PowerPC G5 on a AmigaOne EX systems...

[Floid]

Hey, who wouldn't?  Thing is, the processor bus is an entirely different animal from the MPX(?) bus used on most G3s/G4s, and even with huge caches around the CPU, the math doesn't work out for most memory accesses.

Remember, if you can't fetch or write to RAM, you're spinning cycles doing nothing, and that doesn't equate to better performance.  Prefetching and write buffering can help, but PC133 just doesn't cut it too well for a chip that can actually bust out work at GHz rates, and by the time you work out the bridge chips necessary to go from MPX<->"Elastic bus(?)" [What's Apple's name for the new one, anyway?], you've created something that probably costs the same as a 'straight' quad-CPU box from IBM.

Remember, one of the things that made the original Amiga so great was that the memory ran *faster* than the 68000; the custom chips could sneak in on opposite clocks, and the CPU rarely had to take any wait states.  Contrast this to other designs of the time, where DMA was rare, and the CPU actually had to direct the fetching of every piece of data, and the sending of it to RAM, framebuffer or disk.

[Floid]

BlackMonk wrote:

I think it is API, Apple Processor Interconnect.  Designed by Apple and made/implemented by IBM.

It's something like that, anyway.

Don't forget that the PPC970 supports many multiples.  So one could possibly run a 2.0 GHz PPC970 with, what, a 500 MHz FSB?  I forget the exact multiples but a motherboard manufacturer might be able to take some "shortcuts" if the high FSB is an issue.

(FSB meaning the p2p CPU connections to the northbridge, not to include the memory bus or PCI bus, etc.)

Sounds right; I think they were abbreviating it ApplePI, now that you remind me.  I forgot about that little announcement, and for some stupid reason I've been assuming this was a serialized design (a-la HyperTransport)...  The Ars article says different, though: it's DDR, and "Since the bus composed of two unidirectional channels, each 32 bits wide," ... so maybe it's within the realm of feasibility, if someone (IBM?) coughs up an MPX bridge cheap.

The thing is, though, everyone knows Motorola's lost the desktop race for now, and appear almost happy to be out of it, what with potentially putting the processor unit up for sale.  IBM's going to have to balance the desire to support their own legacy customers ("Hey guys, have we got an upgrade for you!") -- including Mai, Marvell, Eyetech, Genesi, and all the other little and not-so-little guys that form the broader market for PPC -- with the desire to push forward with the competetive chip, avoid paying more 'cooperation fees' to the Motorolan black-hole for MPX -- (wonder how much AltiVec cost them?), and move forward  the architecture that actually *is* competetive, shipping low-clock, low-bus-speed versions to cover the fat end of the embedded market, which could always use more power anyway.  I'm going to bet the "Appleness" of the interconnect is only an encumbrance for northbridge manufacturers, so that aspect should be encapsulated away in epoxy or ceramic by the time Mai or Marvell ship a supporting product.  (One has to doubt whether Apple really *likes* designing all their own bridges, anyway; I'd wager they'd rather be designing added value to hang off them, a-la Firewire and the like.  Of course, that'd force them to *think,* and Jobs is at the helm, so you never know.)

IBM seems to think the 750GX is good enough for now, and they have the resources to run off an entire new *CPU* with a different interconnect if they want.  

But... why is it unlikely to see them bother?  Let's face it; board manufacture itself is cheap.  The Inquirer says it, and whether or not you want to trust them, you can't deny the surprisingly healthy state of the x86 board market, even with new chipsets (and thus new designs) every few weeks, the global economy somewhat in the toilet, the whole SARS thing, and little compelling in the CPU area until the Opteron launched.  What's expensive are the initial investments - setting up the businesses around the boards, ensuring the equipment is up to date (able to operate with enough precision for ever-shrinking components) - R&D costs - and, in the case of outsourcing small runs, paying for all the setup times (and *their* equipment costs) when the company could be running off commodity DIMMs, thermostats, or something else that doesn't change much.  All the big names from Taiwan have already gone through this, probably getting a boost from the .com boom; our favorites in the game have missed that chance, but as long as they live to get over the hump, and pay down past mistakes (each company's had one, after all) there're clear skies, bigger volumes, lower prices, and a  'sustainable' market to look forward to.

Having to redesign for 970 will cause some grumbling, especially as everyone's taken a hit for one reason or another, but meanwhile, the GX doesn't look like a bad little chip, and it's the northbridge companies' jobs to produce implementable solutions, and the board companies' jobs to design boards.  Unless Mai or Marvell screw up, I'd think we could avoid the underclocked buses (hmm, wonder if Apple can impose price-hikes for bus licenses based on MHz?), and stick to playing in the same leagues as the big boys.