the frontiers of robust circuit design - isqed...the frontiers of robust circuit design in...

42
The Frontiers of Robust Circuit Design in Sub28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our success in the past will no longer work in the future.” Cliff Hou, VP R&D, TSMC, at DAC 2014 2

Upload: others

Post on 15-Aug-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

The Frontiers of Robust Circuit Designin Sub‐28nm Process Technologiesg

Jim Dodrill

ARM

“The conventional wisdom that led to our success in the past will no longer work in the future.”

Cliff Hou, VP R&D, TSMC, at DAC 2014

2

Page 2: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

3

Agenda

• The Cost of Design Margins

• FinFET Changes the LandscapeFinFET Changes the Landscape

• Variation– Process, Voltage, and TemperatureProcess, Voltage, and Temperature

– Location and Layout Dependent Effects

• Random Failures– Radiation, Synchronization, and Noise

• Aging– BTI, HCI, EM, and TDDB

• Putting it All Together

4

Page 3: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Focus

• Focus is on SoC implementation aspects

Assume foundry and IP partners have done their– Assume foundry and IP partners have done their best to offer reliable products

• For each failure mechanism:

I d h h i l b k d– Introduce the physical background

– Compare planar vs. FinFET

– Offer practical design advice

5

Device Lifetime

• Testing• defects

• Random Failures• soft errors

• Aging• BTI HCI• defects

• variation• soft errors• synchronization• noise

• BTI, HCI• EM, TDDB

e rate

failure

burn‐in useful life wear‐out

6

log time

Page 4: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

The Heart of the Matter…

“At the heart of reliability engineering is the fact that there is a di ib i f lif i f h f il h i Wi h ldistribution of lifetimes for each failure mechanism.  With low failure rate requirements we are interested in the early time‐range of the failure time distributions. There has  been an increase in process variability with scaling (e.g., distribution of dopant atoms, CMP variations, and line‐edge roughness). At the same time the size of a critical defect decreases with scalingsame time the size of a critical defect decreases with scaling. These trends will translate into an increased time spread of the failure distributions and, thus, a decreasing time to first failure.”

ITRS 2013, Process Integration, Devices, and Structures 

7

Tactics

• Margining

Th l M t• Thermal Management

• Dynamic Voltage and Frequency Scaling (DVFS)

– In‐situ monitors

– Adaptive Supply Voltage (ADV) or Back‐Bias (ABB)

• Redundant Components

– Error Detection and/or Correction/

– Time Multiplexing (Resource or Task Allocation)

– Cannibalization (Resource Sharing)

8

Cannibalization (Resource Sharing)

Page 5: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

The Cost of Design Margins

Moore’s Law

Dark Silicon

Margin Cost

“Traditional VLSI design bypasses the analysis and optimization of such non‐uniform, dynamic network by approximating the problem into optimization of uniform static network with certain 

d b d ”guard band.”

Muhammad Alam,“Reliability and process variationReliability‐ and process‐variation aware design of integrated circuits”2008 Elsevier Ltd.

10

Page 6: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Moore’s Law is Still Alive

• The cost per transistor continues to reduce

Or does it?– Or does it?

11 Source: Bill Holt (Intel), “Intel Investor Meeting,” (2014)

Another View…

• Lower yield will drive costs up in at 20nm and smaller while competition will drive costssmaller while competition will drive costs down at 28nm and larger.

12

Page 7: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

The Dark Silicon Apocalypse“Where once we would spend exponentially increasing amounts of silicon area to buy performance, now, we will spend exponentially increasing amounts of silicon area to buy energy efficiency.”

Michael B. Taylor, “Is Dark Silicon Useful? Harnessing the Four Horsemen of the Coming Dark Silicon Apocalypse ” ACM (2012)Dark Silicon Apocalypse,” ACM (2012)

Will increasing amounts of silicon area be used to buy reliability?

13

Will increasing amounts of silicon area be used to buy reliability?Of course!

Timing Margins

• More accurate modeling of timing variation results in more efficient implementationsresults in more efficient implementations

120%

OCV Margin Methods

106%

94% 96%

86%

80%

100%

60%

80%

OCV

AOCV

LVF

20%

40%

14

1% 1% 1% 1%0%

frequency dynamic power static power cell area

Page 8: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

FinFETs Change the Landscape

Planar vs. FinFET

Multi‐patterned Interconnect

Introducing the FinFET

Source: GLOBALFOUNDRIES

16

Page 9: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Planar vs. FinFET

17 Source: Mark Bohr , et al. (Intel) (2011)

Fins are Very Thin

• This is a TEM image of the Intel 22nm Fin

At 14 Fi ill b b t 20 t id• At 14nm, Fins will be about 20 atoms wide

18 Chipworks (2012)

Page 10: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

The Future…

19

Source: Martin van den Brink, “Many ways to shrink: The right moves to 10 nanometer and beyond,” AMSL (2014)

Delay vs. Voltage

20 Source: Mark Bohr , et al. (Intel) (2011)

Page 11: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Power vs. Speed

21 Source: Shien‐Yang Wu, et al. (TSMC), IEDM (2013)

Leakage vs. Achieved Frequency

• Leakage reduction from FinFET is significant

age

Leak

a

22Source: Leah Schuth, (ARM) (2014)

Achieved Clock Frequency

Page 12: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Interconnect

23Source: S. Natarajan, et al. (Intel), IEDM (2014)

Wire Resistance

• The RC time constant of wires is increasing substantially as line widths reducesubstantially as line widths reduce

24Source: Serkan Kincal, et al., “RC Performance Evaluation of Interconnect Architecture Options Beyond the 10‐nm Logic Node,” IEEE (2014)

Page 13: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Cell vs. Wire Delay

• Wire delay is becoming as big as cell delay

We’ve heard this for years but its real now– We ve heard this for years, but its real now

25Source: Greg Yeric (ARM), “Design, Technology and Yield in the Post‐Moore Era,” ITC (2014)

Double Patterning

26

Source: Andrew J. Hazelton, et al. “Double‐patterning requirements for optical lithography and prospects for optical extension without double patterning,”  J. Micro/Nanolith. MEMS MOEMS (2009)

Page 14: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

RC Variation

• Note the wire thickness and spacing differences in the two metal patterns A & Bdifferences in the two metal patterns, A & B

27Source: ITRS 2013 EDITION: INTERCONNECT

On‐Chip Variation

Process, Voltage & Temperature Variation

Layout Dependent Effects

OCV Modelling

Page 15: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Sources of Variation

• Process (transistor and wire)

V lt• Voltage

• Layout Dependent Effects (LDE)

• Temperaturetemperature variation

LDE

voltage variation

process variation

29

Variation: Planar vs. FinFET

• Single fin FinFETs are not used due to high variationvariation

Source Planar FinFET

Random Dopant (less)Line Edge Roughness Gate Edge Roughness Gate Granularity Fin Edge Roughness Fin Height Fi Sh

30

Fin Shape Source: Greg Yeric (ARM), “Design, Technology and Yield in the Post‐Moore Era,” ITC (2014)

Page 16: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Variation vs. Voltage

• Delay and variation increase as voltage decreasesdecreases

31 Source: Isadore Katz, CLK DA (2014)

Voltage Variation

• Chip dynamic voltage drop based on two different operating modesdifferent operating modes

32 A. Shanmugavel, Ansys Inc. (2013) 

Page 17: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Temperature Variation

• Thermal conduction from the channel

planar transistors: tends toward the substrate– planar transistors: tends toward the substrate

– finFET transistors: tends toward the metal

~12C

~

33 A. Shanmugavel, Ansys Inc. (2013) 

Layout Dependent Effects

• Non‐uniformities in the surrounding context of a transistor adds to delay variationof a transistor adds to delay variation

34

Page 18: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

OCV Modeling

• Traditional OCV

Percentage derate applied to clock paths– Percentage derate applied to clock paths

– Plus a fixed margin added to the clock uncertainty

eren

ce

intercept=margin

of d

elay

diff

e3s

igm

a o

35

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

path depth

AOCV Modeling

• Advanced OCV (AOCV)

Tables of derate values for cells and nets indexed– Tables of derate values for cells and nets indexed by path depth

– These tables are called Stage‐Based OCV (SB‐OCV)– These tables are called Stage‐Based OCV (SB‐OCV)

– May include location‐based (LOCV) derate tables

1 5

1

1.5

1±nσ/μ early

late

0.5

1

derate= 

36

0 20 40 60 80 100

d

path depth

Page 19: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Stage Based‐OCV Limitation

• SB‐OCV is limited to:

One timing arc per cell– One timing arc per cell

– One load/slew point

Delay Variation vs. Load/Slew

n index

e (σ/µ)

1

3

5

put tran

sition

Variance

37

7

1 2 3 4 5 6 7 8

inp

output load index

OCV Modeling

• Liberty Variance Format (LVF)

Table of sigma values for cell delay transition– Table of sigma values for cell delay, transition time, and constraints in the Liberty model

– Derates are calculated by the timing engine based– Derates are calculated by the timing engine based on N*sigma and path depth

LOAD

σ σ σ σ σ

SLEW

σ σ σ σ σ

σ σ σ σ σ

σ σ σ σ σ

38

σ σ σ σ σ

Page 20: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Statistical Constraint Margins

• Timing Constraints also have variation

Usually only hold and removal constraints are– Usually only hold and removal constraints are margined 

lay

rop

agat

ion

de

Setup and hold are defined by a specified increase in

additional hold margin

Prp

CK->Q

39

Data arrival w.r.t. clocktsu_tp_r th_tp_r

Constraint Margins vs. Voltage

• The variation of also constraints increases as voltage decreasesvoltage decreases

40 Source: Isadore Katz, CLK DA (2014)

Page 21: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

PVT Corners

Hold must beHold must be met across the entire 

Full‐yield setup corner

process range

Typical On‐DieMean

FastestSlowest On‐DieMean

41

MeanFast

MeanSlow

Determining N*sigma

• LVF allows users to choose N*sigma

Native die yield should dominate not timing yield– Native die yield should dominate, not timing yield

Step Examplep p

Yield Failure Rate

Estimate manufacturing yield 95% 5/100

Chose a timing yield target that is better

97% 3/100

Estimate the number of near‐critical hold paths

50K 2/100K

multiply 99 99994% 6/10M

42

multiply 99.99994% 6/10M

Derive sigma about 5σ

Page 22: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

“Statisticians, like artists, have the bad habit ofStatisticians, like artists, have the bad habit of falling in love with their models.”

‐‐ George Box

43

Practical Advice

• Use the latest, most accurate variation models availableavailable

– Consult your foundry and IP provider for guidance

B t t f b i i ti• Be sure to account for non‐obvious variation

– Layout Dependent Effects

– Wire Variation

• Consider the cost of design margins

– Yield loss vs. power, area, and design schedule

– Prioritize good hold margins

44

Page 23: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Random Failures

Soft Error Rate (SER)

Synchronization

Random Telegraph Noise (RTN)

The Neutron Strike

• Cosmic radiation includes alpha particles and high energy neutrons that can createhigh energy neutrons that can create ionization which can upset logic

46

Source: Christopher Frost, “How Alien InvadersCan Change Governments,” ISIS (2014)

Page 24: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

SER of Planar vs. FinFET

• The trend is for improvement in SER

FinFET has 3X 10X lower SER than planar– FinFET has 3X‐10X lower SER than planar

47 Anthony S. Oates (TSMC), “Will Reliability Limit Moore’s Law?,” IEDM (2014) 

SER of Systems

• While the SER of individual state elements is improving overall chip SER is notimproving, overall chip SER is not

– Flip‐flops dominate in systems with ECC

48Source: Vikas Chandra, “Cross layer resilience in real world,” DATE (2014)

Page 25: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

SER vs. Voltage in FinFET

• SER increases as voltage decreases

Various FinFET technologies can improve SER– Various FinFET technologies can improve SER

49Source: Huichu Liu, et al., “Soft‐Error Performance Evaluation on Emerging Low Power Devices,“ IEEE (2014)

Practical Advice

• Acquire FIT rate estimates from foundries and IP providersIP providers

– Consider the lowest operating voltage domain

D t i hi h t t l t t ib t• Determine which state elements contribute the most to the probability of program failure

– The “Architectural Vulnerability Factor” (AVF)

– Harden those state elements

• Use appropriate error detection and correction

50

Page 26: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Clock domain crossing

• Synchronizers are used when data crosses between two asynchronous clock domainsbetween two asynchronous clock domains.

– That means data can change during the window between the setup and hold constraintsbetween the setup and hold constraints.

CK1Y

CK2

YD DQ Q

CK1 CK2

Y

CK2

tsu th

51

CK1 and CK2 are asynchronous

Metastability

• When setup and hold constraints are violated, the signals inside the receiving flip‐flop can failthe signals inside the receiving flip‐flop can fail to resolve within a clock period

CK

D

nm

CKCK

m

CK2

tsu th

Dnm

CK

m

m/nm

m fails to resolve to ‘1’ or ‘0’

52

m fails to resolve to 1 or 0

Page 27: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

MTBF for Synchronizers

TS is the resolution time, which is approximately the clock periodτ is the resolution time constantτ is the resolution time constant, 

a function of the latch design and PVT cornerTW is the time window, also a function of the latch design and PVTfd is the data frequencyfd is the data frequencyfc is the clock frequencyn is the number of synchronizers in the entire system

53

MTBF vs. Voltage

• MTBF reduces as voltage decreases– MTBF also decreases as the period decreases– MTBF also decreases as the period decreases

– VT choice is critical

MTBF vs. Voltage MTBF vs. VT and LG

70

80

90

1 E+50

1.E+60

1.E+70

1.E+80

600

700

800

1E+50

1E+60

1E+70

1E+80T G

30

40

50

60

1.E+20

1.E+30

1.E+40

1.E+50

τ(ps)

MTB

F (years)

300

400

500

1E+20

1E+30

1E+40

1E+50

τ(ps)

MTTF (years)

The universe

is 1 38*1010

0

10

20

30

1.E‐20

1.E‐10

1.E+00

1.E+10

0

100

200

1E 20

1E‐10

1

1E+101.38*1010

years old

54

01E‐20

lowest VT highest VT

Page 28: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Practical Advice

• Use specially constructed synchronizer cells

– One cell N‐stages deep– One cell, N‐stages deep

– Choose the lowest VT and shortest LG available

• Obtain the necessary MTBF parameters for your• Obtain the necessary MTBF parameters for your synchronizers

– Use the model for an average transistor on a slow die– Use the model for an average transistor on a slow die

– Calculate the MTBF at each voltage and frequency combination

– Account for the total number of synchronizers in the computing system

55

Random Telegraph Noise

• Random Telegraph Noise (RTN) is caused by the capture and emission of carriers at trapsthe  capture and emission of carriers at traps (defects) in the oxide boundary

56Source: Xiaoming Chen, et al., “Statistical Analysis of Random Telegraph Noise in Digital Circuits,” IEEE (2014)

Page 29: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Random Telegraph Noise

• “The static variability of the source‐induced RDF is found to overwhelm the dynamic on‐current f yfluctuation due to RTN.”

RTN = Random Telegraph NoiseRDF R d D t Fl t ti

57Source: Akito Suzuki, et al, “Source‐induced RDF Overwhelms RTN in Nanowire Transistor: Statistical Analysis with Full Device EMC/MD Simulation Accelerated by GPU Computing,” IEEE (2014)

RDF = Random Dopant Fluctuation

Aging

BTI: Bias Temperature Instability 

HCI: Hot Carrier Injection

EM: Electro‐migration

TDDB: Temperature Dependent Dielectric Breakdown

Page 30: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

BTI: Stress and Recovery

VDD VDD VSSPMOS:Negative Bias

G:0(on)

D:1

G:1(off)

D:1 or 0

G:1(off)

D:1 or 0

TemperatureInstability(NBTI)

D:1 D:1 or 0 D:1 or 0

Stress NaturalRecovery

ProactiveR

D:0 or 1D:0 D:0 or 1

Recovery Recovery

NMOS:Positive Bias

G:0(off)

G:1(on)

G:0(off)

VDD

Positive BiasTemperatureInstability(PBTI)

59

VSSVSS

(PBTI)

Based on: Lin Li, “Improving the Reliability of Microprocessors Under BTI and TDDB Degradations,” University of Pittsburgh (2014)

BTI: Planar vs. FinFET

• The VT shift due to PBTI is lower, and due to NBTI is higher for FinFETNBTI is higher for FinFET

60 Kyong Taek Lee, (Samsung) IEEE (2013)

Page 31: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

BTI: Ring Oscillator Degradation

• VT shifts due to BTI lead to larger propagation delaydelay

61

Source: Halil Kükner, et al., “Scaling of BTI reliability in presence of Time‐zero Variability,” IEEE (2014)

BTI vs. Voltage

• At higher voltage, delay degradation is more 

62

Source: Halil Kükner, et al., “Scaling of BTI reliability in presence of Time‐zero Variability,” IEEE (2014)

Page 32: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

BTI and Variation

• Process variation remains a normal distribution after agingdistribution after aging

63 D. Angot, et. al. , IEDM (2013)

BTI: Stress vs. Relaxation

• Partial healing during relaxation leads to delay degradation that is state dependentdegradation that is state dependent

64 Kyong Taek Lee, (Samsung) IEEE (2013)

Page 33: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

BTI and Duty Cycle 

DutyFactor

Square WaveArbitrary Switching

100%75%50%

Factor

50%25%5%

65

Source: Haldun Kufluoglu, “MOSFET Degradation due to Negative Bias Temperature Instability (NBTI) and Hot Carrier Injection (HCI) and its Implications for Reliability‐Aware VLSI Design ,” Purdue University (2007)

Path Rank Analysis

• Rank paths in fresh and aged design, sorted by slack

• Non‐critical paths can become critical vice versaNon critical paths can become critical, vice versa

Path rank% Timing 

DegradationFreshAged 

(Dhrystone)

Path rank% Timing 

DegradationAged 

(Dhr stone)Freshg

(Dhrystone)

1 14084 7.642 9781 7.94

3 9329 8.02

g(Dhrystone)

1 179394 15.612 145042 15.413 134419 15.183 9329 8.02

4 12345 7.875 6220 8.316 36672 7.16

3 134419 15.18

4 1413427 17.575 272323 15.676 224034 15.46

7 7771 8.198 11580 7.969 28975 7.40

10 20054 7 66

7 331934 15.768 275422 15.569 481425 16.06

10 208561 15 24

66

10 20054 7.66 10 208561 15.24

Source: Vikas Chandra, et al. (ARM), “Workload dependent NBTI and PBTI analysis for a sub‐45nm commercial microprocessor,” IEEE (2013)

Page 34: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Hold Time with Gated Clocks

• Since the rising edge of a gated clock spends most of its time in recovery it does not agemost of its time in recovery, it does not age

• Rising edge of un‐gated clocks does age

• Leading to potential hold failures over time

Rising edge does not age

EN0

1 0 1 0

CK

67

Both edges age

Practical Advice (1/2)

• Set critical range to at least 10% of the clock periodperiod 

– Prevents area and leakage recovery from creating setup paths that will age to become criticalsetup paths that will age to become critical

critical range

percentile

area and leakage recovery

path p

68

0 +10%path slack

Page 35: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Practical Advice (2/2)

• Most gated clocks should connect directly to sequential elementssequential elements

– Identify cases of clock tree stages beyond the gated clock and add extra hold margingated clock and add extra hold margin

• Advocate with EDA vendors for static timing with aging based on time and switchingwith aging based on time and switching activity

Th i f d t i i i t it hi– The issue of determining an appropriate switching activity remains

69

Hot Carrier Injection

Channel Hot‐Electron Injection Drain Avalanche Hot‐Carrier Injection

70

Channel Hot Electron Injection Drain Avalanche Hot Carrier Injection

Source: Robert Entner , “Modeling and Simulation of  Negative Bias Temperature Instability,” Technische Universitat Wien (2007)

Page 36: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

HCI: Planar vs. FinFET

• HCI in FinFET is better than planar transistors

71Source: S. E. Liu, et. al., (TSMC) “Self Heating Effect in FinFETs and its Impact on  Device Reliability Characterization,” IEEE (2014)

Practical Advice

• Limit the maximum transition time to reduce degradation due to HCIdegradation due to HCI

– This will insure HCI has less effect than BTI

Ad t ith EDA d f t ti ti i• Advocate with EDA vendors for static timing with aging due to HCI

– Table based on input transition, output load, and switching activity 

72

Page 37: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Electromigration

(Black’s Equation)

73 Source: Y. L. Hsu, et al., “Failure Mechanism of Electromigration in Via Sidewall for Copper Dual Damascene Interconnection,” ECS (2006)

EM: Resistance vs. Time

• Resistance changes as dislocations form– Failure criteria is specified as a given resistance shift in a– Failure criteria is specified as a given resistance shift in a percentage of samples

74Source:  Zhuojie Wu, “Study of Initial Void Formation and Electron Wind Force for Scaling Effects on Electromigration in Cu Interconnects,” University of Texas  (2013) 

Page 38: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

EM Trend

• With stronger FinFET transistors and thinner interconnect EM is becoming criticalinterconnect, EM is becoming critical

75 ITRS (2013)

EM and FinFETs

• Self‐Heating in FinFETs may lead to worse EM in surrounding wiresin surrounding wires– Self‐heat manifests as a sensitivity to the fin or gate count in switching aging degradationg g g g

76http://semimd.com/blog/tag/rram/

Page 39: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Practical Advice 

• Take advantage of the Blech effect whenever possible, especially in VIA stackspossible, especially in VIA stacks

• Build and verify power grids early– Limit placement density in areas with high powerLimit placement density in areas with high power

• Limit wire length and maximum transition times to reduce RMS EM violations on signalstimes to reduce RMS EM violations on signals

• Follow IP provider guidelines for cell level EM compliancecompliance

• Carefully balance operating temperatures and MTBF due to EM

77

MTBF due to EM

Time Dependent Dielectric Breakdown

78 Source: ITRS 2013 EDITION: INTERCONNECT

Page 40: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

TDDB:

• Electric fields are highest at wire tips

Line edge roughness contributes on wire sides– Line edge roughness contributes on wire sides 

79Source: Ong,  IEEE IPFA (2012)

TDDB vs. Spacing

• TDDB is becoming a concern

m m 4nm

28nm

22nm

20/14

80Anthony S. Oates (TSMC), “Will Reliability Limit Moore’s Law?,” IEDM (2014) 

Page 41: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

Summary

The Balancing Act

The Balancing Act: Temperature

* * * *

rature

Temper

T

equen

cy

Power

ariation

SER

c. M

TBF

BTI/HCI

EM

TDDB

82

Fre Va

Sync B

*Depends on voltage

Page 42: The Frontiers of Robust Circuit Design - ISQED...The Frontiers of Robust Circuit Design in Sub‐28nm Process Technologies Jim Dodrill ARM “The conventional wisdom that led to our

The Balancing Act: Voltage

age

Volta

equen

cy

Power

ariation

SER

c. M

TBF

BTI/HCI

EM

TDDB

83

Fre Va

Sync B

84