dana@locus.com (Dana H. Myers) (01/11/91)
Last night I posted that CMOS power consumption is directly proportional
to frequency of operation. This was in the context of how hot a 386
runs at 25Mhz compared to 16Mhz. In summary, I said a 386 at 25Mhz will
consume roughly 150% of the power it does at 16Mhz (i.e. 50% more).
I'd like to revisit that statement, with some more information.
Technical literature from National Semiconductor and Motorola both
explain that the power consumption of a CMOS gate directly proportional
to frequency of operation. An item generally ignored in the case
of a simple gate is the leakage current - this is the static
current drawn by the gate, and is usually quite small (under 1uA).
For completeness, or when dealing with a lot of gates, this leakage
can not be ignored. So, the power consumption of a CMOS part is
actually:
Pd = Pstatic + K * F
Where:
K = is some constant in ma/Mhz
F = frequency of operation
One person wrote to me saying that relationship of power consumption
to frequency is highly linear in the parts he tests but that it does not
cross 0 at the origin; he then asserts that I was wrong in saying a
386 at 25 Mhz will consume roughly 25/16 the power it consumes at 16 Mhz.
Shucks, why do you think I used the term "roughly" ?
Citing some real numbers, I have a graph of Icc vs. Frequency for an
80376, which is a simplified variant of of the 80386 intended for use
as an embedded processor. Granted, it is not a 386, but it is close
enough for the sake of this discussion. Looking at the graph, one
can see the linear Icc vs Frequency relationship, but the graph only
spans 4Mhz to 16Mhz. Using some algebra, let's determine the static
and dynamic factors...
At 4 Mhz, Icc = 125 mA, at 16Mhz, Icc = 295 mA.
and:
Icc = Istatic + K * F
So:
(295 - 125) mA = (Istatic + 16 Mhz * K) - (Istatic * 4 Mhz * K)
170 ma = 12 Mhz * K
K = 14.166 ma/Mhz
Also:
Istatic = 125 ma - 4 Mhz * K
Istatic = 68.333 mA
At 25 Mhz:
Icc = 68.333 ma + 25 Mhz * 14.166 ma/Mhz
Icc = 422.5ma
I25Mhz/I16Mhz = 422ma/295ma = 1.432
Certainly, for the 376, the relationship I'd first suggested, about
150% or 1.5, is fairly close to the 1.43 we derived from typical data.
I doubt the 386 differs significantly from this.o
One question; where is that 68.333 of static current going? I'd
suggest two places. One may be in on-chip bias voltage generators,
though I suspect this is small. The other place is likely to be
leakage current. Assuming there are 120,000 gates in the 376 (just
guessing, but certainly within less than an order of magnitude),
this would amount to a leakage current of about 500 nA per gate,
which sounds a little high but not too far off.
So there. I WAS right. :-) (and I DON'T do hardware for a living :-)
--
/*
* Dana H. Myers KK6JQ | Views expressed here are *
* (213) 337-5136 | mine and do not necessarily *
* dana@locus.com | reflect those of my employer *