Hi all,
I’m running an RTX 6000 Blackwell (Pro Edition) on a Ryzen 7500X3D (ASUS B850 platform) with a Samsung Gen5 NVMe.
I’ve successfully isolated the RTX 6000 by moving the display to the iGPU and cleaning all Xorg/Gnome-shell processes (verified via nvidia-smi). I am using the 590-series open drivers and CUDA 13.1.
The Problem: > gdscheck reports “Unsupported” for NVMe P2P unless use_compat_mode is enabled. My topology has the GPU and NVMe on separate bridges (00:01.1 and 00:01.2).
Even with manual setpci ACS bypasses and iommu=relaxed, the nvidia-fs driver seems to detect ACS in the capability registers and refuses to establish a P2P path.
Question: > For Blackwell cards on consumer Gen5 platforms where BIOS ACS controls are limited, is there a modprobe parameter for nvidia-fs or a cufile.json setting to force the driver to ignore ACS hardware flags when the user has verified a safe P2P path?
Currently hitting 11 GiB/s in compat mode, but hunting for that 14+ GiB/s P2P limit.
Thanks!