CUDA Heterogeneous Memory Management (HMM) Support

How will the CUDA 12.2 HMM impact my A100-based system?

A good way to think about HMM is that it’s like UVM, but better. So where UVM gives you the ability to do things like using managed memory to make data migration easier, oversubscribe memory, etc. when you allocate memory through specific UVM APIs like cudaMallocManaged(), HMM gives you all of that with standard malloc() / free(). It makes it so much easier to work with large applications and libraries that might have their own internal memory management, but still use your GPU for acceleration.

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.