Without spending any time on it, this is the "most
PCI Express" I could find.
http://www.newegg.com/Product/Product.aspx?Item=N82E16813128508
Note that it still uses the 990FX chipset. I think that chipset
has max total, about 42 lanes or so. Even the official looking
diagrams on the web, aren't complete.
The thing is, you can "fan out" lanes, but the max bandwidth
available doesn't change.
Say, for example, I do this.
x16 ----- Switch ---- x16 (card)
host Chip ------ x16 (card)
If both cards start transferring at the same time, each will
average x8 transfer rate. You cannot make something from nothing.
So when the chipset starts with 42 lanes, you cannot use chips
downstream of that, to make more bandwidth. The number of lanes
may *seem* larger, but depending on the transfer conditions,
the average rates achieved on the cards are still influenced
by the "narrowest piece of pipe".
To get a summary of the wiring pattern there, we don't trust
the Newegg table. This calls for the motherboard manual.
ftp://download.gigabyte.ru/manual/mb_manual_ga-990fxa-ud7_e.pdf
And even that might not be enough.
This is the block diagram. The official diagram is missing a
detail. This is my best guess now. I got a double-check
on ALink from here.
http://compare-processors.com/wp-content/uploads/2012/03/platform-architecture.png
CPU
|
| 6.4GT/sec * 2bytes/T = 12.8GB/sec
|
Northbridge --- x16 ---- slot
| 990FX ---- slot
| --- x16 ---- slot
| ---- slot
|
| ---- x6 (flexible)
|
| ALink 2GB/sec
|
Southbridge ---- x4 (flexible)
SB950
The first thing to notice in such a diagram, is the CPU
connection to the outside world, is limited by Hypertransport.
Now, the claim I see, is the current transfer rate is 6.4 giga-transfers
per second. And that bus is 16 bits or 2 bytes wide. So the processor,
when sending bus traffic in one of the directions, cannot
exceed 12.8GB/sec when doing so. If I take 12800 and divide
by the 500 per PCI Express lane, that equals 25.6 lanes. That
means, the bottom part of the diagram has 42 lanes, but the
processor "bottleneck" supports at maximum 25.6 worth.
The second bottleneck, is the ALink. There are 4x500 for the
four PCI Express lanes off the Southbridge. But, there are
other "loads" on the Southbridge, such as your SATA drives.
Say, for example, you start a read on your 550MB/sec SATA III
SSD, at the same time as the x4 on the side are doing something.
There is a small chance of a bottleneck there.
When there is a bottleneck, nothing breaks, it just runs
slower. Like, when you step on a hose, and there is a lesser
trickle of water from the hose nozzle.
The impressive motherboard above, has the same restriction
as the motherboard you were just looking at.
Two video cards Two video cards
and two blanks and two TV tuners
-------------- -----------------
x16 x8
None x8 <--- just installed Tuner
x16 x8
None x8 <--- just installed Tuner
So that covers the restrictions on four of the slots. It
has twice as many slots for video, as the cheaper motherboard,
but the same style of restriction when all four have something
plugged in.
Now, try to figure out what remains. There are five
PCI Express peripheral chips. We have x6 left on top,
x4 left on the bottom. 10 - 5 = 5 left. Yet, there
are two x4 PCI Express slots, for a total of 8 lanes.
That means those 8 lanes can't be "real".
The Gigabyte diagram says something like this, but
there isn't a "switch" shown feeding the two x4 slots.
NB x6 --- x1 Etron EJ168 USB3
--- x1 88SE9172
--- x4 ------------------ switch --- x4
--- x4
SB x4 --- x1 Etron EJ168 USB3
--- x1 88SE9172
--- x1 RealTek RTL8111E NIC
--- x1 (unused???)
Notice that the two x4 slots, are sharing. If, for
some reason, one wants to run x4, it can. But if both
start transferring at the same time (unlikely, actually),
then they'd be limited to x2 each.
So even a fancy board, using the best (42 lane) chipset
on AMD you can find, is left wanting in places. But
at least you can say, "two video cards can be run
at x16 each". Subject to the bottleneck where the CPU
connects to things. So each card can have 6.4GB/sec average,
if both were transferring at the same time (say, some
Crossfire operation).
HTH,
Paul