I think managed memory is very useful. I’m wondering how to use managed memory for multi GPU cards. please kindly advise me
To quote https://devblogs.nvidia.com/parallelforall/unified-memory-cuda-beginners/ for Pascal and later devices, “it just works”.
Unified Memory transparently enables oversubscribing GPU memory, enabling out-of-core computations for any code that is using Unified Memory for allocations. It â€œjust worksâ€� without any modifications to the application, whether running on one GPU or multiple GPUs.
For pre-Pascal architectures see: http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#um-advanced-topics-hd