CPU performance is not only about arithmetic throughput. Modern workloads spend a large part of their time waiting for data, so the path from a load instruction to the memory subsystem can dominate execution time. That is why memory-latency measurement is useful: it exposes behavior that a CPU-only benchmark can hide.
A latency benchmark should distinguish access patterns and working-set sizes. Small working sets may remain in a cache, while larger ones spill into deeper cache levels or main memory. Sequential and random access can also produce very different results because the hardware can prefetch some patterns more effectively than others.
Latency and bandwidth are related but not interchangeable. Latency asks how long the system takes to service an access under a given pattern. Bandwidth asks how much data can be moved over time. A system can show strong aggregate bandwidth while still delivering poor latency for pointer-heavy or irregular workloads.
The important engineering step is interpretation. Benchmark results should be tied to a workload, architecture, compiler settings, and measurement method. The number is evidence about a system under stated conditions — not a universal score for the machine.