Jump

When a loop's branch instruction crosses a 32-byte aligned boundary, the instruction fetch can be significantly slower due to how the frontend decodes instructions. With 15 NOPs, the loop body is larger and the backward jump crosses an alignment boundary, causing extra cycles.

Usually not a concern in practice, as the compiler handles alignment. Note that this benchmark can be unstable since instruction addresses may change between builds.

References:

• Intel Optimization Manual - Code Alignment
• Agner Fog - Microarchitecture