schmidtbag:

No, they aren't. Only a handful of GPUs can max out 2.0 @ x16 or 3.0 @ x8.

All mid to high end GPU's since Kepler can max out PCI-E 2 and 3 16 links under complex workloads

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080_PCI_Express_Scaling/3.html Older and tested on a 1080, but this gives you some example of PCIE scaling. Not really a huge difference.

Pcie4 would benefit from multi gpu systems. Too bad they are dead.

Astyanax:

All mid to high end GPU's since Kepler can max out PCI-E 2 and 3 16 links under complex workloads

Talking about games, most of current GPUs do not take more then 1-2% performance improvement in most AAA titles and most of time the performance improvement is 0% or just system noise.

Undying:

Pcie4 would benefit from multi gpu systems. Too bad they are dead.

Multi-GPU is mostly limited by latency times, not bandwidth. They aren't also dead, simply AMD and NVIDIA decided wisely to limit the linked node adapter support to 2 adapters (ie: crossfire and sli are limited to 2 cards), however low-overhead/abstraction APIs like Direct3D 12 and Vulkan do not need linked node adapter to take advantage of multiple GPUs at all.

Alessio1989:

Multi-GPU is mostly limited by latency times, not bandwidth.

Isn't [modern] Crossfire more heavily dependent on bandwidth? That uses the PCIe bus for cross-communication between the GPUs, whereas SLI uses a discrete hardware bridge.

schmidtbag:

Isn't [modern] Crossfire more heavily dependent on bandwidth? That uses the PCIe bus for cross-communication between the GPUs, whereas SLI uses a discrete hardware bridge.

It depends on the single workloads, of course more bandwidth is always welcomed, but the most complex aspect, which is not resolved at all on todays hardware, is accessing on a memory resource which is resident on another physical adapter. This is why most of multi-GPU rendering techniques are restricted to decouple the rendering resources (buffers, vertices and indices, and textures) on both GPUs and use AFR or using one GPU for the main rendering and the second for heavy post-processing. Again, higher bandwidth can help and is always welcomed, but it's not the solution to improve multi-GPU performance and programming facility.

Astyanax:

All mid to high end GPU's since Kepler can max out PCI-E 2 and 3 16 links under complex workloads

There are certain workloads like cuDNN and Caffe that show large differences between PCIE-3 x8 and x16. However I don't know of any significant differences in gaming which is probably 99.99% of anyone reading this forum. PCIE-4 likely will be a boon for reducing latency due to its faster signaling which will also have some fringe use cases let alone the obvious bandwidth increases. Even with a RTX 2080 ti you only lose 2% performance in gaming moving down to PCIE-3 x8 from x16. You can see significant bottlenecks in PCIE-3 x8 when you SLI 2 RTX 2080 ti's without the NVlink ie over PCIE but that is more theory than applicable since nobody would do that in the real world.