6 J.R. Smith
Since the MSP430 has very low idle power, and transitions efficiently between
active and idle, this is not a bad approximation for the systems that motivated
this chapter. For energy-constrained systems such as these, the “efficient idling”
approximation will presumably get better over time as increased attention is paid
to reducing idle power. Even if leakage current rises as a fraction of total power
budget, which naively might increase idle power consumption, it will be possible to
gate power to most of the system, thereby keeping idle power low.
Note that energy efficiency values are greater in the microcontroller data than
in the microprocessor data. For example, in 2006, the microcontroller data shows
around 10,000 instructions per μJ, while the plot from [6] only shows about 100
instructions per μJ. Part of the explanation for the discrepancy is that the data
presented in [6] and plotted in Fig. 1 is for complete computer systems (including
memory, support circuitry such as display drivers, etc), whereas Fig. 2 is for the
microcontroller in isolation. Also, the absolute computation rate (in instructions per
second) is not constrained to be the same in the two plots, so energy efficiency
(in instructions per μJ) is not the entire story. Running the microprocessors at the
speeds of the microcontrollers would provide improvements in energy efficiency
but would make the absolute computation rate insufficient for the requirements of
general purpose computing. Finally, the data points in Fig.1 were selected because
they were commercially available computer systems; the data points in Fig. 2 were
selected because they are particularly energy efficient. Despite the differences, it
is interesting that the functional form of the scaling curves is the same, that is, a
straight line on a semilog plot. The exponential energy scaling behavior appears to
be quite generic.
2.2 Scaling of Workload Power Requirements
For the purposes of this chapter, we will consider a 6,000 instruction computation,
executed ten times per second, for a net workload of 60K instructions per second.
This is approximately the computational requirement of the Saturn accelerometer
demo, described in the chapter “The Wireless Identification and Sensing Platform”
[11] of this volume, or the musical input device described in the chapter “RFID-Vox:
a Tribute to Leon Theremin” of this volume [8].
If we use F to denote the energy efficiency of microcontrollers (measured in
instructions per μJ), then a fit yields F ∝ 2
0.46t
or F = c
1
2
0.46t
. For a given workload
W, a fixed number of instructions per second, the power required is given by
P
REQ
=
W
F
. With W = 60,000
Inst
s
,thenP
REQ
=
60,000
Inst
s
c
1
2
0.46t
Inst
μJ
=
60,000
c
1
2
0.46t
μJ
s
. Combining
the constants c
1
and 60,000 into a new constant c,weget
P
REQ
= c2
−0.46t
.
我的内容管理
展开
