9.1 WHAT DOES ENERGY GET DISSIPATED FOR IN CMOS CIRCUITS use software code 39 encoding torender barcode 3 of 9 on software Web app Again, this overall barcode 3/9 for None gure is meant per computation cycle and averaged over many cycles. Equation (9.10) has two bene ts.

Firstly, it quanti es what almost always is the most signi cant contribution to a VLSI circuit s overall energy dissipation. Secondly, the few arguments on its righthand side are fairly straightforward to come up with. Remember that it has been obtained under ve simple assumptions: All node capacitance values Ck are xed.

The switching devices (MOSFETs) are of resistive nature when on . Supply voltage Udd is constant and the same throughout the (sub)circuit considered. Nodes are always fully charged and discharged (full swing from rail to rail).

All activity numbers refer to the same computation cycle and clock signal.. Several observations about node activities are due. 1. A circuit node s activity is not the same as the probability of nding that node in the opposite logic state at the end of a computation period as digital circuits are subject to glitching.

Investigations on various adder structures fed with random numbers have resulted in activities 10% to 20% above what is anticipated with glitching ignored. While that much extra unrest is typical for many circuits, data activities can even grow beyond the intuitive bound of k 1. Node activities in excess of 6 have been reported in circuits where signals propagate along paths of markedly di erent depths before converging in combinational operations.

Unbalanced delays occur in multipliers and even more so in cascades of multipliers or other arithmetic units with no registers in between. As an example, consider the isomorphic architecture of a lattice lter in g.9.

2. How to mitigate this e ect by way of delay balancing and signal silencing will be a subject of section 9.2.

2.. 3.29 > 2.14 3.29 > 2.14 DecIn1 DecIn2 A A B D 1.15 4.28 > 2.

25 4.28. B A B E C A B E C A D 8.36 > 7.25.

F D D k1 G k2 G E E k3 G k4 4.28 > 2.25.

A 1.15 B: 0.5 < < 1 E A A D: 1.5 < < 2 E: 2 < < 3.5 F: 3.5 < < 5 G: 5 < Saturation and Truncation A: < 0.5 C: 1 < < 1.5 Fig. 9.2 In ation of node activities in a lattice lter (reprinted from [238]).

. 2. Node activities ar Software bar code 39 e statistical data. They are typically obtained during gate-level simulation runs, see g.

9.3.6 For every circuit node, the toggle counts get collected over what is believed to be a representative sequence of operation for the circuit being analyzed.

. The alternative of probabilistic p ower estim ation ignores glitching and is less accurate. Design of VLSI Circuits stimuli vectors gate-level simulation run extracted netlist accumulated toggle counts estimated power power computation gate-level netlist node capacitances Fig. 9.3 Simulation-based power/energy estimation. 3. Almost all circuit s exhibit signi cant temporal (over time) and spatial (across signals and bits) correlations among the toggling of their nodes. This is because most ip- ops are part of registers and because many registers are updated on a regular basis.

Also, activities may greatly di er among the bits of a data word. Consider an accumulator, for instance, and watch how bits evolve from one computation cycle to the next. LSBs will typically behave like random variables whereas MSBs will be strongly correlated.

. Numerical examples Statistical data coll ected from the entire population of ip- ops in six benchmark circuits that have been fed with representative stimuli. Benchmark circuit ARES Bongo FIR SST Shiva CCDChip Average node activities D inputs Q outputs 0.178 0.

076 0.50 0.40 0.

54 0.38 0.066 0.

054 0.30 0.11 0.

24 0.26. Figure 9.4 plots the Software bar code 39 average bitwise activities of a noisy speech signal quantized with 16 bit resolution, 16 kHz sampling rate, and SNR = 40 dB. Note the huge di erence between MSB and LSB.

Also observe that 2 s complement (2 C) encoding entails higher node activities than sign-and-magnitude (S&M) representation in signals such as speech that uctuate around zero for much of the time with modest amplitudes.. In more general terms , the nding is Observation 9.2. Node activities are distributed very unevenly.

Circuits typically include a number of ip- op outputs that toggle with k 1 while most other circuit nodes exhibit signi cantly lower 2 activities. Nodes with an activity close to zero are also common..

Copyright © . All rights reserved.