So after countless mind-boggling threads over the last two days, I’ve realized most people have no clue how important underlying technologies work, and accordingly, I forgot to mention quite a bit in my critique of the new Mac Pro, so here’s part two!
To alleviate fears of zero expansion options, Apple is going to ship the new Mac Pro with six Thunderbolt 2.0 ports, so let’s recap a little on thunderbolt. Thunderbolt is basically an external pcie interconnect, and it runs at pcie 2.0 x4 link-speeds. Thunderbolt controllers can present one or two ports, depending on the controller used, though there’s little to no technical data on how the controller is “doubling” bandwidth to present two ports, my best guess is built-in, intel designed, pcie multiplexing akin to popular PLX 8747 chips. Thunderbolt 2.0 and 1.0 run at the same pcie link speeds, though 2.0 combines previously separate data and display channels into a single bi-directional channel to effectively double bandwidth from 10Gbps (or ~1GB/second) to 20Gbps (~2GB/second.) You can expect that 2.0 will require all new cables, and all new peripherals, else the ports will almost certainly drop back down to 1.0 compatibility modes, or simply fail to work with 1.0 peripherals at all.
Now, an Intel socket 2011 CPU exposes 40 pcie 3.0 lanes to the host system. We have to assume that apple is connecting both AMD GPU’s via dedicated pcie 3.0 x16 links (I sure hope they are at least! EDIT: this is an important point to clarify, as many “enthusiast” or “gaming” computers will run dual-GPU’s at pcie 3.0 x8/x8 links, and have no issues with gaming or daily use. However research has shown that, specifically with AMD GCN/79XX/W9000 at least, for GPGPU/Compute heavy workloads, dropping to x8 links lowers performance significantly. This would be flat out unacceptable for a “Pro” or “Workstation” computer.) And that leaves us with 8 pcie 3.0 lanes to distribute to the rest of the system. Based on the reported “1250MB/second” speeds of the pcie SSD apple is using, we have to assume that 2 pcie lanes are being routed for the SSD (a single pcie 3.0 lane can’t break 1000MB/second.) That leaves 6 pcie lanes leftover for thunderbolt controllers (I’m ignoring anything that might be exposed via the X79/C606 PCH for three reasons; first, PCH lanes operate at pcie 2.0 or 1.0 speeds; second, it seems clear Apple is leaving the PCH mostly intact with dual 1Gbe NIC’s, USB 3.0, on-board audio, etc. and simply disabling the SATA/SAS controllers; and third, using all native pcie 3.0 lanes, directly connected to the CPU is the best possible scenario for performance.)
While there’s no technical data available for falcon-ridge (thunderbolt 2.0) we can assume that intel has both single-port and dual-port variants like they did with previous generations. So now we need to try and figure out if apple is using 1-port or 2-port controllers for the new Mac Pro, and it’s immediately clear that apple is most likely using dual-port controllers. This is due to the fact that we have 6 pcie 3.0 lanes (equivalent to 12 pcie 2.0 lanes) leftover, exactly the number required to run three dual-port controllers (based on previous generation requirements.) So we have to expect that all six thunderbolt ports will have added latency from whatever internal multiplexing intel does on dual-port controllers.
Ok, all of that out of the way, let’s actually get back to some of Apple’s new reality-distortion-field marketing! Apple claims that the Mac Pro can support three 4K displays at the same time. Ok, well, I’m going to assume they mean Ultra HD (3840 x 2160) resolution and not Cinema 4K (4096 x 2160) even still, the bandwidth required to drive a single Ultra HD display, at only 60Hz, is ~1.5GB/second. In other words, if you want to use a 4K monitor with the new Mac Pro, you won’t want to daisy chain any device between the computer and the display, else you might bottleneck the display of the bandwidth it needs, especially since there’s already added latency from the dual-port controller.
EDIT: To add to the above, this assumes that the “4k” displays are all thunderbolt 2.0 compliant, since thunderbolt 1.0 lacks the bandwidth necessary to push 4k @ 60hz.
And now if we take that “three 4K displays” claim a bit further, and we see that of the six thunderbolt ports, we now have only three left for actual expansion. So far this seems reasonable enough, though it’s questionable if even dual FirePro W9000 series GPU’s could truly drive three 4K displays without problems, especially while trying to perform GPU heavy workloads concurrently.
Next assume that you need a thunderbolt video capture unit, the best I can find is the Blackmagic Design UltraStudio 4K. This is also only thunderbolt 1.0, and thus another device you won’t want to daisy-chain since uncompressed Ultra Hd @ 24p is ~900MB/second (Cinema 4K @ 24p is ~955MB/second) and thunderbolt 1.0 has max transfer speeds of 1GB/second. That takes us down to 2 ports. Lastly, let’s assume that you need 10Gbe network connectivity to deal with all the uncompressed 4K footage you’re editing that you can’t store locally. This is another thunderbolt 1.0 device, and thus another you can’t daisy-chain since a 10Gbe adapter requires all the bandwidth of a thunderbolt 1.0 port. And that leaves us with a single thunderbolt port for all other connectivity (better hope you don’t have legacy firewire 800 drives with your music, photos, etc., else there goes your last port for an adapter, or you need new USB 3.0 or thunderbolt drives.)
Now, here’s the real interesting part. Going back to the fact that all six thunderbolt ports most likely run over only six shared pcie lanes. Assume all 5 of those ports in-use are operating at near capacity, that’s 1.5GB/second for each display (so 4.5GB/second total) and another 1GB/second for the capture card, and another for the 10Gbe NIC. That’s already 6.5GB/second. Now assume you actually have all 2.0 compliant peripherals, and all six ports can operate at full 2.0GB/second, that would be 12GB/second, all running over 6 pcie 3.0 lanes, with a hard bandwidth limit of ~6GB/second. Yep, absolutely no way of avoiding that latency and potential bottlenecking, both very valid (and worrying!) concerns for would-be workstation class computers.
Anyway, I just hope this causes someone to see through Apple’s reality-distortion-marketing and look at the Mac Pro objectively. It’s an “epeen” toy for rich people, or people that have absolutely no choice but to use Apple hardware due to contractual obligations, corporate investments, etc.