Register-Transfer Level Estimation Techniques for Switching Activity and Power Consumption
Anand Raghunathan, Sujit Dey, and Niraj K. Jha
Outline Motivation RTL power estimation methodology Estimating glitching activity at the RT level RTL power models Experimental results
Motivation
1.64mW with CSIM 1.32mW with no gltich (19.5% under estimate) 1.53mW with glitching activity and power estimation
RTL power estimation methodology First phase
Word- and bit-level Mixed-level High speed
Second phase Zero-delay + gate-delay
Third phase Build power models Bit-level
Estimating glitching activity at the RT level Models for glitch generation Propagation through various components Two sub sections
Glitching activity at the control signals Glitch generation and propagation in data path
Glitching activity at the control signals Using control expressions
Glitching activity at the control signals-Example 1 S3 to S4
x4↑and x3↓ If x4 is later than x3
1-0-1 on contr[1]
Two conditions Logic
Correlation between rising and falling transitions
Simultaneous occurrence
Temporal Related to controlling value
011
100
Glitching activity at the control signals-Example 1 (cont.) Logic conditions
At least one rising and at least on falling At the gate’s input
No input assume a steady controlling value Temporal conditions – AND for example
Earliest falling after the latest rising An interval great than the gate’s inertial delay
Glitching activity at the control signals-Example 1 (cont.) Glitch counter for each product term
Increment when two conditions satisfied In each cycle
Not easy to know accurate manner Pessimistic assumption
Whenever the logic conditions Leads to over-estimates of glitching
Glitching activity at the control signals-Example 2 contr[2]=x0+x1C11+x3C10
Logic conditions Case 1: Count(x1↓,C11↑)=15
Case 2: Count(x1↑,C11↓)=20
Case 3: Count(x3↓,C10↑)=35
Case 4: Count(x3↑,C10↓)=30
100 estimated glitches; while 52 by CSIM
Glitching activity at the control signals-Example 2 (cont.) Partial information
Toutputs of comparators > Tpresent state signals
Divide input into three group Early Late Unknown ─ pessimistic approach
Glitching activity at the control signals-Example 3 contr[2]=x0+x1C11+x3C10
Logic conditions Case 1: Count(x1↓,C11↑)=15
Case 2: Count(x1↑,C11↓)=20
Case 3: Count(x3↓,C10↑)=35
Case 4: Count(x3↑,C10↓)=30
50 estimated glitches; while 52 by CSIM
Modeling glitch generation and propagation in data path blocks Glitch generation and propagation in multip
lexers
Mux_gl_gen[] Ai(t), Bi(t), Seli(t), Ai(t-1), Bi(t-1), Seli(t-1)
1
( ) _ ( ) _ Pr _
_ Pr _ _ Pr _
n
ii
i
Gl OUT Gl Gen i Gl op A
Gl op B Gl op Sel
RTL Power Models Datapath Controller
Datapath Power Models Bit-level signal Transition probabilities Correlations Glitching activity
Controller Power Models Assume
2-input AND 2-input OR Inverter
Experimental results