kelsey hudson wrote:
Keep in mind that voltage and current are inversely proportionate and directly relate to how much power the chip consumes: a basic principle of electronics (Ohm's law) states that voltage equals current times resistance;

Did you ever wonder why chips are all trying to lower their voltages given that VI=P and V=IR? It would seem to be a net wash.

*However*, CMOS technologies that we use in modern VLSI chips are, to a first order approximation, purely capacitive loads (not so true as we get down to 45nm, but that's for a different day). They behave a bit differently.

Basically, CMOS gates behave as two switches (one each to Vdd and GND) which either dump charge onto a capacitive load from Vdd or pull charge off of a capacitive load and dump it to ground.

If your grind through a bit of (not too hard) mathematics, you wind up with an interesting result:

P = k C V^2 f where k is a constant of proportionality

Therefore, we have:

Power is linearly proportional to Capacitance

Capacitance is pretty much set by technology and size of chip. Not a lot you can do.

Power is linearly proportional to Frequency

This is why we don't see a lot of 4GHz chips floating around. They burn twice as much power as 2GHz and the overhead of trying to run something that fast eats into the time to do actual work. So you burn twice the power but don't quite get twice the performance. There is a "sweet spot" where 60-70% of your on chip logic arrives "just in time". That pretty much sets your optimally efficient performance and power.

Power is *quadratically* proportional to Voltage

This is *huge*. If you manage to halve your voltage, you cut your power to 1/4 of the original consumption. However, running chips at lower voltages hurts performance in various ways, so, again, there is a point at which you stop going much lower. For now, the limit appears to be about .8V. Much lower than that and you have to start doing some major tricks to keep your performance up.

And now you know ...

-a





--
[email protected]
http://www.kernel-panic.org/cgi-bin/mailman/listinfo/kplug-list

Reply via email to