Transcript talk12-1

University of Toronto
Minimization of Delay
Sensitivity to Process
Induced Vth Variations
Georges Nabaa
Farid N. Najm
University of Toronto
(georges,najm)@eecg.utoronto.ca
Outline
Introduction
 Problem formulation and goals
 Methodology
 Standard simulations







Static gates
Dynamic gates
Transmission gates
Standalone simulations
Sizing simulations
Conclusion
Nabaa-Najm
NEWCAS-05
2
Introduction

The threshold voltage is a fundamental operational
parameter of a MOSFET

For the past 30 years, performance improvements in
semiconductors have been achieved by decreasing
channel length

This decrease had to be accompanied by a


Decrease supply voltage

Decrease threshold voltage (Vth)
This Vth decrease has not been followed by a
corresponding decrease in threshold voltage variations
Nabaa-Najm
NEWCAS-05
3
Random Dopant Fluctuations

Threshold Voltage is a
function of the dopants in
the channel

Due to the decrease in the
number of dopants in DSM
processes there is increased
variability
Nabaa-Najm
IBM: ISSCC 2004
Threshold Voltage Variations

Threshold voltage variations
(δVth) cause variations in
circuit delay that impact the
chip timing yield


Can cause up to 30%
variation in chip frequency
[BKN02]
Threshold Variations can be
divided into
Within-die
 Die-to-die

Tschanz 2002
Nabaa-Najm
NEWCAS-05
5
Problem Formuation

Previous work applies chip wide compensation schemes




Unsuitable for the within-die component
Within-die variations become larger as the feature length gets
smaller
We study design techniques that minimize the effects
of threshold variations on circuit delay variability
(minimize delay sensitivity)
Specifically, we explore:



Topology issues, e.g., series vs. parallel arrangements
Design style, e.g., static vs. dynamic (NAND vs NOR)
Optimization issues, e.g., sizing
Nabaa-Najm
NEWCAS-05
6
Goals



Evaluate these styles based on performance penalty,
area overhead, and delay variability minimization
This per gate approach tackles within-die variations
intrinsically
Design δVth aware Libraries
Nabaa-Najm
NEWCAS-05
7
Methodology

We model Vth variations (δVth ) as normally distributed random variables
(RVs)
The 3σ limits of the Normals are from the technology files
 Adjusted using the Law of Area: (Horstmann99)

 vth  A


WL
The larger the transistors, the smaller the input Vth variations
For all the transistors in a given logic gate consider δVth variations as:
Independent Normals (n transistors -> n independent normals)
 Fully Positively Correlated Normals

Nabaa-Najm
NEWCAS-05
8
Methodology (cont)
Used 0.13um UMC process
 Generated 1000 sweep points and link it to the
DELVTO parameter in SPICE
 Run the simulation and record propagation delay



Absolute delay is input dependent
For each gate we choose the worst-case input vectors
Nabaa-Najm
NEWCAS-05
9
Methodology


Normal Score for Output Delay (NAND)
Assume Linearity between
process and delay
From each sweep, the
sensitivity is recorded as:
Delay  3
4
3
2
1
0
-4
-3
-2
-1

-1
0
1
2
3
4
-2
-3
0.045
0.04
0.035
0.03
0.025
Normalized Probabilty Delay
Normalized Probability Vt
-4
0.02
0.015
0.01
0.005
0
1
9
17
Nabaa-Najm
25
33
41
49
57 65
73
81
89
97
NEWCAS-05
0.045
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
1
9
17
25
33
41
49
57 65
73
81
10
89
97
Static Gates: Series better than
Parallel

Series stacks exhibit less delay sensitivity than their
parallel counterparts.

Explanation: body-effect minimizes the impact of Vth

Design: insert series transistors to create series stacks
18
16
Series
Parallel
Delay Variations (%)
14
12
10
8
6
4
2
0
Nabaa-Najm
Independent
NMOS
Correlated
Independent
NEWCAS-05
PMOS
Correlated
11
Hybrid Gates


The fact that series are better than parallel led us to insert a
“serializing” dummy transistor into the structure of a gate
For a 2-input NAND gate, two potential configurations:
Standard
Nabaa-Najm
Configuration 1
NEWCAS-05
Configuration 2
12
Hybrid gates: Independent

Hybrid gates exhibit less delay variability
than standard gates
Delay Variations (%)
18
16
14
12
Original
Hybrid
10
8
6
4
2
0
Rise Delay
Nabaa-Najm
Fall Delay
NEWCAS-05
13
Hybrid Gates: Correlation

Even with correlation
Delay Variations (%)

Hybrid gates exhibit less delay variability than
standard gates
18
16
14
12
10
8
6
4
2
0
Original
Hybrid
Rise Delay
Nabaa-Najm
Fall Delay
NEWCAS-05
14
Hybrid Gates: Limitations


These gains come at the expense of larger absolute delays.
These delays can be recovered by a corresponding increase in area:


This overhead is reduced as the number of inputs increases;


In order for the hybrid NAND to match the nominal delay
performance its area must be 2.1 times the area of a standard NAND
gate
For a three input hybrid NAND, the area overhead required to match
the nominal performance of a standard 3 input NAND gate is 1.5×.
To further minimize area overhead, we use a low Vth for the dummy
transistor.

Area overhead required to match similar performance is down to 78 %
Nabaa-Najm
NEWCAS-05
15
Dynamic Logic

Performed similar experiments on dynamic gates

NOR gates



Very susceptible to variations
The footer in standard dynamic logic helps to reduce variability
Still has large variability



Nabaa-Najm
NEWCAS-05
16
Dynamic Gates: NAND

NAND Dynamic gates exhibit less variation than NOR
Dynamic Gates


But footerless dynamic NAND gates are better than
those with footer


Dynamic NAND has more variations than Static NAND
Can be attributed to the fact that the footer transistor is also
subjected to the normal δVth (and the circuit is already in series)
Use NAND Logic instead of NOR logic whenever
possible.

Footerless NAND logic is fastest and less prone to variability
Nabaa-Najm
NEWCAS-05
17
Transmission gates

Transmission gates display the best delay variability
robustness in both



Correlated simulations
Independent simulations
Can be explained through the intrinsic structure of the
gate



NMOS and PMOS have opposite Vth values (in sign)
In correlated simulations, when subjected to similar ΔVth , the
contribution (faster or slower) that results from say the NMOS
device is counterbalanced by an equal contribution from the
PMOS device and vice versa.
In independent simulations, the variability still remains low
(half that of a NAND gate)
Nabaa-Najm
NEWCAS-05
18
Transmission Gates
Transmission Gate AND
vs Static NAND
(Independent)
Static NAND
Transmission Gate AND
Rise Delay
Nabaa-Najm
Delay Variations (%)
Delay Variations (%)
18
16
14
12
10
8
6
4
2
0
Transmission Gate AND
vs Static NAND
(Correlated)
18
16
14
12
10
8
6
4
2
0
Static NAND
Transmission Gate AND
Rise Delay
Fall Delay
NEWCAS-05
Fall Delay
19
Standalone tests

Second Type of Test


Goals: Find (if any) the critical transistor in a gate



Input threshold voltage variations on only one transistor at a
time:
Can be made wider to minimize the Vth variations
Can be used in the context of a multiple Vth solution
Results:



Bottommost transistor of a stack constitutes the bottleneck
This transistor can be made larger to minimize variability
Can also be used in the context of a multiple Vth solution
Nabaa-Najm
NEWCAS-05
20
Sizing

Third set of tests


Simulated gates with different transistor sizes
The sizing simulations show that:
Larger gates demonstrate less variability
 Optimal widths are twice the width of standard gates

Nabaa-Najm
NAND Δrise 35
NAND rise 34
33
32
31
30
29
28
18
16
14
12
10
8
6
4
2
0
x1
x1.5NEWCAS-05
x.2 x.2.5
Sizing
x.3
Absolute Delay
(ps)
Delay Variability
(%)
Delay Variability vs. Sizing 2-input NAND
21
Sizing
Delay Variability
(%)
10
38.5
38
37.5
37
36.5
36
35.5
35
34.5
NOR Δrise
NOR rise
8
6
4
2
0
x1
x1.5
x.2
x.2.5
Absolute Delay
(ps)
Delay Variability vs. Sizing (2-input NOR)
x.3
Sizing
Nabaa-Najm
NEWCAS-05
22
Conclusion
We studied the delay sensitivity of major design
families with respect to Vth variations
 Series stack are less sensitive than parallel
configurations
 Serialized standard gates: hybrid NAND and NOR
gates
 NAND footerless logic is “better” than standard
dynamic logic.
 Transmission gates are intrinsically robust with respect
to Vth variations
 Optimal sizing of gates seems around 2x that of
standard gates

Nabaa-Najm
NEWCAS-05
23