Originally Posted by nht
Hmmm...12x? That's in the data spec and they control both GPU and MB completely.
Maybe, Thunderbolt bandwidth is confusing. There are two controller types: 2-channel and 4-channel. The 2-channel ones have 40Gbps aggregate, and the 4-channel 80Gbps. PCIe 2 x4 is 40Gbps:
"(4 x 10Gbps bidirectional = 80Gbps aggregate bandwidth)"
The Macbook Air uses the 2-channel one, which could be why it got the update at WWDC.
They exceeded the 10Gbps data in one direction in a test here:
According to Anandtech, Apple's been running the controller off the CPU lanes so they might have been using PCIe 3 lanes already, possibly just two lanes.
But Thunderbolt 2 gets rid of the reserve for display bandwidth so that should mean it's possible to have 80Gbps aggregate PCIe per pair of ports (240Gbps aggregate external - same as the old Mac Pro had internally, just divided differently), which is double the old Thunderbolt data limit. Technically it's the same overall aggregate as before but you have the ability to use the previously reserved bandwidth for something other than display data.
That would require 12 PCIe 3 lanes for the Thunderbolt on the MP, leaving 28 to divide between two GPUs and the storage. The GPUs could be on x12 each as you say, leaving x4 for storage and anything else.
It definitely has 40 lanes as the marketing says 40GB/s PCIe 3 bandwidth (1GB/s per lane). There are legacy PCIe 2 lanes on some motherboards though so they could be getting bandwidth elsewhere for things like the SSD.
I'm sure when it comes out we'll know for certain where the bandwidth is allocated. For heterogeneous computing, it makes sense to have them on as fast an interface as possible but again, it depends on what's being done. I don't see why data would have to be copied across the interface a lot when there's enough memory on each side because even on PCIe 3 x16, that's less than 1/10th the GPU memory bandwidth so copying data around isn't efficient.
Originally Posted by PhilBoogie
Huh? I put my MP5.1 with 10.8 to sleep instead of shutting down. That's hibernation, right? I'm missing something here...
Hibernation is distinct from sleep - it's used to mean dumping the contents of RAM to the hard drive so the RAM can be powered down for long standby (RAM is volatile so it has to stay powered for the contents to stay in there, SSD/HDD is non-volatile so it can be shut off from power and keep the memory state). Hibernation is used on laptops to maintain standby time (battery life while the computer is asleep) but desktops are powered all the time so they just sleep without hibernation.