Does cuGraph have UVM support?
Yes, through RMM (rapids memory manager) we support use of managed memory for both single-GPU and multi-GPU workflows. Obviously, performance with managed memory will not be as high as performance with device memory. In addition, cuDF has a spilling feature that can be very helpful when working with RMM pools, which are pre-allocated chunks of memory. And finally, for multi-GPU workflows, dask has its own spilling features that can be configured. There is also additional research going on for further expanding use of host memory in GPU workflows.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.