Hey all
I am planning to build a 4 GPU PC like Nvidia's devbox
https://developer.nvidia.com/devbox for deep learning work (Not for gaming, SLI is not needed).
Motherboard: Gigabyte x99-sli
GPU: 4 GTX 1080ti
CPU: 6800k (Max # of PCI Express Lanes: 28) OR 6900k (Max # of PCI Express Lanes: 40)
SSD: Samsung 850 EVO Series 1TB M.2 (to be connected on PCI)
I suspect there would be a problem with the Max # of PCI Express Lanes for each CPU.
My guess is that when the 6800k is used, then the 4 GPUs will be connected as (x8, x8, x8, x4), while on the 6900k they will be connected as (x16, x8, x8, x8). But I don't know how the PCI SSD would affect this setup.
Can the GPUs and the SSD work in parallel, or one of the GPUs connections would be suspended?
Also, will the x4 connected GPU slow down the other gpus while transferring the data between each iteration?
Will there be a bottleneck generally in the PCI lanes due to the 4 GPUs and the PCI connected SSD if I used the 6800k CPU instead of the 6900k, and are there any benchmarks that gives quantitative results?
I have seen other motherboards that use PLX for switching between the PCI connections, does it improve performance or the Gigabyte x99-sli is sufficient?
Thanks
