Nano-CMOS MixedCMOS Mixed-Signal CircuitSignal Circuit ...Create a Metamodel Metamodel Perform...
17
Nano Nano-CMOS Mixed CMOS Mixed-Signal Circuit Signal Circuit Nano Nano CMOS Mixed CMOS Mixed Signal Circuit Signal Circuit Metamodeling Metamodeling Techniques Techniques: : A A Comparative Study Comparative Study O G it l 1 S PMh t 2 E K i 3 P Pt 4 O. Garitselov 1 , S. P. Mohanty 2 , E. Kougianos 3 , P. Patra 4 NanoSystem Design Laboratory (NSDL, http://nsdl.cse.unt.edu), University of North Texas Denton TX 76203 USA 1,2,3 University of North Texas, Denton, TX 76203, USA. Intel Architecture Group, Intel Corporation, USA. 4 E-mail: [email protected] E mail: [email protected] Presented By: Garima Thakral University of North Texas Acknowledgment: This research is supported in part by SRC award P10883 and NSF awards CNS-0854182 and DUE-0942629. University of North Texas 12/21/2010 Mohanty 1
Transcript of Nano-CMOS MixedCMOS Mixed-Signal CircuitSignal Circuit ...Create a Metamodel Metamodel Perform...
NanoNano--CMOS MixedCMOS Mixed--Signal CircuitSignal CircuitNanoNano CMOS MixedCMOS Mixed Signal Circuit Signal Circuit MetamodelingMetamodeling TechniquesTechniques: :
A A Comparative StudyComparative StudyO G it l 1 S P M h t 2 E K i 3 P P t 4O. Garitselov1, S. P. Mohanty2, E. Kougianos3, P. Patra4
NanoSystem Design Laboratory (NSDL, http://nsdl.cse.unt.edu), University of North Texas Denton TX 76203 USA 1,2,3University of North Texas, Denton, TX 76203, USA. , ,
Intel Architecture Group, Intel Corporation, USA.4
E-mail: [email protected] mail: [email protected] Presented By:
Garima ThakralUniversity of North Texas
Acknowledgment: This research is supported in part by SRC awardP10883 and NSF awards CNS-0854182 and DUE-0942629.
University of North Texas
12/21/2010 Mohanty 1
Outline of the talkOutline of the talk Introduction Fast Design Exploration through Metamodeling Metamodeling vs. Macromodeling Metamodeling vs. Macromodeling Metamodeling based Design Flow
S li T h i Sampling Techniques Ring Oscillator Example Circuitg p “Golden” Surface
E i t l R lt Experimental Results Conclusions and Future Research
12/21/2010Mohanty 2
IntroductionIntroduction Complex computations for analog circuits to
IntroductionIntroduction
include parasitics Physical layout and simulation analysis is very y y y y
costly processes in design flow Metamodeling is mathematical formula that Metamodeling is mathematical formula that
represents circuit’s behavior within a given range i li i tusing sampling points
This paper targets sampling techniques which are technology independent and the amount that is needed to create an accurate metamodel
12/21/2010 Mohanty 3
Fast Design Exploration Through Fast Design Exploration Through g p gg p gMetamodelingMetamodeling
a. Traditional Slow Approach
b Metamodeling Based Fast Approachb. Metamodeling-Based Fast Approach
12/21/2010 Mohanty 4
MetamodelingMetamodeling vs. vs. MacromodelingMacromodelinggg gg Metamodeling mathematical representation of output prediction equation prediction equation can be used in different tool like MATLAB Macromodeling simplified version of the circuitsimplified version of the circuit used in same simulation tool
h d hard to create
12/21/2010 Mohanty 5
Metamodeling Design Flowg gRegular design flow is
altered for usingCreate Logical Design
Input Specifications
Schematic
No Specifications
Create Logical Design
Input Specifications
Schematic
No Specifications altered for usingmetamodels.
Advantages for usingCreate Physical Layout Layout
Perform DRC/LVS/RCLK Extraction
Yes
No
Netlist with Parasitics
Specifications met?
Create Physical Layout Layout
Perform DRC/LVS/RCLK Extraction
Yes
No
Netlist with Parasitics
Specifications met?
Advantages for usingmetamodels: Reusability
DONEYes
No
met?Specifications
Parameterized Parasitic Netlistwith Design Variables
Parasitic AwareParameterized Netlist
DONEYes
No
met?Specifications
Parameterized Parasitic Netlistwith Design Variables
Parasitic AwareParameterized Netlist
Reusability Speed Accuracy
Create a Metamodel Metamodel
Perform Design OptimizationUsing the Metamodel
Optimized PhysicalDesign Variables
Create a Metamodel Metamodel
Perform Design OptimizationUsing the Metamodel
Optimized PhysicalDesign Variables
AccuracySpecifications met?
Yes
Create Physical Design
SimulationPerform DRC/LVS/RCLK Extraction Final Optimized Layout
No Specifications met?
Yes
Create Physical Design
SimulationPerform DRC/LVS/RCLK Extraction Final Optimized Layout
No
Physical design is doneonly 2 times in the
Specifications met?
DONE
No
Yes
Specifications met?
DONE
No
Yes
only 2 times in theproposed design flow.
12/21/2010 Mohanty 6
DONEDONE
Ring Oscillator: 45nm CMOS Design
pNtf
21
Where - frequency of oscillations, N - number of inverters, - delay of each inverterpt
f
I1 I2 I3 Output
Vdd
Wp=240nmL=45nm
Wp=240nmL=45nm
Wp=240nmL=45nm
Output
L=45nmWn=120nmL=45nm
Wn=120nmWn=120nmL=45nm
Gnd
12/21/2010 Mohanty 7
Ring Oscillator: CharacterizationgEye Diagram for Parasitic Aware Netlist Simulation
12/21/2010 Mohanty 8
Sampling Techniques Exploredp g q p Exhaustive – evenly distributed large amount of
samples Monte Carlo – random samplingp g Latin Hypercube – random sampling within each
Latin squareLatin square Middle Latin Hypercube – middle point sampling
within each Latin square Design of Experiments – min, mid, max samplingg p , , p g
for each parameter
12/21/2010 Mohanty 9
Sampling Techniques: Applied to Ring Oscillator Circuit
MLHSMonte Carlo
LHS DOE
12/21/2010 Mohanty 10
“Golden” SurfaceFrequency without Parasitic
1.2
1.3
x 1010 10,000 sampling surface
was created forexhaustive analysis
1
1.1
1.2 exhaustive analysis.
-RMSE < 0.01%
80.8
0.9
“Golden” Surface is used
1
2
3
4
x 10−7
0
2
4
6
8
x 10−7
WnWp
as actual results tocompare to the samplingtechniques10
WnWp techniques.
12/21/2010 Mohanty 11
The The MetamodelsMetamodels The generated sample data can be fitted in
many ways to generate a metamodel.y y g The choice of fitting algorithm can affect the
accuracy of the metamodelaccuracy of the metamodel. Metamodel has the form:
Where y is the response being modeled (e.g.y p g ( gfrequency), x = [Wn,Wp] is the vector ofvariables and ij are the coefficients.variables and ij are the coefficients. The polynomial regression determined k = 2
for DOE and k = 4 for other cases12/21/2010 Mohanty 12
for DOE and k 4 for other cases.
Error Analysisy
N
RMSE 2))(ˆ)((1
k
kk xyxyN
RMSE1
2 ,))()((
N
kk RMSExyxy 2 ,))(ˆ)((1 k
kk yyN 1
,))()((
Where:
is the set of parameterskx
Where:
is the predicted equation
is the number of sampling points
y
N
12/21/2010 Mohanty 13
Experimental Results10
8RMSE for MLHS
108
RMSE for MC
107
Fre
quen
cy(f
)
107
Fre
quen
cy(f
)
101
102
103
104
106
Number of Simulations
101
102
103
104
106
Number of Simulations
108
RMSE for LHS
107
Fre
quency(f
)
101
102
103
104
106
Number of Simulations
Fr
12/21/2010 Mohanty 14
Number of Simulations
Conclusions A design flow for metamodeling is
proposedproposed A 45nm ring oscillator was subjected to
the proposed design flow Uniform sampling techniques has better Uniform sampling techniques has better
performance than DoE or thanrandomizedrandomized.
Designers should choose LHS or MLHSgover MC but the trend in typical designenvironments is the opposite.environments is the opposite.
12/21/2010 Mohanty 15
Future Research
Our future research will include the Our future research will include theoptimization techniques part of the proposeddesign flow We will conduct thorough studydesign flow. We will conduct thorough studyby using metamodel optimization for higheramount of variables by the example of a morecomplex nano-CMOS circuit.p
12/21/2010 Mohanty 16
Thank you !!!y
12/21/2010 17
