Originally posted by wizard69
Intel is to be feared only as long as they are well managed. There is a reasonable question as to that being the case anymore. They still (after a couple of years) have not responded to AMD's challenge.
Intel's fab capability is very forgiving to management and design mistakes. I don't see that going away. AMD better be scared because fab capability will win the dual-core wars, and Intel is going all dual-core from server to laptop afterall.
I could be completely wrong here, but elastic bus just doesn't impress me as something that a manufacture would want to support across all of its lines.
Bah. This is Apple we're talking about. They design hardware based on considerations more important to them than hardware performance. Laptop G5? Sure, Apple just needs to make it 1.5" thick. Cheap desktop G5? Use 1.6 970 CPUs with 1/3 FSB. They didn't even do the little effort to support PC4000 DDR SDRAM or better for the 2nd gen Power Mac G5s. That could have squeezed in another 5% performance improvement, maybe more for FPU.
Well not exactly. They could still keep e-bus and simply handle memory transfers over the other bus. This owuld require modest changes to the north bridge. Better yet have memmory hanging off both buses with the block managed by the controller (northbridge) kept as share memory for the second processor module.
I think this is too much of a kluge. If they are going to implement a non-uniform memory architecture they should at least arbitrate inter-processor traffic with a switched fabric. I think that is a major change in core logic and board design. On top of that, not quite sure how much of a performance improvement that would be compared to what they have now.
I just don't see the memory being so far from the processor as being a good thing at all for Apple. Especially when the get trounced by implementations that take care of that issue.
Memory architecture is about choices, and I don't see the 970 architecture as limiting. If latency is hurting 970 performance, IBM can increase L2 to 1 or 2 MB and add a backside L3. It did wonders for Power5. Also, the elastic bus is what allows Apple to place the memory so far from the processor.
On bandwidth, the elastic bus does extremely well. Double its width to 64 bit would give it even more bandwidth.
IBM and its 970 series is another joke all together. It is basically an OK processor that Apple can't use in half its hardware. Not exaclty inspiring. Further one could make the argument that for some machines like the iMac the 970 was shoe horned in and might not be the best of choices. So the question is can IBM deliver what Apple and the rest fo the industry needs.
The 970 holds its own. It's about equivalent to the Athlon 64 on a per MHz basis. Where it is hurting is integer performance and memory performance. If they can give it more integer resources, more L2 cache, a backside L3, more main memory performance and dual-core, it will be fine for the next year.
I think in Cell we can see the possibilites for future Apple processors. Hopefully soon we will see if IBM can transform the technology in Cell into hardware with broad appeal.
Not going to comment on Cell until more info is out.
As to Apple doing its own in house design, well one only has to mention SUN. For processors to be viable they have to have broad apppeal plus a bit of committed marketing muscle. I'm not sure that that would be a good long term stradege. IBM did well with the 60x series but we have seen less take up with the 970 outside of Apple. This lack of take up appears to be the result of the processor being to specific to Apples needs. Or maybe IBM simply isn't marketing it effectively due to a possible short life span.
Being fabless means you can take your design and go to different manufacturers. Apple already does it for its core logic chips, the CPU isn't that big of a difference to me. Ie, if IBM doesn't meet their needs, they can go to Intel, UMC, TSMC, TI or whoever has the best process. Freescale is out of the picture for high performance processors because that are about a year behind in fab technology. IBM and Intel have been shipping 90nm chips for about a year now and Freescale is yet to ship. They will only fall farther behind.