I wonder if that benchmark actually loaded the PCIe bus to any significant degree after the initial benchmark startup, or just updated a couple small parameters

I wonder if that benchmark actually loaded the PCIe bus to any significant degree after the initial benchmark startup, or just updated a couple small parameters on a single scene and thus mainly just tested the local computation on the GPU? You'd want to somehow monitor the PCIe bus error rate - with a marginal signal and lots of errors -> retries, something that loads the bus harder (loading new textures etc) could suffer way more. They do briefly show a different PCIe riser made out of generic ribbon cable [1, 3:27], and say that one failed after chaining only two ~200mm lengths. The quality of the riser cable certainly matters. [1] https://youtu.be/q5xvwPa3r7M?t=207 https://youtu.be/q5xvwPa3r7M?t=207