PCIe 4.0 to be finalized in 2017, with new connector and 16 GT/s data-rates

[Bouzoo]

because of the massive bandwith increase, meaning we are less reliant on VRAM meaning we can potentially devote more transistors to logic instead of memory controllers, meaning more powerful gpus.

 

I didnt see anyone complaining over mantle being a closed vendor locked solution that potentially increases FPS in games. fucking amd fanboys and idiots who dont think stuff through /rant

The underdog has to compete so it's okay. /s

But with the HBM2 and DX12 coming, do we really now need NVLink if those 2 are gonna be such big game changers? I'm not really sure we do.

Mind you it wouldn't hurt if we got it.  :lol:

[LukaP]

The underdog has to compete so it's okay. /s

But with the HBM2 and DX12 coming, do we really now need NVLink if those 2 are gonna be such big game changers? I'm not really sure we do.

YES. The more stuff you can leave in system memory, the better the game can look, if you want to look at it from that perspective (and with it being 12x faster than PCIE that means 192GT/s which is the bandwith of midrange videocards memory interfaces. that is ALOT of bandwith)

[Bouzoo]

YES. The more stuff you can leave in system memory, the better the game can look, if you want to look at it from that perspective (and with it being 12x faster than PCIE that means 192GT/s which is the bandwith of midrange videocards memory interfaces. that is ALOT of bandwith)

Yes but iirc current consumer cards can't even max PCIe 3.0 (I think not even 2.0), so how are they gonna use NVLink to it's fullest? I don't remember, but was HBM able to utilize the full bandwith of PCIe?

Maybe fiber will do the trick, since they are thinking using it for 4.0.

[LukaP]

Yes but iirc current consumer cards can't even max PCIe 3.0 (I think not even 2.0), so how are they gonna use NVLink to it's fullest? I don't remember, but was HBM able to utilize the full bandwith of PCIe?

its not the memory bandwith that matters here. its how you utilise the VRAM + interconnect bandwith. so far we didnt need to utilise the interconnect that much, since VRAM was plenty and the APIs didnt really support that high of a fidelity to saturate that. With DX12/Vulcan and HBM limiting the capacities somewhat, its ever more important to leave stuff in RAM and just access it through the interconnect when necessary.

[BirdyTheMighty]

they better come with an adaptor if its somewhat backwards  compatible