PCIe bandwidth requirements (RTX 50xx edition)

[Albuquerquefx]

Just some napkin math: each lane of PCIe 3.0 is a touch shy of a gigabyte per second of bidirectional throughput (eg it can be transmit, or receive, or both at the same time at the full ~985MB/s.) Meaning, at least in theory, a PCIe 3.0 x4 slot would be able to manage, with some to spare, the throughput needs of @arisu's GTX970M.

Part of me wonders if that particular project was perhaps VRAM heavy? Most of the 970M's were shipped with 3GB of VRAM, so maybe combined with the compute kernel and dataset, it needed to continually page in the data? That seems unlikely to be honest, but I can't imagine why else there would be a continuous stream of 3GB/sec to a 3GB VRAM card. Would be interesting to see that same WU on a card with larger VRAM, just for A/B comparison sake on the PCIe throughput.

[CaptainHalon]

But for what it's worth to anyone, I also (surprisingly) noticed quite a difference going from x8/x8 pcie 4.0 to x8/x8 pcie 5.0. At least 2-3m PPD on the combined total for both cards. These are trends I've noticed over multiple days. So take it with a grain of salt if you will, but from what I'm seeing it matters more than I would have thought.

Something else changed, like OS, SATA/NVME or something in BIOS.
I tried RTX 2080 with pcie2.0 x8 and pcie3.0 x8. Found no difference in PPD.

Negative, ghostrider.