occupancy

**Occupancy** is the **ratio of active warps on an SM relative to its architectural maximum capacity** - it estimates available parallelism for latency hiding, but optimal performance depends on more than occupancy alone. **What Is Occupancy?** - **Definition**: Active-warp fraction determined by block size, register use, and shared memory allocation. - **Resource Limits**: High per-thread register or shared-memory use can cap active blocks and warps. - **Not Absolute**: Maximum occupancy does not guarantee maximum throughput if kernels are compute-bound differently. - **Measurement**: Reported by profilers alongside issue efficiency and stall breakdown. **Why Occupancy Matters** - **Latency Hiding**: Higher occupancy often helps mask long memory and synchronization delays. - **Launch Tuning**: Occupancy analysis guides block-size and resource tradeoff decisions. - **Performance Diagnosis**: Low occupancy can explain underutilization in memory-sensitive workloads. - **Portability**: Occupancy-aware kernels adapt better across GPU generations with different limits. - **Optimization Balance**: Helps choose between aggressive unrolling and resident-warp count. **How It Is Used in Practice** - **Kernel Resource Audit**: Measure register and shared-memory usage per thread block. - **Launch Sweep**: Benchmark multiple block dimensions to find best throughput and occupancy balance. - **Combined Metrics**: Interpret occupancy together with memory and instruction-efficiency counters. Occupancy is **a key parallelism indicator for GPU kernel tuning** - best results come from balancing occupancy with instruction efficiency and memory behavior, not maximizing one metric blindly.