Do you have access to Nvidia hardware to test the behavior ? Even though the GPU driver does the allocation of memory for the graphics resources you mentioned...driver memory management is tightly coupled with the OS memory management scheme. The GPU driver is also able to allocate memory in the host address space and thus have to negotiate memory transaction with the operating system. Also is it possible that memory is being allocated at a faster rate than its being de-allocated. Keep in mind that if there are pending transaction referencing the memory region, they will not be freed immediately or this may even force the driver to clone the memory internally before freeing..It would be interesting to see if this behavior can be replicated on other HW..
↧