Last week’s feature on the NextIO PCI-Express Virtualization & Expansion appliance drew some questions on what kind of bandwidth you could expect from the device. Some people believed that you could get a full x16 of bandwidth out of it, so I sent the question on to NextIO and here’s their response:
GPU density and performance per server is application dependent. The NextIO appliances support multiple x8 Gen2 PCIe connections per server and a Gen2 PCIe connection per GPU. As a user adds more GPUs to the cluster, more bandwidth to the server can be added based on application demand. By way of comparison, the nVidia Tesla S1060 GPU appliance supports 4 GPUs with 2 x16 Gen2 server connections. This is an average of x8 Gen2 per GPU. The same server bandwidth is available with the NextIO appliance. Additionally, if an application does not require full x8 per GPU, more GPUs may be consolidated per server connect.
So essentially they say that the existing “top of the line” offering from NVidia, the S1060, offers 4 GPU’s on 2 x16 connections, giving you an effective x8 to each card, which is the same as their offering.
NextIO – PCIe Expandability, Virtualization, & Hot Swap | VizWorld.com.
