How does this memory model change with Nvidias addition of unified memory?
Metalbird
@rohany so from what I can tell, with the new unified memory assuming that you are using at least CUDA 8 (whatever the newest version is), and you have access to a pascal architecture NVIDIA GPU, you can have access to memory from both the CPU and the GPU! This is super useful in some contexts because transferring data between the CPU and the GPU can be a bottleneck in a lot of problems, cudaMemcpy, while now you should be able to get around that issue.
So I think the way the memory model changes is that you are able to use more of a shared address space approach as opposed to using a message passing, where you have to rely on transferring data through the PCI-E between the device and the host, and now you can just access a shared address space from either the device or the host. Again the main reason this matters is that transferring data is usually the slow step so instead of waiting for messages to pass, now you will have to make sure that you are properly managing your data and not violating data dependencies because of the shared memory system.
How does this memory model change with Nvidias addition of unified memory?
@rohany so from what I can tell, with the new unified memory assuming that you are using at least CUDA 8 (whatever the newest version is), and you have access to a pascal architecture NVIDIA GPU, you can have access to memory from both the CPU and the GPU! This is super useful in some contexts because transferring data between the CPU and the GPU can be a bottleneck in a lot of problems, cudaMemcpy, while now you should be able to get around that issue.
So I think the way the memory model changes is that you are able to use more of a shared address space approach as opposed to using a message passing, where you have to rely on transferring data through the PCI-E between the device and the host, and now you can just access a shared address space from either the device or the host. Again the main reason this matters is that transferring data is usually the slow step so instead of waiting for messages to pass, now you will have to make sure that you are properly managing your data and not violating data dependencies because of the shared memory system.