unified memory

**Unified memory** is the **shared virtual memory model that allows CPU and GPU to access a single logical address space** - it simplifies programming by automating page migration, but performance depends heavily on access locality. **What Is Unified memory?** - **Definition**: Managed memory system where runtime migrates pages between host and device memory on demand. - **Ease-of-Use Benefit**: Developers can avoid manual memcpy choreography for many workflows. - **Migration Behavior**: Page faults trigger data movement over interconnect such as PCIe or NVLink. - **Risk**: Poor locality can cause page thrashing and severe slowdown under repeated bidirectional access. **Why Unified memory Matters** - **Development Productivity**: Reduces complexity for prototypes and irregular data-structure workloads. - **Memory Flexibility**: Can handle datasets larger than device memory through managed paging. - **Portability**: Unified programming model simplifies code maintenance across hardware tiers. - **Operational Simplicity**: Fewer explicit transfer paths reduce integration bugs. - **Selective Utility**: Useful in targeted scenarios where convenience outweighs migration overhead. **How It Is Used in Practice** - **Access Pattern Planning**: Design for locality so most accesses occur from one processor side at a time. - **Prefetch Hints**: Use managed-memory prefetch APIs to move pages before compute phases. - **Profiling**: Track page-fault counters and migration volume to catch thrashing early. Unified memory is **a productivity-focused memory model with locality-dependent performance** - when migration behavior is managed carefully, it can simplify complex host-device workflows.