Splitting the interface signal is the way to go with such high end internally connected arrays . And the next part of the project was to do exactly that . But hitting a brick wall so soon with three modern nVidia GPUs running a recent OS X version has dampened my spirits .
Now I am considering System configurations that would permit splitting a single host interface for the video rendering and freeing up the other host interfaces for other peripheral cards (EFI UI graphics card , controller cards, etc.) .
There should be more than enough bandwidth here , in a cMP .
In fact , with the Nehalem Mac Pro available bandwidth should be :
PCIe slot 1 = 8 GB/s (16 x 500 MB/s)
PCIe slot 2 = 8 GB/s (16 x 500 MB/s)
PCIe slot 3 = 2 GB/s (4 x 500 MB/s)
PCIe slot 4 = 2 GB/s (4 x 500 MB/s)
With keeping in mind slots 3 and 4 share resources .
So , you see , there's a lot to work with here . And the GPGPU array I have already built should , in theory , work fine connected to a single x16 PCIe slot with the appropriate splitting . We're talking 4 lanes of PCIe Rev 2 per graphics card .
For comparison purposes : the nMP (Cylinder) has just a 2.5 GB/s connection through a single TB2 port , which is how an eGPU expansion chassis would connect to this System .
Last edited:
Retail price? Who pays that? For HP?
