Hello,
We have an RTX PRO 6000 Blackwell GPU in our datacenter running on SLES 15 with driver version 580.65.06. In previous generations, we disabled ECC for rendering, but now it can’t be turned off. Could you please advise on this? Is there any way to disable it? Or is it an issue with the drivers? Thank you.
Based off my experience with the RTX PRO 6000 Workstation cards, ECC is an inherent feature of the GDDR7 RAM and is always enabled (unlike older cards where turning on ECC seemed to be a function of the GPU and came at a performance cost and used up a small percentage of RAM capacity).
nvidia-smi will report ECC as being Off if you have your GPU in graphics mode (default setting from the factory for Workstation cards), and it will report ECC as being enabled if you have your GPU in compute mode (default setting from the factory for Server cards, as reported in some other threads I have seen).
nvidia-smi has options for enabling or disabling ECC (as does nvidia-settings), but they don’t do anything AFAICT. I may be wrong, but I believe ECC cannot be disabled as it is how the hardware works, but only reports error counts when in compute mode in current driver versions.