Event announcement Topic: Thermal-Aware Design of 2D/3D Many-Core - - PowerPoint PPT Presentation

1/31

Event announcement

Topic: Thermal-Aware Design of 2D/3D Many-Core Servers with Inter- Tier Liquid Cooling Speaker:

• Prof. David Atienza, École polytechnique fédérale de Lausanne

(EPFL), Switzerland Time/Location: Monday, July 7th, 16:00, Room H120 Technologiefabrik

2/31

Bachelor/Master Thesis:

Build your own manycore System-On-Chip

• FPGAs
• Network on Chip
• Verilog
• SystemC
• Qt

OptimSoc http://www.optimsoc.org/ DFG SPP 1500 Dependable Embedded Systems http://spp1500.itec.kit.edu/ ZTEX FPGA Module 1.15

src: http://www.ztex.de/usb-fpga-1/usb-fpga-1.15.e.html

interested? volker.wenzel@kit.edu

3/31

Dennard Scaling and Power

• see blackboard

4/31

The end of Moore's law?

src: http://en.wikipedia.org/wiki/Moore%27s_law

„Prediction is very difficult, especially about the future.“

Niels Bohr (1885 - 1962)

5/31

Aging Mechanisms in Integrated Circuits

probabilistic empirical physical Aging Phenomena

src: de.wikipedia.org/wiki/Tesla-Transformator src:de.wikipedia.org/wiki/ Entropie_Thermodynamik

2nd law of thermodynamics

6/31

Aging Mechanisms in Integrated Circuits

• Electromigration (EM)
• Negative Bias Temperature Instability (NBTI)
• Hot Carrier Degradation (HCID)
• Time Dependent Dielectric Breakdown (TDDB)
• Soft Errors
• Process Variation

7/31

Electromigration What is it?

• very old reliability concern
• physical migration of atoms

through current

• „electron wind“ = momentum

transfer of electrons to atoms/ions

• affects metal-interconnects
• power and ground wires (DC

current flow)

• voids (internal failure)
• hillocks (hillock failure)

src: http://en.wikipedia.org/wiki/Electromigration

8/31

Electromigration How to model EM?

• Black's equation (empirical

model)

– A constant – j current density – n parameter – Q activation energy in eV – k Boltzmann constant – T temperature in K

• Homework Challenge: Derive

standard deviation of Black's equation.

9/31

Electromigration How to deal with EM?

• copper better than

aluminium for inter- connects

• design rules
• TCAD Simulation of

wires

10/31

NBTI: Preparation Threshold Voltage

src:http://en.wikipedia.org/wiki/Threshold_voltage

11/31

NBTI: Preparation Hydrogen?

• Hydrogen-terminated silicon surface
• chemical passivation for silicon substrate
• remove natural SiO

2 surface

• terminate dangling bonds
• realized by treatment with HF

12/31

NBTI What is it?

• = Negative bias temperature instability
• mainly p-channel MOSFETs
• increases Vth over long periods of time (years)
• decreasing transconductance, linear drain current,

saturation current, channel mobility,...

• stress means

– negative gate voltages – at elevated temperatures

• signal probability (SP) important

parametric degradation

13/31

NBTI

Breaking Bonds

• Model 1: Reaction-Diffusion

– breaking of SiH at SiO2/Si

substrate interface

– H diffuses away

• Model 2: Charge Trapping
• creation of

– interface traps

(permanent traps)

– oxide traps

(can recover)

src:http://www.iue.tuwien.ac.at/phd/entner/ Quiz: How could you check if the hydrogen at the Si surface is really responsible for NBTI?

14/31

NBTI

Breaking Bonds

src: [Schro04]

15/31

NBTI shift of threshold voltage

• power law time

dependence

– initially quick degradation

& recovery

• Arrhenius law

temperature dependence

• highly device-dependent

16/31

effect on latch:

• reduced SNM
• increased switching

time

NBTI

impact on different electronic components

effect on SRAM:

• reduced SNM

effect on logic gate:

• increased delay

effect on ring

• scillator:
• frequency

deterioration

17/31

NBTI

„Have you tried turning it off and on again?“

• NBTI recovers (partially)
• makes measurement

difficult

• mitigation techniques
• NBTI depends on

– voltage stress – signal probability – temperature Vth shift [V] Time

Stress Recovery

Vg [V]

• 1

18/31

Static Noise Margin: a 6-T SRAM cell...

src:http://www.iue.tuwien.ac.at/phd/entner/node34.html

19/31

Static Noise Margin What are Butterfly Curves?

• Voltage Transfer

Curve (VTC)

• SNM is length of

biggest square

• SNM quantifies

resiliency to noise

src:http://www.iue.tuwien.ac.at/phd/entner/node34.html

20/31

Static Noise Margin What are Butterfly Curves?

NBTI decreases SNM

src:http://www.iue.tuwien.ac.at/phd/entner/node34.html

21/31

Hot Carrier Degradation What is HCID?

• acceleration of carriers

(electrons/holes) under lateral electric fields

• hot electrons cause

damage

• dependent on

– supply voltage – channel geometry src: http://www.iue.tuwien.ac.at/phd/entner

22/31

Hot Carrier Degradation Maxwell Boltzmann Distribution

• electrons at right tail

causing damage in device

– „hot electrons“ – „lucky electrons“ – electrons w/ high

kinetic energy

23/31

Time Dependent Dielectric Breakdown

What is TDDB?

• old reliability concern
• degrade oxide layer's insulating property

by breaking of Si-O bonds

• leads to excessive gate drain current
• depends on

– dielectric thickness – manufacturing quality of gate oxide

src:images.nationalgeographic.com

24/31

Time Dependent Dielectric Breakdown What is TDDB?

src: https://nanohub.org/resources/17208/download/2013.03.01-ECE695A-L21.pdf BINGO!

25/31

Time Dependent Dielectric Breakdown What is TDDB?

• more recent: breakdown ist not instantaneous

hard breakdown vs. soft breakdown aka Progressive Breakdown

• degrade drain current by 10%
• shift threshold voltage by 75mV

(like NBTI, but without recovery)

• ref. „Circuit-level delay modeling considering both TDDB and NBTI“ (Yu Cao)

26/31

What is TDDB? Mitigation

• TDDB is Voltage Accelerated Process

(Reducing gate voltage helps)

• Power gated circuits do not degrade from TDDB.
• Soft breakdown can partially be tolerated in thin
• xides
• Thickness of gate oxide, voltage chosen at

transistor design time for desired chip life time

• Reducing Temperature helps

27/31

Comparison of TDDB and NBTI

NBTI:

• diffusion of H, H2 in gate/poly
• creates traps in Si/SiO2

interface

• recovery
• shifts Vth
• magnitude depends on duty

cycle aka signal probability

TDDB:

• increases leakage
• traditionally: hard

breakdown due to conducting channel in dielectric

• no recovery
• recently:soft/hard TDDB
• shifts Vth

28/31

Comparison of TDDB, NBTI, HCID

src: https://nanohub.org/resources/17208/download/2013.03.01-ECE695A-L21.pdf

29/31

Conclusions

• Electromigration
• NBTI
• Hot Carrier Injection
• TDDB

Quiz: Which part affects what?

src: http://de.wikipedia.org/wiki/Metall-Oxid-Halbleiter-Feldeffekttransistor

30/31

Sources

• „Intrinsic Transistor Reliability Improvements from 22nm Tri-

Gate Technology“ Ramey, S. et al

• „A Lifetime Projection Method Using Series Model and

Acceleration Factors for TDDB failures of Thin Gate Oxides“ Shiono, N. et al

• „Method of Determining Reliability Screens for Time Dependent

Dielectric Breakdown“ Crook, D.L. et al

• „Circuit Reliability: From Physics to Architectures.“ Fang, j. et al

(good overview, ICCAD 2012)

• [Schro04] „Negative Bias temperature instability: What do we

understand?“ Dieter K. Schroder

31/31

Sources

• http://www.iue.tuwien.ac.at/phd/entner/
• „Reliability Physics and Engineering: Time-To-

Failure Modeling.“ J. W. McPherson, Springer Science & Business Media, 2010