It is crazy how much people talk about the importance of PCIe lanes, and how allegedly x40 lanes offered with s2011-3 and the s2011 platforms before it were giving an obvious advantage over “mainstream” sockets. Puget puts this under scrutiny with a pair of very powerful Titan Xs.
The claim that PCIe bandwidth is “never enough” has been put to the test multiple times before: when we transitioned to PCIe 2.0, then to PCIe 3.0 and everytime, the bottleneck was re-invented to the minds of people: 16x PCIe lanes are not enough – regardless of TFlops and regardless if the available bandwith could write and re-write the available VRAM multiple times per second.
In this latest test, Puget tries to see if 2016 is the time that will help these claims hold water for the first time. Spoiler: not really. Included are some GPU rendering tests too.
And before someone jumps into it thinking NVMe SSDs would make a difference, well, I dare you to try it and report about it. I highly doubt it, simply because no realistic scenario applicable to 99.99% of situations can stress the PCIe lanes with 2+ GB/s of reads/writes to and from the SSD, while pushing tenths of GB/s to the cards…just doesn’t happen.