This is true that the latency of each operation won't change much as we increase concurrent accesses, but the amortized latency would (be expected to) reduce as

This is true that the latency of each operation won't change much as we increase concurrent accesses, but the amortized latency would (be expected to) reduce as the graphs show. When we did benchmarks, the Go benchmark framework provides us with ns/ops which we directly used to plot the graphs. We could calculate ops/sec too but that would just be another step of calculation using throughput = 1/(time taken/ops), may be more clear to understand though. The graphs would still flatten out after 20 concurrent accesses.