Clarification needed on PCIe lanes, cross generation usage and bandwidth

[FeeNiX]

I may not be using the technically precise proper terms in places, so I’m trying to over explain the question so that I am hopefully properly understood.

I know that speed roughly doubles generation over generation, so theoretically an x16 PCIe 4 port would have the same bandwidth as an x8 PCIe 5 port.

For this question, lets assume my video card is a 3090TI (PCIe 4) and that it saturates fully a PCIe 4, x16 slot. My motherboard is an AM5 platform, with two PCIe Gen 5 x16 (electrically and physically) slots, that operate in x16/x0 or x8/x8 directly off the CPU.

Once I utilize the second x16 slot, the first is operating at x8. While this would be enough bandwidth for the 3090’s full speed, is the throughput halved because less “pins” in the slot are utilized? Basically do lanes go directly to pins in the slot, or is there some data buffer between the slot and lanes to the CPU so that the data can be transferred from all the contacts on the card and then sent over fast lanes to the CPU?

I’m also assuming if you put an x16 card into an x16 slot wired only for x8, the throughput is constrained by the fact that half of the contacts on the card aren’t used. Is this correct?

[igormp]

@FeeNiX

so theoretically an x16 PCIe 4 port would have the same bandwidth as an x8 PCIe 5 port.

Yes, in theory.

@FeeNiX

While this would be enough bandwidth for the 3090’s full speed, is the throughput halved because less “pins” in the slot are utilized? Basically do lanes go directly to pins in the slot

Since your 3090TI is a PCIe 4.0 device, that slot would only work as a PCIe 4.0 slot, so you’d have the bandwidth provided by a PCIe 4.0 x8 slot.

@FeeNiX

I’m also assuming if you put an x16 card into an x16 slot wired only for x8, the throughput is constrained by the fact that half of the contacts on the card aren’t used. Is this correct?

Yes, one can see it that way.

[FeeNiX]

Good point on the generation thing regarding the video card. That was a pebkac question. Knowing that lanes go to pins clears things up though. Thanks.

[cowphrase]

When talking about video card PCIe bandwidth, you also need to factor in VRAM. When gaming most resources are preloaded, so if your Game fits within the VRAM then the bandwidth requirements are going to be lower. If the Game doesn’t fit in VRAM it needs to constantly load new resources to the card, and bandwidth requirements are higher.

@FeeNiX

Once I utilize the second x16 slot, the first is operating at x8. While this would be enough bandwidth for the 3090’s full speed, is the throughput halved because less “pins” in the slot are utilized? Basically do lanes go directly to pins in the slot, or is there some data buffer between the slot and lanes to the CPU so that the data can be transferred from all the contacts on the card and then sent over fast lanes to the CPU?

When both slots are in use only 8 lanes are electronically active. There is a chip (commonly called a PLX chip) which switches the lanes like a railway switch. So a PCIe 4 16x device would operate at PCIe4 8x, since only 8 lanes are connected.

Here is an example of an older motherboard with a block diagram https://overclock3d.net/reviews/cpu_mainboard/gigabyte_b550_aorus_master_preview/6. You can see the section highlighted in red can be a 16x slot, or an 8x slot with two 4x M.2 slots. Many motherboards have a similar block diagram in their manual.

A PCIe4 GPU can downgrade to earlier versions of PCIe if the slot/motherboard is an older version. A slot/motherboard can also downgrade to an earlier PCIe version if the GPU only supports the older version.

[FeeNiX]

Thanks to both of you.