Actually video playback perhaps isn’t a great example :)
Places where you need PCIe bandwidth:
Video playback (unless you can store entire video sequence on GPU which is pretty rare although theoretically possible if you had a CUDA h264 decompressor. With HAP you’d run out of GPU memory too quickly)
Loading / reloading images (e.g. FileTexture)
Copying between GPU’s (e.g. if you have a single renderer window which spans monitors attached to different GPU’s)
You’re doing a lot of reading back from GPU (e.g. Pipet, Writer)
(3) should be definitely avoided, and also applies when any window is visible on a monitor other than the one it is being rendered on
If you’re just pipeting a few values (e.g. <1kB per frame) then PCIe bandwidth probably isn’t your bottleneck
In general, VVVV is assigned to one GPU (google the forums for startup flags to assign dx11 a specific GPU). In this case, if you use just one vvvv.exe process, then all your other GPU’s are generally just (relatively slow) output splitters.
Still the question stands…
Would any of this be any use for a graphics server?
My instinct is that the software just isn’t there (in VVVV land) to cope with many-GPU. But if it was, then you could do a hell of a lot as long as you didn’t require many inter-GPU copies
I don’t think the celeron cpu – whatever it is exactly – will be able to satisfy all those GPUs. As far as I know most if not all celerons only have 2 Cores. But maybe you can swap it for something more powerful. Also 4 GB of Ram are most likely insufficient and there seems to be only one slot on the motherboard.