Impact of Process Variations on Computers Used for Image Processing

molassesitalianΤεχνίτη Νοημοσύνη και Ρομποτική

6 Νοε 2013 (πριν από 3 χρόνια και 7 μήνες)

60 εμφανίσεις

Impact of Process Variations on
Computers Used for Image Processing

Suraj

Sindia
,
Fa

Foster Dai,
Vishwani

D.
Agrawal

Auburn University, AL, USA

Virendra

Singh

Indian Institute of Technology Bombay, Mumbai, India

ISCAS 2012, Seoul, South Korea

22 May 2012

Outline


Why are computers becoming “noisy”?


Background & motivation


How can we model this noise?


Our work in this project


What is their impact on simple image
processing operations?


Results

2

Transistor: Basic Building Block of
Computers

Transistor: SWITCH

(1) Transistors are used as switches.

(2) Computers are built using complex logic networks of these
switches.

D

S

ON

OFF

D

S

S

D

V
GS
>V
TH

V
GS
<V
TH

G

V
GS

3

Transistors Are Shrinking: Moore’s Law

Ref.: S.
Borkar
, “Design Perspectives in 22nm CMOS and Beyond,” DAC’09

Strained Si

High
-
K metal gate

Fin FET

Ref.: D. Patterson, et. al., “Big Data, HPC, and Cancer,” Intel Developer Forum’11

4

Manufacturing Variation in Transistors


As transistor dimensions shrink they no longer
behave “exactly” as intended.


Random
dopant

fluctuation in transistor channel


Line edge roughness from lithography



Threshold Voltage,
V
TH
,

of two transistors on
the same chip is no longer constant.


Line Edge Roughness

Random dopant fluctuation

5

Gate Delay and Threshold Voltage


Variation in transistor threshold voltage leads
to variation in delay offered by logic gates
built using such transistors


V
DD
: Supply Voltage (constant)

t
D
0
: Delay offered by zero threshold voltage gate

t
d
: Delay offered by a logic gate

6

Histogram of
V
TH

and
t
d
at

L=32nm

Variation in delay offered by logic gates on a chip leads
to errors in computation.

5 10 15 20 25 30

t
d
(ns)

0.1 0.2 0.3 0.4 0.5 0.6 0.7

V
TH

(
V
)

150


120


100


80


60


40


20


0

No. of gates

No. of gates

150


120


100


80


60


40


20


0

V
TH

(
σ
/
μ
)=16.82%

Delay (
σ
/
μ
)=21.64%

7

Evolution of
V
TH

and
t
d

Variance

8


20 40 60 80 100 120 140 160 180

Effective channel length,
L

(nm)

0.35

0.30

0.25

0.20

0.15

0.10

0.05

0

Delay variation is increasing at a higher
rate than variation in threshold voltage


delay,
t
d


threshold voltage,
V
TH

σ
/
μ

Outline


Why are computers becoming “noisy”?


Background and motivation


How
can
we model this
noise?


Our work in this project


What is their impact on simple image
processing operations?


Results

9

Statistics of Computation Error Due to
Underlying
V
TH

Variation

µ
VTH

Delay in
logic
circuits

Normal distribution of
V
TH

µ
Delay

Log
-
Normal distribution of

Delay (
t
d
)

0/1
thresholding

due to

flip flop
metastability

(Uniform distribution)

Only if D > µ
Delay

, else it is correct.

Statistics of
errors in
computation is
deduced

0

1

10

p(
V
TH
)

p(
t
d
)

a

b

a

b

a

C
in

b

C
out

C
in

C
in

S

+

b

a

C
in

S

C
out

+

a
7

b
7

S’
7

N(1,
σ
)

S
7

+

a
0

b
0

S’
0

C’
1

N(1,
σ
)

S
0

C
7

C
0

8 bits

N(1,
σ
)

C
1

Our

Unit

element

256
×
256 array Adder

Synthesis of Unreliable Hardware

11

Image Processing Tasks Used for
System Simulation

12

Low pass filtering or image smoothening

1/9

1/9

1/9

1/9

1/9

1/9

1/9

1/9

1/9

Input image

Convolution Mask

2D Convolution

Output image

Scaled output image

High pass filtering or Edge Enhancement

-
1

-
1

0

1

1

0

24

1

-
1

Input image

Convolution Mask

2D Convolution

Analytical Model for Computation
Noise

,
i

,
i

t
d,i

: Delay offered by logic along
i
th

bit line

t
d,th

: Maximum allowed delay across all bit lines.



(Determined by operating clock frequency)

13

0

Outline


Why are computers becoming “noisy”?


Background and motivation


How
can we model this noise
?


Our work in this project


What is their impact on simple image
processing operations?


Results

14

Low Pass Filtering: Comparison Across
Different Technology Nodes

15

Processed on ideal computer

without any process variation

Noisy image

Processed image

Processed on 45nm

Processed on 32nm

Processed on 22nm

High Pass Filtering: Comparison Across
Different Technology Nodes

16

Processed on ideal computer

without any process variation

Noisy image

Processed image

Processed on 45nm

Processed on 32nm

Processed on 22nm

SNR as a Function of Technology Node

17

20
40
60
80
100
120
140
160
180
10
20
30
40
50
60
70
Signal to Noise Ratio (dB)

Technology node (nm)

Low
pass

filter

High pass filter

Threshold of

visual tolerance

Recovering Images using Hardware
Fault Tolerance


We notice that the images in the processed
technologies have noise that is identified as salt
-
and
-
pepper type.


By using simple neural networks in hardware, we can
recover these images online.
[
Sindia

et. al. @ VTS’2012]

18

Processed images on 22nm node

SNR=58dB

SNR=17dB

With hardware fault tolerance

Without hardware fault tolerance

Conclusion


We studied the impact of process variation in
processors on common image processing tasks such
as low pass and high pass filtering.


We synthesized an adder array of size 256 by 256 to study these
image processing tasks.


We saw that there is a decrease in SNR of processed images by
as much as 7.5dB with the advance of every technology node.


We motivated future work to recover images processed on
computers where the underlying hardware has process
variation.


We also proposed analytical model for process
variation in underlying hardware manifesting as
computation noise at a high level of abstraction.

19

20