A code efficiency improvement of 5% (represented as a 5% improvement in CPU utilization) does not equate to a 5% improvement in equipment power draw though.
You at least need to factor in the percentage of an individual server's power draw that goes to the CPU, which is about 30% (again, on average).
So a 5% equipment power draw improvement would need at least a 15% code efficiency improvement (0.05 / 0.3 = 0.1666), assuming you're CPU bound on whatever your code is doing. I factored in the average CPU load (50%) to account for this, but it's heavily application dependent.