From the 05/07/2025 document "MC/DC Plans Through the End of CEMeNT" :
MI300A unified memory space allows for a single copy of GPU data to be shared between the host (CPU) and device (GPU). Enabling this functionality would result in dramatic memory savings (upto 2X) and would enable us to run our challenge problems on fewer nodes
The very same functionality that would support the conveyance of larger data structures to the GPU would additionally facilitate the use of unified memory on MI300A.
From the 05/07/2025 document "MC/DC Plans Through the End of CEMeNT" :