Question What are my options for the most powerful GPU when can be run only at PCIEx8 lanes?

[macgyver27]

Hello.

My scenario is that I have a build where I have 2 GPUs. Machine is natively running Linux and using Nvidia Quadro as its GPU. Then I have a VM where I passthrough 2nd GPU.

My CPU is Ryzen 9 5900X. Motherboard is x570-aorus-master and when I have occupied both PCIE slots they both run at x8 lanes - see the motherboard manual:

Integrated in the CPU (PCIEX16/PCIEX8):

3rd Generation AMD Ryzen™ processors:
- 1 x PCI Express x16 slot, supporting PCIe 4.0 and running at x16 (PCIEX16)
- 1 x PCI Express x16 slot, supporting PCIe 4.0 and running at x8 (PCIEX8)
2nd Generation AMD Ryzen™ processors/3rd Generation AMD Ryzen™ with Radeon™ Graphics processors:
- 1 x PCI Express x16 slot, supporting PCIe 3.0 and running at x16 (PCIEX16)
- 1 x PCI Express x16 slot, supporting PCIe 3.0 and running at x8 (PCIEX8)
    * For optimum performance, if only one PCI Express graphics card is to be installed,
    be sure to install it in the PCIEX16 slot.
    * The PCIEX8 slot shares bandwidth with the PCIEX16 slot. When using the 3rd
    Generation AMD Ryzen™ processors/2nd Generation AMD Ryzen™ processors/3rd
    Generation AMD Ryzen™ with Radeon™ Graphics processors, the PCIEX16 slot
    operates at up to x8 mode.
2nd Generation AMD Ryzen™ with Radeon™ Vega Graphics processors/1st Generation AMD Ryzen™ with Radeon™ Vega Graphics processors:
- 1 x PCI Express x16 slot, supporting PCIe 3.0 and running at x8 (PCIEX16)
Integrated in the Chipset (PCIEX4/PCIEX1):
- 1 x PCI Express x16 slot, supporting PCIe 4.0 (Note 1)/3.0 and running at x4 (PCIEX4)
- 1 x PCI Express x1 slot, supporting PCIe 4.0 (Note 1)/3.0

What is the most powerful gaming GPU parameters (or example of cards from AMD/Nvidia) I can use based on this limitations - so I don't overpay the performance I wont be able to utilize?

 

[pete_101]

You could run any current GPU and not see much of a drop in performance. PCIe 4 running at x8 is the same bandwidth as PCIe 3 x16. This is fast enough for all the top end GPUs, however, just about all GPUs run a few percent slow at x8 than x16. I don't know the exact reason why they run slower even if there is enough bandwidth, I suspect using 16 lanes rather than 8 is just optimal, this is the small price you pay for not having x16.

 

[bill001g]

When the 4090 first came out there were a number of benchmark people trying to see how much difference pcie4 vs pcie3 made. The way they simulated it was to run pcie4 x8 rather than x16 since that is the same bandwidth drop between pcie4 and pcie3.

There conclusion was basically even the 4090 could not use the total bandwidth. This got into the discussion why it also doesn't matter that cards don't currently support pcie5. In theory if the 4090 supported pcie5 it could run on 4 lanes.