I am using three titan rtx.
I want to get a stable performance.
So, i tried to set gpu, application clocks and memory clocks with nvidia-smi.
gpu and application clocks setting command are fine but memory clocks does not work.
Is there any solution to set gpu memory clocks?
or Is there any reason why is it not supported?
I have not used this particular GPU, but on various recent GPUs the memory clock does not appear to change dynamically for any performance state, and in particular the highest performance state (typically P0). This is in contrast to the GPU core clock that is adjusted dynamically in the highest performance state based on environmental factors, in particular temperature and power draw.
My hypothesis is that the memory clocks have been carefully tweaked to optimal values in the context of the board design (e.g. trace routing) and that allowing the memory clocks to change would lead to an increase in instability. So now memory clocks are fixed for each power state and only a few clock frequencies have to be qualified for a particular SKU.
Based on my experience with older GPUs where memory clocks could be changed, adjusting memory clocks is much more finicky than adjusting core clocks, in that stable operation is much more quickly compromised with the former.
My second hypothesis is that with GPUs whose memory clocks are already locked by design, an attempt to lock them via
nvida-smi will fail in the manner shown.
These hypotheses are based purely on observation. Your observations may differ or you may disagree with my reasoning.