[tylerhou]

For a strictly increasing array, assuming the branch predictor always predicts correctly (guesses taken), the CPU does not have to wait for the test to resolve (v > maxV) before it executes the assignment (maxV = v). So effectively the CPU is just running

  maxV = v
  maxV = v
  maxV = v
  maxV = v

over and over (with tests running in parallel as well, but those matter less, because they are mostly just "checking" whether the branch predictor guessed correctly.).

Then, in that above sequence, there is no /read/ of maxV after each write, so there is no true data dependency. (We don't need to wait for the last write to finish before we write again, unlike the read after write case.)