PCI-E 3.0 8x now and in the future

Skakruk said:
Because of XDMA?

What about a single 295 in 16x vs 8x.

Has anyone seen benchmarks for this with cards like the 295 or Titan Z? The most recent numbers I could find were from the original Titan (which showed no effect), but they only tested gaming, no compute.
Click to expand...

I tested my 780ti classy in pcie 2.0 x16 v 3.0 x16 and was barely a difference. 30 pts I think in firestrike.

I guess to get closer, I can OC to the highest I've had the card then run the two pcie gens.
 
high end single gpu cards aren't saturating the pci-e bus. and i can't see the titan z as a factor in determining potential v3.0 8x vs v3.0 16x bottlenecks largely due to the fact that sli still utilizes a bridge interconnect to move data between two cards. the r9 295x2 is different in that it exclusively uses the pci-e bus to communicate with another r9 295x2.

the real question is whether a 2011 socket solution (with dual v3.0 16x multi gpu capabilities) is better than an 1150 or 1155 socket solution (with dual v3.0 8x multi gpu capabilities) at handling the enormous data exchange required by two r9 295x2's in the current crossfire method.
 
It's not just a factor of the how powerful the GPU is but the workload as well to determine whether or not pcie bandwidth can lead to performance improvements/limitations, since it governs data transfer between the GPU and rest of the system.

Skakruk said:
Because of XDMA?

What about a single 295 in 16x vs 8x.

Has anyone seen benchmarks for this with cards like the 295 or Titan Z? The most recent numbers I could find were from the original Titan (which showed no effect), but they only tested gaming, no compute.
Click to expand...

Some compute tests for pcie 2.0 vs 3.0 with the 7970 - http://en.expreview.com/2012/01/18/pci-e-3-0-vs-pci-e-2-0-test/20390.html
 
Last edited:
Skakruk said:
I guess that could have just been within the margin of error, but if there was a difference, regardless of how small, that kind of suggest to me that the difference for the next couple of gens might end up being significant.
Click to expand...

I've never seen a significant gain from PCI Express standards going from one generation card to the next and I doubt the next generation will be any different, or the generation after that really.
 
There is nothing that saturates the lanes yet. PCIe 3.0/8x operates at PCIe 2.0/16x, which is still very fast,
 
The Cobra said:
There is nothing that saturates the lanes yet. PCIe 3.0/8x operates at PCIe 2.0/16x, which is still very fast,
Click to expand...


Even then PCIe 1.0 16x is pretty damn solid (i.e. PCIe 2.0 8x). The main issue is bandwidth will always suck going from GPU to CPU, hence one of the reasons why everyone is working on a Unified Memory architectures and stacked memory in the long-term.

As of today, most GPU processing is meant to be localized as much as possible with as minimal data being transferred across those lanes as you can get. Distributed computing may be the exception, and even then system architectures are specially built.
 
I think latency is the main issue, not so much bandwidth. If bandwidth was a major bottleneck, I would think we'd see a much larger performance delta between the different PCI Express modes, since that's essentially what PCI Express is, data lanes going from the CPU to the GPU.
 
Is there really anything that requires large amount of bandwidth once the GPU starts rendering? From my understanding, you would want to fit everything in the VRAM before you start rendering anyway, otherwise if the card have to swap textures in and out, you'll experience stuttering regardless of how fast the PCIE bus is.

Therefore I don't see how the PCIE bandwidth will play a role in the GPU performance.
 
For single gpu configurations, bandwidth is rarely a concern with an 8x+ slot, whether it's pci-e 2.0 or 3.0. However, for multi-gpu setups (specifically 3 and 4 gpu's) the bandwidth impact can be much more significant. Just ask Vega
 
RamonGTP said:
I think latency is the main issue, not so much bandwidth. If bandwidth was a major bottleneck, I would think we'd see a much larger performance delta between the different PCI Express modes, since that's essentially what PCI Express is, data lanes going from the CPU to the GPU.
Click to expand...

latency is so frequently overlooked. the thing you have to keep in mind though is the timings. waiting 10 cycles at 1ghz is the same amount of time as 20 cycles at 2ghz. So while bumping the clock speed is a bandwidth win every time, if you have to increase the timings at the same time that you increase the clock speed you might end up with a net latency increase. This happens every time new ram formats are released (DDR2,3,4) where the first parts from the new generation have great bandwidth but high timings means that the older formats actually have better latency. Of course, with RAM as time goes on they improve the latencies until they are better than the previous gen.
 
You must log in or register to reply here.
Back
Top