Does not seem to be anyone who is able to tell how effective this will be? From my digging around I have come to find that it probably means if you got two of the same RTX cards (above 2070) you can use NVLink with them and it should make it possible to use them as one virtual card… but it also probably requires the software using these GPUs to support this feature.
Nowhere does it seem to be explained clearly though, so which is it?
They do seem to say that this will benefit deep learning. Somewhere I read it might also mean that it would make the RAM count as double, making it effectively twice the RAM. If it can help with deep learning, then it seems possible that it could possible help with the features the Tensor cores are used for in RTX games?
I have pre-ordered an RTX card that could be used for SLI and NVLink. I am likely to buy a second RTX card at some later point even if the benefits are not going to benefit game performance, SLI alone would benefit quite a few games anyway, both VR and none VR. Most games probably wont benefit much from it though.
Maybe someone here knows if there are any reviewers planning on testing this once these cards are released? The whole “make them into one virtual card” just really sounds like they could possibly have a way of making it usable in games without the games supporting SLI and such.
Firstly I will tell you that deep learning and whatever benefits nvlink has will only matter to a handful of people. It’s safe to say that virtually no one that reads this thread is involved with “deep learning”, however now that I’ve mentioned it I’m certain to get that handful of users to reply.
You as a gamer should know to wait for benchmarks. How many pixels are you trying to push that you think would justify a second card? Do you use vr or are you only bringing it up for theoretical discussion?
Yes, I use VR. Its a GPU that will be out in 10 days, I might even have the GPU in 10 days, but cards already exist that can use NVLink, so maybe some knew if NVLink had been tested with games. NVLink does require the cards to be SLI compatible… but being SLI compatible is not enough, so not sure why they mentioned SLI in this regard as only some SLI compatible can be used with NVLink.
Deep Learning is a lot more accessible today than it was just a year ago. Also some of the games that make use of RTX are not using the ray-casting chips, but the Tensor Core chips, and with that being the case it seems likely that NVLink might help with whatever the Tensor Cores will be used for in those games.
I will be using it for VR, games, game development, deep learning, VR game development, some advanced physics systems like Nvidia Flex and at least one alternative to Nvidia Flex that is so far mainly CPU based rather than GPU based.
What will justify the second card? I am trying to find out if it is justified. The second card will probably be the last part of the PC build I have planned. Certainly is way too expensive to just buy it straight away because pretty sure there will be people testing this. For now the only part I have ordered is the RTX 2080Ti, which I will use in my current computer, then sometime next year I will buy a new system that I will put this GPU into instead.
A second GPU is very likely to make sense for me either way though and if I get that second GPU, well then NVLink in itself is not that big an investment and I might as well get that as well. Several regular PC games and VR games support SLI, some VR games also support VR SLI and pretty sure some of them would benefit from two GPUs.
So I am mainly speculating about what the NVLink might add, if anything to the gaming performance, whether it could add something to regular SLI. I suspect it would not add anything, except maybe when it comes to RTX supporting games.
This LTT video says that GTX/RTX cards with NVLink will run SLI on top of the NVLink connection, not the normal NVLink protocol itself and therefore cannot do the memory sharing you asked about: