Non-linear Equations Are Our Friends.

If subsystem performance were largely linear in terms of its design parameters we would only get limited insight from high-level models.

Fortunately, system performance is non-linear and trends can be easily observed even if coefficients are inaccurate.

Obvious examples are

• Using a memory hierarchy: e.g. dont have an enormous fast cache, instead use a fast L1 and a large L2 to get same hit rate at almost the same latency but with far less energy use.
• Use multiple cores when you can: a high-performance (Intel Nehalem style) core consumes far more power than a set of simple cores with similar total throughput.

Interesting paper: Andre DeHon in FPGA 2013 Location, location, location: the role of spatial locality in asymptotic energy minimization. »PDF