My group is considering building a stack of 1 one more S870s. As I understand it, in the S870 configuration you can specify either one PCI Express adapter for the (4) GPUS, or optionally two adapters - driving two cards each.
To verify - there should be no issue in writing single-host applications to be simultaneously allocated to each of the (4) Teslas? I have read postings that refer to a recommended “one GPU per host” - but this seems to be counterintuitive to the architecture of the S870, which would offer a minimum of two GPUS per host…
By using two PCI adapters, can one drive two S870’s with one host? Does anyone have any experience or empirical data that would suggest what the optimum host per GPU ratio should be?
Are there any software issues to worry about when writing multiple-GPU applications?
I also understand that the S870’s are Linux only - is this going to change anytime soon? (We also have win-only apps that we’d like to run on the host for this unit…)
I appreciate in advance your feedback!