "crossbar switch") or a combination of a bus and a switch.
In practical terms it will permit the the memory controller to be housed on the die, communicating at full clock speed. So potentially, there's no need for L3 on-die cache. With DDR-SDRAM at 266Mhz and heading for 333Mhz, this should result in a comfortable improvement in throughput.
Our source suggests that Apple will have "little choice" but to adopt RapidIO, given that the bulk of Motorola and IBM PPC volume is in the embedded business. But there's much to be gained too, as RapidIO becomes the internal wiring for knitting together the various bus protocols a desktop PC needs to support: ATA, USB, FireWire and so on.
With Intel's latest Northwood systems employing a 400Mhz front side bus (up to 533Mhz by the end of the year), Apple loyalists should take some cheer from these developments. System performance owes a great deal to the internal bandwidth, and running a high frequency processor on a slow frequency bus has never been considered optimal.
More importantly for Apple, it allows the company plenty of potential for system improvements to the consumer iMac range, and far more scope for differentiating between the pro and consumer lines. Right now, they look awfully similar. ®
Sponsored: Benefits from the lessons learned in HPC