NVLink/PCIe Bottleneck

[rarifiedbovine]

Basically, I have heard and seen some numbers suggesting that whether you're doing x8/x8 or x16/x16 doesn't matter anymore when you're using NVLink for SLI.  That doesn't make a ton of sense to me but some results obtained by GN testing seem to suggest the same.  I'd love to avoid having to go with an HEDT CPU since those are a bit slower per core, resulting in overall worse gaming performance for most games--in the past, I've used the HEDT line only because of the x16/x16 possibility.  Is it true that PCIe bandwidth no longer matters, and that games will have absolutely no advantage from going with an HEDT platform?

[Jurrunio]

Graphics cards don't care about PCIe 3.0 x8 or x16 these days unless you do crazy things like 2080ti in DX12 multi GPU without any NVLink bridge whatsoever. so yes, there's no advantage for going with an HEDT platform for extra PCIe lanes to the graphics cards for a REALLY LONG time.

 

[JoostinOnline]

You seem to be under the impression that SLI and NVLINK communicate via PCIe lanes. With the exception of a few early SLI cards (low budget ones that you'd never use in SLI anyway), they don't. That's what the bridge is for, and that's why it doesn't get the same micro-stutters that Crossfire (which DOES use PCIe lanes) suffers from. So no, it's not a bottleneck, but it never has been one in the past either.

[rarifiedbovine]

1 minute ago, JoostinOnline said:

You seem to be under the impression that SLI and NVLINK communicate via PCIe lanes. With the exception of a few early SLI cards (low budget ones that you'd never use in SLI anyway), they don't. That's what the bridge is for, and that's why it doesn't get the same micro-stutters that Crossfire (which DOES use PCIe lanes) suffers from. So no, it's not a bottleneck, but it never has been one in the past either.

Thanks, that's not quite what I mean though.  I guess my question is that the GPU has to get instructions from the CPU or directly from the application.  That's done through PCIe, right?  In that case, if you have two plugged in to a 26-lane motherboard, each gets x8.  Right?

[Enderman]

With PCIe 3.0 a GPU doesn't use more than x8, so there is no performance difference if you give it x16.

[JoostinOnline]

20 minutes ago, rarifiedbovine said:

Thanks, that's not quite what I mean though.  I guess my question is that the GPU has to get instructions from the CPU or directly from the application.  That's done through PCIe, right?  In that case, if you have two plugged in to a 26-lane motherboard, each gets x8.  Right?

Yes, but a single GPU doesn't even use up eight 3.0 lanes (although the RTX 2080 Ti comes close).

 

Again though, this isn't a new occurrence. PCIe bottlenecks are incredibly rare. I suspect the idea that 16 lanes wasn't enough came from AMD fanboys who needed something to brag about before Ryzen came around. I can only guess though.

[rarifiedbovine]

56 minutes ago, JoostinOnline said:

Yes, but a single GPU doesn't even use up eight 3.0 lanes (although the RTX 2080 Ti comes close).

 

Again though, this isn't a new occurrence. PCIe bottlenecks are incredibly rare. I suspect the idea that 16 lanes wasn't enough came from AMD fanboys who needed something to brag about before Ryzen came around. I can only guess though.

That seems reasonable and I WANT to believe it's the end of the story, but see the following: 

 

https://www.gamersnexus.net/guides/3366-nvlink-benchmark-rtx-2080-ti-pcie-bandwidth-x16-vs-x8 (Ashes of the Singularity results); and

 

 (some posts indicating worse results with x8/x8.

 

 

[JoostinOnline]

10 minutes ago, rarifiedbovine said:

That seems reasonable and I WANT to believe it's the end of the story, but see the following: 

 

https://www.gamersnexus.net/guides/3366-nvlink-benchmark-rtx-2080-ti-pcie-bandwidth-x16-vs-x8 (Ashes of the Singularity results); and

 

 (some posts indicating worse results with x8/x8.

 

 

They removed the bridge in those tests just to see if they could reach the limit. Basically they ran it the same way that Crossfire is run, which is an inferior method. See this line:

Quote

For the 2080 Tis, we removed the NVLink bridge and just tested them via explicit multi-GPU via the PCIe bus.

If you're planning on doing SLI, use the bridge. It's got a purpose.

[rarifiedbovine]

50 minutes ago, JoostinOnline said:

They removed the bridge in those tests just to see if they could reach the limit. Basically they ran it the same way that Crossfire is run, which is an inferior method. See this line:

If you're planning on doing SLI, use the bridge. It's got a purpose.

Ok, so basically, if you’re playing games, even SLI,  go with 9900K over cascade lake-x?

[JoostinOnline]

11 minutes ago, rarifiedbovine said:

Ok, so basically, if you’re playing games, even SLI,  go with 9900K over cascade lake-x?

I wouldn't get either of them. The 9900k may be the best, but it's a terrible value. For the context of your question though, yes.