Oh, I agree... It is WAY more likely that Apple will continue to under-provision the PCIe lanes bandwidth. Probably in the ballpark of the current 64 lanes the Mac Pro (2019) has.
Probably not 64 lanes out of the SoC package itself.
There is a design tradeoff between total aggregate off-die RAM bandwidth and total aggregate PCI-e bandwidth .
There are a pile of unknowns:
- I'm still guessing that Apple won't utilize any of the M* dies for PCIe slots, they'll add an independent bridge chip to the UltraFusion fabric? But that is just pulled out of some dark place, and is probably wrong.
Just because the PCI-e controller goes on another chiplet doesn't really mean going to get a huge block allocated it.
If it is just a 'shim' chiplet between two "Max sized" dies in an enhanced "Ultra" configuration , then there will be heavy constraints on just how big that 'shim' can separate those two dies and not introduce any substantive NUMA problems.
It is more likely they are going to try to match the aggregate bandwidth of the MP 2019 , not the lane count. More so don't 'backslide' on general I/O lane count than get into a lane count number size pissing match.
If Apple folds the default display GPU into the SoC and the Thunderbolt provisioned PCI-e bandwidth into the SoC than there is no "bandwidth" loss there even if slide backwards on general slots. So don't need "slot 1" and the two MPX connector PCI-e contributions to 'keep up' with the MP 2019 in useful bandwidth provided.
- Whether Apple will still utilizes bridge chips on the motherboard like on the Mac Pro (2019)?
Errr .... extremely unknown really? The MP 2009-2012 had a on mainborad PCI-e switch to provision the two x4 slots. The MP 2013 .. had a mainboard PCI-e switch . The MP 2019 had a mainboard PCI-e switch. ( didn't take time to look up schematics of 2006-2008 MP but good chance they had a switch too. )
For about last decade, it is closer to the peculiar case of when the Mac Pro has not had a PCI-e switch.
Yes it is not 100% certain, but Apple's track record here is pretty solid. If more than 2 slots, the very probably getting one.
- Whether the M* Ultra and M* Quadra Mac Pros would have the same PCIe configuration or same PCIe bandwidth? (it could if they use an on package bridge chip)?
Slot count is very likely the same because probably don't want major differences in main motherboard costs across the two variations. The mainlogic board should have a high degree of overlap. ( Apple went through large efforts to make single and double package MP 2009-2012 look the same. Unit number likely not any higher here at all so have same economies of scale problems.)
The Mac Studio Max and Ultra boards mostly the same. (more gap space around the package on the Max version due to smaller package. )
- How is the 4x M* Max scaling working with UltraFusion (given that it sounds like Apple hasn't gotten it shippable, I expect this is not going well), and will any nasty bottlenecks show up?
UltraFusion isn't the core problematical issues as much as the chiplet die design itself. The laptop optimized 'Max' die isn't really a good scalable chiplet. Whether the Mac Pro is successful or not is more about whether Apple can get past the fixation of trying to use a chip primarily designed to be monothic as a chiplet. Pick one or there other, but the 'half effort chiplet' is not likely going to succeed at scaling up. To scale need a more deliberate functional deomposition into multiple dies. A good real chiplet would fail if only used just one single, solitary die.
Similarly the Mac Studio doesn't necessarily need a laptop optimized Max die.
www.macworld.com
