I have been attempting to set the GPUMemoryTransferRateOffset. I have already set coolbits to 31. The command ‘nvidia-settings -a “[gpu:0]/GPUMemoryTransferRateOffset=x”’ and variants result in ‘Unknown Error’. Setting the offset in GUI appears to work; the change is accepted without error and ‘nvidia-settings -q “[gpu:0]/GPUMemoryTransferRateOffset”’ gives the new value. However, any value other than 0 causes nvidia-smi to report a memory frequency of ‘N/A’.
I have tested memory offsets in Windows 10 with MSI Afterburner and hwinfo, and no problems occur. The memory offset is applied, power usage increases slightly, and the new frequency is correctly reported by monitoring tools. I believe the offsets are also taking effect under linux as, using the decidedly non-scientific method of running a demanding game, a memory offset of 1000 increases average power draw from slightly under to slightly over 330 watts. Is this just a problem with nvidia-smi?
I am running the Manjaro distribution with kernel 5.1.15-1, nvidia drivers 430.26-6, and cuda 10.2 or 10.1.168-4. (pacman and nvidia-smi give different results for cuda version.) My graphics card is a MSI Ventus OC 2080ti (UUID GPU-1a644d63-ae0e-bb53-62a7-efb38b82d293).
nvidia-bug-report.log (1.78 MB)