and what is the plan for the high end workstation that needs Multi screen video? useing the TB bus can kill IO for non video needs.
People who may need multi screen more basic video out and don't want USB based video (did apple lock that out yet?)
The TB Bus isn't a particularly relevant constraint. The Display Port goes into the TB controller on DP bus.
The Utlra Studio has 6 TB ports. Even if have to 'sacrifice' four of those solely for DP v.14 DSC ouptput that still leaves two TBv4 ports for I/O.
"...
- Support for up to four Pro Display XDRs (6K resolution at 60Hz and over a billion colors) over USB-C and one 4K display (4K resolution at 60Hz and over a billion colors) over HDMI
..."
support.apple.com
If the new Mac Pro starts off with the Ultra Studio specs as the baseline and then
moves up from there in display output (e.g., two HMDI 2.1 ports and a DisplayPort 2.0 port ). How many folks are really going to be doomed being limited to 4 6K displays and 3 more 4-5K ones? Relatively not many.
The part that would be slacking in the M2/M3 Mac Pro SoC isn't number of video ports out, but upgrading the compliance with the current leading specification of the ports. Three points.
A. HDMI 2.1a (stop being capped on some part buying for the Apple TV 4K to provision HDMI )
B. DisplayPort 2.0 ( going to be provisioned on the AMD GPUs by end of the 2022. Apple should be in the mix also. If Mac Pro is pragmatically going to be a 2023 product not having DP v2 is slacking. )
C. Better DP Multiple stream support ( don't need to necessarily blow all of DPv2 bandwidth at just one humongous screen. If can daisy chain 4-5K displays then don't need as many output ports to drive multiple displays. )
It should be relatively easy to do if Apple isn't doing "exact twins of MBP 16" dies to do the Ultra and "double Ultra-class". Dial back on one of the dies Thunderbolt ( shouldn't need more than 6 (and could just do with 4 ). In the space for 2-4 TB controller complexes provision PCI-e v4 and/or DisplayPort. 90+% of the rest of the die is the same ( for twin , quad die packages).
A discrete card would allow to work around their slacking in standards compliance later , but why pay for a slacking card now? ( shipping with the 580 in late 2019 was kind of weak. Doing that again in 2022 would be setting a pattern of being weak. )
The trend lines in monitors is heading toward fewer , bigger , higher resolution monitors driven by one card more so than some kind of discrete card explosion to drive lots of medium scale resolution monitors.
high end workstation with big GPU compute needs?
That doesn't necessarily need to come with a display stack attached to it. A memory of the device driver "family" that plugs into PCI and provides just compute probably should be part of Apple's line up. There is already a M-series family class for PCI devices. That would just a be a more refined subclass that would send certain Metal compute jobs to.
As long as there is a single x16 PCI-e v4 (or better) slot then could couple a PCI-e enclosure break out box to a new Mac Pro and provision it with 2-3 triple wide slots and another 1,400W of power. Would be pulling in MI250 and A
There are some apps that need the display out on the same card that does the compute , but there are others that don't. ( for the latter the Infinity Fabric links point toward distributed loads being OK for some workloads. )
Right now apple sells the mac pro with an max gpu config of twin 64GB Video cards what is the plan to deal with work loads that really need an big ram pool just for video + an big system ram pool?
This niche just plays directly into the "Unified" , Uniform memory approach that Apple is on. In contrast to discrete GPU cards the integrated GPUs can use a huge fraction of the main RAM. If the Ultra class is going to 192 ( or 144) GB ( 16 X 12GB or 16 x 9GB with ECC ) and the quad die is going to 384 (or 288) GB then fitting 128GB "VRAM" workloads on those systems works just fine. That would be a net 100% increase in VRAM workload head room. If started off only using 16GB more per year in VRAM model then the 64 to 128 progression would be 4 years long.
If on a 4-5 year depreciation schedule that isn't going to be a huge issue to buy another one on that cycle.
These abnormally large VRAM footprints is exactly the corner case that Apple is going to point to showing that their approach is works extremely well. It is not a huge upside for the discrete GPU cards.