PCI-SIG Releases PCIe 4.0 Specs

[Guest]

PCI-SIG released the specs for PCIe 4.0, which promises data transfer rates of 16GTps, and said work on PCIe 5.0 is well under way.

PCI-SIG Releases PCIe 4.0 Specs : Read more

 

[bit_user]

In the best case, I think this could usher in a new era of multi-GPU gaming & VR. Especially if we see GPU performance start to plateau, like CPUs have.

 

[alextheblue]

In the best case, I think this could usher in a new era of multi-GPU gaming & VR. Especially if we see GPU performance start to plateau, like CPUs have.

I honestly don't think that's very likely until we start bumping into silicon limits.

For VR it might be not be too big of a mess at least for twin identical cards, one GPU per eye. Otherwise it's the same old story, poor software support, all-around fustercluck.

 

[Icepilot]

"We don't expect Intel to support PCIe 4.0 until mid-2019, and AMD has said it plans to add support in 2020."
"All signs lead us to believe we will have PCIe 4.0 in the consumer space next year."

Running on what?

 

[bit_user]

Especially if we see GPU performance start to plateau, like CPUs have.

I honestly don't think that's very likely until we start bumping into silicon limits.

Yeah, that's basically what I'm talking about. It's not just feature size that's a limiting factor, but also things like leakage that limit how much benefit you get from a node shrink.

 

[JonDol]

I hope that for your estimations (?) about its adoption are not true and that for once AMD will rather 'set the trend' than 'catch up'. I hope this too with the adoption of the USB 3.2 so that will help us to forget the proprietary Thunderbolt.

 

[mapesdhs]

Alas Jondol, hope is not a plan...

 

[samer.forums]

5 years for PCIe 3.0 and only 2 years between 4.0 and 5.0 ? Just skip the 4.0 altogether. Otherwise companies will never produce many cards on the 4.0 and will wait until 2019. 2 years is bad for R&D and companies wont waste money on them if the 5.0 is only 2 years after.

Having said that , I am waiting for the PCIe 5.0 so much ... using x4 PCIe 5.0 instead of x16 slots will save TONS of space on motherboards .. and will allow full speed external PCIe GPU with ease using just 4 lanes.

 

[10tacle]

So we go from 3 years -> 4 years -> 7 years -> 2 -> years between PCIe 1.0 to 5.0. PCI-SIG first announced 4.0 specs back in December 2011! Why the huge gap between 3.0 and 4.0? Is the Enterprise market (non-GPU devices specifically) getting in the way of things in demands?

 

[samer.forums]

So we go from 3 years -> 4 years -> 7 years -> 2 -> years between PCIe 1.0 to 5.0. PCI-SIG first announced 4.0 specs back in December 2011! Why the huge gap between 3.0 and 4.0? Is the Enterprise market (non-GPU devices specifically) getting in the way of things in demands?

because we simply did not need that bandwidth increase... there are no GPU cards that can saturate x16 PCIe 3.0 lanes today , and no SSD as well ... unless you are talking about $100K cards...

The M2 NVME SSD saturated the x4 PCIe 3.0 lanes ... so now we can make a use of Gen 4.0

IMO

 

[10tacle]

^^Oh I understand that, especially for gaming. There are countless tests out there showing nearly zero improvement of x8 vs. x16 in gaming except in high end multi-GPU tests and extreme crazy non-realistic GPU setting like 4K in surround on low quality (https://www.pugetsystems.com/pic_disp.php?id=40772&width=800&height=800). My question was more for the professional development progress side of things (ie: Enterprise).

 

[DerekA_C]

Yeah, I feel you guys are forgetting that AMD has Infinity Fabric, meaning they can put (up to) 4 gpu's on a single PCB within a single die "theoretically" and eventually 8 when they hit below 7nm. Nvidia is working on something similar. mGPU is no longer going to be limited to multiple PCIe slots it will also be for single slots and PCIe4/5 will allow such possible upgrade path. With high enough bandwidth and low enough latency (mGPU) will have a slight latency increase like the Threadrippers, but nothing like SLI/Xfire or dual chip single PCB's.

Look on the bright side PCB's should get smaller, and we could see a 80% if not more increase in GPU Performance over it's single GPU chip and 20 to 30% over the current mGPU dual chip PCB, being on the same die instead of tied together through small wires with a lot of distance between, makes a lot of sense for the next step in GPU acceleration.