Of course Nvidia chose PCIe 4.0...
... and efficiency, scalability, performance, more PCIe lines...
Apple chose PCIe 3.0, vintage technology and PCIe switches.
Apple chose Thunderbolt ... ( PCIe switch). Yeah they did that a long time ago. So did the rest of the surviving Mac line up. The Mac pro wasn't going to jump off and wildly deviate from that.
Apple also extremely likely wasn't going to pick two CPU sockets so a single EPYC wouldn't have "solved" the switch problem.
Nvidia took EPYC also because it is cheaper. If get to what Nvidia is pitching in the roll out demo ( about timestamp 17:00 or so here )
https://www.nvidia.com/en-us/gtc/keynote/?nvid=nv-int-cwmfg-12643&video=6
They are throwing their $199K server at getting rid of many other Xeon SP and AMD EYPC servers. That is mainly by shifting much of the workload onto the GPUs. The AMD CPU there is more so a tangential issue of what runs the base OS to dispatch work from The primary point of the pitch is that stop buying as many of those x86 cores and spend more on Nvidia cores.
AMD's priority on Threadripper ( Single GPU workstation subsector target by AMD) was about last in the generational roll out . That is was very bad for the Mac Pro since it was already chronically late (last update in 2013 and it was not 2017-18 when earnestly started to work again.) . AMD's tepid support of Thunderbolt and the kind of "we do but not fully" to Threadripper as a targeted workstation processor.
"... on AMD's Ryzen and Threadripper platforms. ECC lies in this region of 'it kind of works' but isn't validated. There is what's called 'unofficial support', which is different to 'official qualification'. Technically, none of the Ryzen and Threadripper CPUs are 'officially qualified' for ECC, however most of them (if not all) will exhibit unofficial support. This means that it might work, but AMD won't give you assistance for it. There are two caveats to this: ..."
www.anandtech.com
Plus some GPU delivery slides that AMD had ( roadmaps to more energy efficient ones for a follow on Mac Pro 2013 ) .
Apple wasn't making a selecdtion solely based on PCI-e version number. The Vega II modules have PCI-e v4 present, but not really used. That "PCI-e" version is not the primary reason why they got picked.
P.S. Intel has problems for general box where just want to throw as many x86 cores in a single box as can for a not too crazy high budget. But doing a Mac means need to dot some i's and cross some t's on other dimensions that just a minimal container for x86 cores. Intel has major issues, but they have also jumped throw lots of hoops to do what Apple asked them to do.
P.P.S. Also much of Intel's additions to Xeon SP 2nd (and upcoming) generations are exactly for inference which is exactly one of the jobs Nvidia wants to "take". AMD has no inference accelerator so therefore no probability of someone getting 'distracted' with trying to push some hyper low latency inference onto the x86 cores and away from the custom Nvidia format/libraries ecosystem. AMD is being selected because not as competitive in the space Nvidia is most interested in here.
