I was wondering if anyone had experience with Tesla bandwidth.
My question is the following - if a 8800 GTX you use for CUDA that plugs in to your desktop requires a PCI X16 slot, how is that a Tesla 1U server with 4 GPUs can interface to the host computer via only one PCI X16 or even X8 slot?
Does this mean that the bandwidth to the Tesla GPUs is reduced compared to a G80 plugged into your machine?
Ecah Tesla S870 interface card is a PCI-E-gen2 card supporting 2 GPUs. Hence, each G80 is getting the equivalent of PCI-E-gen1 bandwidth. This, of course, assumes that you’re using a PCI-E slot in your motherboard which has PCI-E-gen2 support.