David Kanter (dkanter@realworldtech.com) on 9/1/11 wrote:
---------------------------
>Let's take your example. Suppose you are doing work where the maximum sustained
>frequency (due to TDP/power) for several cores is actually 1000MHz. You could achieve
>this by oscillating between 600MHz and 1400MHz and paying the cost of adjusting
>frequency, etc. or you could have a better designed system that runs closer to the sustained levels.
>
>Also, keep in mind we are solely talking about turbo here, where you are beyond max frequency.

[...]

>We are talking about turbo bins beyond the base frequency. What I'm saying is
>that for turbo you want to be able to run at {1.2f, 1.4f, 1.6f...) where f is the base frequency.

Aha. Yes, in that case you are totally right. I had
the impression the article meant DVFS in general, not
specifically turbo.

Do you have any latency numbers for switching frequency?
I guess it can happen in a few cycles, to give everything
time to settle, but it's probably more complicated in
reality. Of course: changing the voltage takes more time.
But I suppose the CPU doesn't have to wait for the voltage
to drop when lowering it, only when increasing it?

If voltage is the troublesome part, then it seems there
doesn't have to be much difference between going from slow
to fast in one step, or in many small steps, as everything
has to wait for the voltage to catch up anyway. Or in other
words, changing a small step should be much faster.

---

(Could you please delete my duplicate post below? I hit
"stop" because I wanted to see the preview. I reloaded the
forum to check if the post made it before reposting, but
apparently it takes more time for a post to get through.)