I know we can only set CU_EXEC_AFFINITY_TYPE_SM_COUNT affinity when creating a CUDA Context and it will be fixed during the lifetime of it. But I was wondering if there is any hope that this affinity would support dynamic changes after creation. Is there a hardware limitation that does not allow this to happen, or, it’s just a matter of software and could be achieved with a new CUDA update?
Generally speaking, NVIDIA employees don’t make forward-looking statements on topics like this, in these forums. There are various reasons for this. Yes, there are exceptions, but not typically for a question like this.
If you have a feature that you would like to see implemented in the future, the suggestion I have is to file a bug. If you use the word “enhancement” in your description, it will be clear you are not pointing out a defect but requesting a feature.