I've been there and done that.
Even with the same hardware :-)

Never got it to work beyond 2 cards or 16 screens (4 screens per card and 4 cards also doesn't work properly) in Windows.
Worked fine with the free Nvidia drivers in Linux, but not with Nvidias own proprietary driver. But that wasn't a solution as we needed to run Windows only software on these.

We concluded that the Nvidia drivers are really crappy and badly tested (if at all) for these configurations even though, in theory, it should be possible.

We ended up using 2 computers. On for the top 2 rows, one for the bottom rows.

Another thing to consider: The cards can handle camera streams from that many monitors, but Windows really doesn't like streaming more than 20 or so simultaneously. Get's really choppy even though the hardware didn't get stressed. Seems a limitation of the Windows video codecs or the Windows desktop manager.
Splitting over 2 computers also allowed us to prevent that happening.


I bet you're running into limitations on the memory bridge and other bottlenecks not typically monitored under windows or UNIX since it's things like the CPU and GPU that normally cap out... but since you're pushing the PCIe bus to its maximum you're seeing it.

This is along the lines of what Tony and others who have tried this say: "things get choppy even though the hardware isn't stressed". But it is stressed, just not in a way that is monitored ie; taskmanager and GPU tools.

North/south bridges and the inter-CPU communication pathways all have limitations and you're hitting them with that much bit-slinging.

For this reason I think switching between AMD/nVidia or Windows/Linux is going to make no difference.

My suggestion: Break this up into 3 or 4 machines and then run something like Mulitplicity so it's all seamlessly controllable from one keyboard/mouse.