Accuracy Is Relative

For those trying to tame the power beast, accuracy typically comes up as a necessity for power analysis and optimization. But should it?


I’ve been doing some thinking on the concept of accuracy lately in my article, Does Power Analysis Need To Be Accurate? And, I’ve come to the conclusion that there isn’t a single conclusion to be had. Accuracy is very relative to the task at hand, as well as the use case, and the very specific power requirements. And maybe the focus shouldn’t be always on the accuracy, but on what the changes are that are needed, and how to figure that out.

Interestingly, there aren’t many people in the industry that are not thinking about this — and each person has a different approach.

So where are we headed?

On the tool side, it’s all about helping the engineer.

Drew Wingard, CTO of Sonics believes the way to help designers is to first of all make sure that they are getting very straightforward solutions to implement the ‘cheap model’ — based on things like area, and clock rates, etc. — and then to provide a platform by which they can bring in better knowledge if they have it.

There is a ‘but,’ of course. “What we think is the practical reality today is that so few chips are ground up, blank sheet of paper designs, that maybe the most useful thing to do is to provide instrumentation so that if you use [an Energy Processing Unit], you get an instrumented design that helps you in the lab, at least, gather the information to build that model so on the next revision of the chip, you’ve got a much better version of the data to work with to help make improvements. That’s pretty straightforward to do. That’s being done, and it allows people to go incrementally,” he explained.

It’s really comes down to moving the ball forward.

“While we’re all worried about the day in which we have chips with 150 power gated domains on them, most engineering teams aren’t anywhere near there yet. And, if we can help in the chip that has 10 power gated domains, gathering up data so they can figure out which one of those should they break into 3 for next time; if we can help them figure out how to make those choices based on data gathered off of the current design, then we are moving the ball forward, and those models would begin to come in a way that’s inexpensive, and therefore has the right ROI kind of benefit to it,” Wingard added.

Sonics is obviously not the only player here, and doesn’t want to be. It would seem likely that subsystem-focused IP and APU companies alike are also feverishly working on similar approaches given the urgency to tame the power beast. I will be watching for those approaches along with you.

Leave a Reply

(Note: This name will be displayed publicly)