Think of this classic example: a global i integer is being incremented within a concurrent code path, one within which multiple threads of execution can simultaneously execute. A naive understanding of computer hardware and software will lead you to believe that this operation is obviously atomic. However, the reality is that modern hardware and software (the compiler and OS) are much more sophisticated than you may imagine, thus causing all kinds of invisible (to the app developer) performance-driven optimizations.
We won't attempt to delve into too much detail here, but the reality is that modern processors are extremely complex: among the many technologies they employ toward better performance, a few are superscalar and super-pipelined execution in order to execute multiple independent instructions and several parts of various instructions in parallel (respectively), performing on-the-fly instruction and/or memory reordering, caching...