Would you please tell me if we have to use the same PCIe bus width for both of the GP100 cards connected with NVLINK bridge?
For example, does the CUDA driver support for the following configurations?
I understand that narrow bus width cannot bring GP100’s high performance.
PCIe x16 GEN3 --- GP100 #1 ---+ NVLINK Bridge PCIe x8 GEN2 --- GP100 #2 ---+
PCIe x8 GEN3 --- GP100 #1 ---+ NVLINK Bridge PCIe x1 GEN2 --- GP100 #2 ---+