LDPC Research

Design of Energy-Efficient LDPC
Codes and Decoders
Elsa Dupraz
16/04/2019
Section 1: Introduction 2
OUTLINE
Introduction LDPC codes and decoders Perf. analysis of faulty decoders Effect of faults in the decoders Conclusion
1. Introduction
2. LDPC codes and decoders
3. Perf. analysis of faulty decoders
4. Effect of faults in the decoders
5. Conclusion
D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Elsa Dupraz 16/04/2019
Acknowledgments
I ANR JCJC project EF-FECtive (January 2018 - December 2020)
I Fangping Ye, Mohamed Yaoumi, Zeina Mheich

I François Leduc-Primeau, David Declercq, Valentin Savin, Bane
Vasic, Lav Varshney, Emanuel Popovici, Frederic Guilloud ...

Moore’s Law
I In 1965, Moore predicted that the number of transistors on

processors was going to double every 2 years
1e+12
1e+10
Number of transistors
Core i7
Core 2 Duo
1e+08
Pentium 4
Pentium
1e+06
80386DX
8088
1e+04
4004
Moore
Actual Intel
1e+02
1970 1980 1990 2000 2010 2020
Year
I What about energy consumption ?

5G
In the 5G standardization process
I Huge increase of number of users, terminals, etc.

I Need to improve environmental footprint, battery lifetime

Faulty computation operations
I Hardware energy consumption has become a major issue
I Energy consumption can be reduced by

• Aggressive voltage scaling
• Increased sampling frequency
I Problem : this may introduce faults in the computation operations
Vdd
faulty
Value 1
Value 0
pxor = P(c̃ 6= a ⊕ b)
td t

Energy-efficient LDPC codes
I In this talk, focus on channel coding
E Channel D
faulty faulty
I Noisy vs Faulty
I Family of error-correction codes : LDPC codes

Energy-efficient LDPC codes
I In this talk, focus on channel coding
E Channel D
faulty faulty
I Noisy vs Faulty
I Family of error-correction codes : LDPC codes
I Objectives
• Study the effect of faults in LDPC decoders
• Design fault-tolerant LDPC decoders

Outline
1. Introduction
5. Conclusion

Section 2: LDPC codes and decoders 9
OUTLINE
1. Introduction
5. Conclusion

Block channel codes
E Channel
D
I Channel : P(Y |X )

Block channel codes
E Channel
D
I Channel : P(Y |X )
Encoding Decoding
uk : information sequence (k) xn : codeword (n)

G : generator matrix (n × k) H (n × m) : parity check matrix
xn = Guk H T xn = 0
LDPC codes : H sparse, optimized

for good perf.

Code construction
I H is a sparse parity check matrix, H T xn = 0.
 
1 0 1 1 0 0
H T = 0 1 0 0 1 1
0 1 0 1 1 0
I v1 , v2 , · · · , vn : Variable Nodes (VN), degrees dv

i
c1 , c2 , · · · , cm : Check Nodes (CN), degrees dcj
I Regular codes : constant degrees dv , dc

k dv
R = =1−
n dc
LDPC code construction
I H is a sparse parity check matrix, H T x6 = 03
 
1 0 1 1 0 0
H T = 0 1 0 0 1 1
0 1 0 1 1 0
I v1 , v2 , · · · , vn : Variable Nodes (VN), degrees dv

i
c1 , c2 , · · · , cm : Check Nodes (CN), degrees dcj
I Regular codes : constant degrees dv , dc

k dv
R = =1−
n dc
LDPC decoders
Ex : Gallager decoder, hard-decision decoder
CN c1 VN v1
1
1
0

LDPC decoders
CN c1 VN v1
1 1
1
0 0
0 0

LDPC decoders
CN c1 VN v1
1 1
1
0
0 0 0
0 0

LDPC decoders
CN c1 VN v1
1 1
0 1
0
0 0 0
0 0

LDPC decoders
CN c1 VN v1
1 1 0
0 1 1
0
0
0 0 0 0
0 0

LDPC decoders
CN c1 VN v1
1 1 0
0 1 1 0
0
0
0 0 0 0
0 0

LDPC decoders
CN c1 VN v1
1 1 0
0 1 1 0
0
0
0 0 0 0
0 0
I Hard-decision decoders : binary messages

I Soft-decision decoders : LLR messages, e.g., log P(X =0|y)
P(X =1|y)

LDPC decoders
I VN update function : γ (`) = (`) (`) (`)

Φv (µ0 , µ1 , · · · , µdv −1 )
...
I CN update function : µ(`+1) = Φc (γ (`) , · · · , γ (`) )

1 dc −1
...

LDPC decoders

Φv (µ0 , µ1 , · · · , µdv −1 )

1 dc −1

LDPC decoders

Φv (µ0 , µ1 , · · · , µdv −1 )

1 dc −1
I APP computation : α(`) = Φa (µ(`) , µ(`) , · · · , µ(`) )

0 1 dv
Decide X̂ = 0 if α(`) > 0

LDPC decoders

Φv (µ0 , µ1 , · · · , µdv −1 )

1 dc −1
I APP computation : α(`) = Φa (µ(`) , µ(`) , · · · , µ(`) )

0 1 dv
Decide X̂ = 0 if α(`) > 0
I Hard-decision decoders : binary messages

I Soft-decision decoders : LLR messages, e.g., µ0 = log P(X =0|y)
P(X =1|y)

Faulty LDPC decoders
I Faulty VN update function :
(`) (`) (`)

γ (`) = Φv (µ0 , µ̃1 , · · · , µ̃dv −1 ), P(γ̃ (`) |γ (`) )
Faults
I Faulty CN update function :
(`) (`)
µ(`+1) = Φc (γ̃1 , · · · , γ̃dc −1 ), P(µ̃(`) |µ(`) )
Faults

Section 3: Perf. analysis of faulty decoders 22
OUTLINE
1. Introduction
5. Conclusion

LDPC codes performance
I Two areas of performance :

BER
Waterfall
Error ﬂoor
SNR
I Error Floor : Avoid short cycles in the code (PEG algorithm)
I Waterfall : Optimize the code threshold (density evolution)

All-zero codeword assumption
Symmetry conditions [Richardson01], [Varshney11]

I Channel : P(Y |X = 0) = P(−Y |X = 1)
I VN function : Φv (−µ0 , −µ1 , · · · , −µd −1 ) = −Φv (µ0 , µ1 , · · · , µd −1 )
v v
I CN function : Φc (b1 γ1 , · · · , bd γd ) = (P bi ) Φc (γ1 , · · · , γd )
c−1 c−1 i c−1
I Fault model : P(−µ̃|µ) = P(µ̃| − µ)


I Channel : P(Y |X = 0) = P(−Y |X = 1)
v v
c−1 c−1 i c−1
Examples
I BSC : α = P(Y = 0|X = 1) = P(Y = 1|X = 0)
I Φv (µ0 , µ1 , · · · , µd −1 ) = P µi
v


I Channel : P(Y |X = 0) = P(−Y |X = 1)
v v
c−1 c−1 i c−1


I Channel : P(Y |X = 0) = P(−Y |X = 1)
v v
c−1 c−1 i c−1
All-zero codeword assumption [Richardson01], [Varshney11]

I The decoder performance does not depend on the codeword xn
I All-zero codeword assumption : xn = 0

Density Evolution
I VN messages probability distributions
P(γ (`) |X = 0), P(γ̃ (`) |X = 0)
Faults
I CN messages probability distributions
P(µ(`) |X = 0), P(µ(`) |X = 0)
Faults

Error probability
I Message error probabilities recall LLR : log P(X =0|y)

P(X =1|y)
n,(`)
Pe (α) = P(γ (`) < 0|X = 0)
n,(`)
P̃e (α, ) = P(γ̃ (`) < 0|X = 0)
[Richardson01], [Varshney11], [Huang14], [Ngassa15], [Leduc18], etc.

Performance criterion
I Noiseless threshold : worst channel parameter α for which
n,(`)
lim Pe (α) = 0
n,`→∞
[Richardson01]
I Faulty threshold : worst channel parameter α for which
n,(`)
lim P̃e (α, ) < η
n,`→∞
[Varshney11], [Dupraz15]

Threshold comparison
I Ex : Binary Symmetric Channel
α = P(Y = 0|X = 1) = P(Y = 1|X = 0)
100 0.12
(3,4)-code
(3,4)-code All Zero
0.1 (3,5)-code
10-1 (3,5)-code All Zero
(3,6)-code
0.08 (3,6)-code All Zero
Threshold
(3,12)-code
Pe 10-2 0.06 (3,12)-code All Zero
10-3 0.04
0.02
10-4
0 0.02 0.04 0.06 0.08 0.1
0
1e-3 1e-2 1e-1
eps01

Section 4: Effect of faults in the decoders 30
OUTLINE
1. Introduction
5. Conclusion

Decoder optimization for fault-tolerance
I Min-sum decoder with LLR messages quantized on q bits

I i.i.d. fault model, P(µ̃|µ) , P(γ̃|γ)

I Min-sum decoder with LLR messages quantized on q bits

I i.i.d. fault model, P(µ̃|µ) , P(γ̃|γ)
Method [Dupraz15], [Nguyen-Li16]

• A wide range of quantization functions
• Performance evaluation with faulty Density Evolution
• Optimization of the quantization function for fault-tolerance

I Optimization results
Histogram of thresholds for 5192 decoders 0
10
(different sets of quantization parameters)
1000 −2
10
Bit Error Rate (BER)

With faults
Without
800 faults
Number of decoders
−4
10
Φ(v,SP)
non−robust
noisy
600 Φ(v,SP), noisy

opt
(v,SP)
Φrobust, noisy
−6
10 Φ(v,SP) noiseless
non−robust
400 Non-robust Robust Φ(v,SP), noiseless
robust
(v,SP)
Φopt , noiseless
decoders Decoders −8
200 10
0 0.01 0.02 0.03 0.04 0.05 0.06
Channel error probability
0
0 0.02 0.04 0.06 0.08 0.1
Functional threshold

I Optimization results
Histogram of thresholds for 5192 decoders 0
10
(different sets of quantization parameters)
1000 −2
10
Bit Error Rate (BER)

With faults
Without
800 faults
Number of decoders
−4
10
Φ(v,SP)
non−robust
noisy
600 Φ(v,SP), noisy

opt
(v,SP)
Φrobust, noisy
−6
10 Φ(v,SP) noiseless
non−robust
400 Non-robust Robust Φ(v,SP), noiseless
robust
(v,SP)
Φopt , noiseless
decoders Decoders −8
200 10
0 0.01 0.02 0.03 0.04 0.05 0.06
Channel error probability
0
0 0.02 0.04 0.06 0.08 0.1
Functional threshold
I Conclusion : Careful quantizer design is sufficient to ensure

fault-tolerance

LDPC decoders under timing errors
I Timing errors in the decoder [Brkic15]
. . V
. .
. .
Vdd
Value 1
P(z (t) = w (t) |w (t) , w (t−1) )

= 1 − ε, Value 0
P(z (t) = w (t−1) |w (t) , w (t−1) ) = ε. td t

I Timing errors in the decoder [Brkic15]
. . V
. .
. .
Vdd
Value 1
P(z (t) = w (t) |w (t) , w (t−1) )

= 1 − ε, Value 0
P(z (t) = w (t−1) |w (t) , w (t−1) ) = ε. td t
Main result [Dupraz17]

(`)
If lim`→∞ Pe (α) exists, then ∀,
(`) (`)
lim P̃e (α, ) = lim Pe (α)
`→∞ `→∞

I L = 100 iterations
1e+0
1e-1
1e-2
1e-3
BER 1e-4 (4,16), ε=0

(4,16), ε=0.2
(3,5), ε=0
1e-5 (3,5), ε=0.2
(4,8), ε=0
1e-6 (4,8),ε=0.2
(3,6), ε=0
(3,6),ε=0.2
1e-7
0 0.02 0.04 0.06 0.08
p

I L = 100 iterations
1e+0
1e-1
1e-2
1e-3
BER 1e-4 (4,16), ε=0

(4,16), ε=0.2
(3,5), ε=0
1e-5 (3,5), ε=0.2
(4,8), ε=0
1e-6 (4,8),ε=0.2
(3,6), ε=0
(3,6),ε=0.2
1e-7
0 0.02 0.04 0.06 0.08
p
I Conclusion : Timing errors do not affect the decoder performance

Noisy Gallager B decoder
I Hard-decision decoders [Sundararajan14],[Vasic15]

I Fault model P(µ̃ = 1|µ = 0) = P(µ̃ = 0|µ = 1) =


I LS(155,64)-code
0
10 epsilon = 0
-1 epsilon=1e-2
10
-2
10
-3
10
BER
-4
10
-5
10
-6
10
-7
10 -3 -2 -1
10 10 10
p


I LS(155,64)-code
0
10 epsilon = 0
-1 epsilon=1e-2
10
-2
10
-3
10
BER
-4
10
-5
10
-6
10
-7
10 -3 -2 -1
10 10 10
p
I Conclusion : Faults in the decoder sometimes improve the decoder

performance
Section 5: Conclusion 36
OUTLINE
1. Introduction
5. Conclusion

Conclusion
Conclusions
I Density Evolution permits to analyze the performance of faulty
LDPC decoders
I The robustness to faults depends on the decoder and on the fault
model
Other existing works

I LDPC encoders [Hachem13],[Yang14],[Dupraz16]
I LDPC decoders for faulty computation [Grandhi16],[Yang16]
I LDPC decoders in faulty memories [Chilappagari07],[Vasic07]
I Other families of error-correction codes [Balatsoukas18]
I Machine Learning Algorithms under faulty hardware [Yang16],
[Leduc18] ,[Dupraz19]

Perspectives
Ongoing works and Perspectives

I Energy optimization of LDPC codes and decoders [Yaoumi19]
I Realistic energy-vs-faults models
I Practical implementations
I Energy-efficient Machine Learning algorithms


LDPC Research

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

LDPC Research

Uploaded by

Copyright:

Available Formats

Design of Energy-Efficient LDPC

Codes and Decoders

2. LDPC codes and decoders

3. Perf. analysis of faulty decoders

4. Effect of faults in the decoders

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I ANR JCJC project EF-FECtive (January 2018 - December 2020)

I Fangping Ye, Mohamed Yaoumi, Zeina Mheich

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I In 1965, Moore predicted that the number of transistors on

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

In the 5G standardization process

I Huge increase of number of users, terminals, etc.

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I Hardware energy consumption has become a major issue

I Energy consumption can be reduced by

I Problem : this may introduce faults in the computation operations

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I In this talk, focus on channel coding

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I In this talk, focus on channel coding

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

2. LDPC codes and decoders

3. Perf. analysis of faulty decoders

4. Effect of faults in the decoders

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

2. LDPC codes and decoders

3. Perf. analysis of faulty decoders

4. Effect of faults in the decoders

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

uk : information sequence (k) xn : codeword (n)

LDPC codes : H sparse, optimized

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I H is a sparse parity check matrix, H T xn = 0.

I v1 , v2 , · · · , vn : Variable Nodes (VN), degrees dv

I Regular codes : constant degrees dv , dc

I H is a sparse parity check matrix, H T x6 = 03

I v1 , v2 , · · · , vn : Variable Nodes (VN), degrees dv

I Regular codes : constant degrees dv , dc

Ex : Gallager decoder, hard-decision decoder

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Ex : Gallager decoder, hard-decision decoder

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Ex : Gallager decoder, hard-decision decoder

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Ex : Gallager decoder, hard-decision decoder

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Ex : Gallager decoder, hard-decision decoder

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Ex : Gallager decoder, hard-decision decoder

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

Ex : Gallager decoder, hard-decision decoder

I Hard-decision decoders : binary messages

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I VN update function : γ (`) = (`) (`) (`)

I CN update function : µ(`+1) = Φc (γ (`) , · · · , γ (`) )

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I VN update function : γ (`) = (`) (`) (`)

I CN update function : µ(`+1) = Φc (γ (`) , · · · , γ (`) )

D ESIGN OF E NERGY-E FFICIENT LDPC C ODES AND D ECODERS

I VN update function : γ (`) = (`) (`) (`)