My problem with the memorization approach is that the technology is still moving quickly enough that the orders of magnitude still "wiggle," particularly in the middle numbers.

For example, at a glance, the disk seek numbers are off by orders of magnitude for SSD drives. And the 1MB of RAM read numbers appear not to have used the SSE streaming load/store instructions. I don't know anything about datacenters, but I'd certainly want to measure that one myself before building a system assuming that number was in its correct place in the hierarchy.