Memory Disaggregation

Memory disaggregation is the architectural vision of treating memory as a pooled, network-attached resource rather than a component bound to a specific processor. In a disaggregated system, a CPU or GPU acquires memory capacity dynamically from a shared pool via high-speed interconnects like CXL, paying latency for capacity rather than being constrained by the memory physically attached to its package. This model inverts the traditional assumption that memory locality is paramount, raising the question of whether future systems will optimize for locality at all, or whether the network will become the new memory bus.