Everywhere on internet i read about NVLink is more powerful than PCI lanes connection.
OK, right. But now, i would use this power on my future computer for deeplearning (especially Resnet/TensorFlow training), mainly with GPU scripts based (Cuda). And why not on my current computer.
I'm reflecting about to buy this new base (sorry, i don't have the money to buy a DGX-1, neither a V100):
But the PCIe is badly limited to x16/x16 with dual gpu, and x8/x8/x8/x8 with quadrial gpu... So, i'm not satisfied at all.
Is it possible to active NVLink instead of PCIe to improve the performance ?
Is there another solution to override/improve the number of lanes ?
At least, is there another mobo/cpu supporting x16/x16/x16/x16 with dual CPU ?
Bios modding or whatever, i would want to go beyond this limit.
[Edit] The motherboard seems to handle 2 PCIe x16 for each cpu. The i7 are 1S only, i changed to Xeon 2600 v4 series, thanks Alpha3031
[Edit] Finally, there is no possibility to use nvlink for now, i took a Supermicro Superserveur 4U 7048GR-TR with 6 pcie v3 x16/x16/x16/x16/x8/x8 with dual Xeon 2640 v4 and 1080ti gpus