At the recent Si2 conference, there was an interesting presentation by IBM’s David Hathaway on what is hoped to be a better way of approaching power modeling at the technology level.
He said that power modeling can be approached differently from delay modeling. With delay, there are numerous effects that combine in complex, non-linear ways, and so a full characterization of each cell is necessary. But with power, because interpolation is risky, many more points are needed, making full characterization a really time-consuming chore.
The good news, he proposed, is that the elements contributing to power can be separated out as more or less orthogonal to each other. Specific power contributors can be isolated, and then each cell can be defined in terms of its contributors. Only the contributors have to be characterized (tens of tests rather than hundreds), and then they can be summed cell by cell.
In an experiment to test this theory out, they compared the calculated value with full-up actual values. 95% of the simulations that would have normally been needed were eliminated, and the average error was 0.073%, with the worst-case error being 3.64%.
There’s more work to be done both at the dynamic and leakage level, but it felt like there’s some promise to this approach, with the potential of making it easier to create new technology models.