Unintrusive Aging Analysis based on Offline Learning Frank Sill - - PowerPoint PPT Presentation

Sill Torres - Aging Analysis

Unintrusive Aging Analysis based on Offline Learning

Frank Sill Torres*+, Pedro Fausto Rodrigues Leite Jr.*, Rolf Drechsler+ *Universidade Federal de Minas Gerias, Belo Horizonte, Brazil

+University of Bremen, Bremen, Germany

2 Sill Torres - Aging Analysis

Motivation

• Aging of integrated systems of rising

importance

• But:

– (Still) less critical for customer applications – Interest in low weight solutions (S.M.A.R.T. for HDDs, …)

• This work:

– Low-weight aging monitoring / remaining lifetime prediction – Based on (offline) learning

V

3 Sill Torres - Aging Analysis

Aging Monitoring

• In-situ slack sensors

– Detection / preview of failing timing – Added invasively to (selected) critical paths

• Online self-testing

– Built-In Self-Test (BIST) during test mode – Additional circuitry (Scan chains, …)

• Aging sensors

– Report experienced aging – Ignores system’s activity

C

4 Sill Torres - Aging Analysis

Software Layer

Prediction Reporting APDB MDB Compression Counter- measures

Unintrusive Aging Analysis

Architecture

• APDB, MDB: Databases

Simulations

Profiling

Stress Test Field Data

VDD, Freq., Sleep

Hardware

Stress sensors Temp, V, Activity

5 Sill Torres - Aging Analysis

• Sensors

– Temperature, voltage, activity, … – Low area offset, unintrusive

• Profiling

– Simulations

• Aging characterization at design time
• Various scenarios (Temp, VDD, activity, …)
• Parameter can vary

– Also possible: Data from stress test / field

Unintrusive Aging Analysis

Profiling

6 Sill Torres - Aging Analysis

Unintrusive Aging Analysis

Compression and Profile Storage

10 20 30 40 Set 4 Set 3 Set 2 Set 1 Set 0

… Sensor ST,4 … MTTF in Set 0 [%] … in Set 4 [%] 20 % 32 % 2e2 h

• Compression of

simulated / measured data

• Insertion in

Databases

Sensor Value Time Set 4 Set 3 Set 2 Set 1

• Data bases for

– Profile Data (APDB) – Measured Data (MDB) MTTF – Mean Time To Failure

7 Sill Torres - Aging Analysis

• Prediction

– Relate Measured data (MDB) to Profiling Data (APDB) for prediction of current Remaining Useful Lifetime (RUL) – Three Models (Linear, Euclidean Distance, Correlation)

Unintrusive Aging Analysis

Prediction Models

Prediction APDB MDB

8 Sill Torres - Aging Analysis

Results

0 % 20 % 40 % 60 % 80 % 100 % INV c499 c880 c1355 c5315

Accuracy of Prediction

Linear Euclidian Correlation Static

Best (Linear): 90.4%

9 Sill Torres - Aging Analysis

Conclusions

• Methodology for low weight prediction of aging of

integrated systems

• Application of profiling data
• Consideration of varying parameters
• Simulation results: Prediction accuracy ca. 90 %

→ Not exact but – Enables proactive counter measurements – User can be warned

10 Sill Torres - Aging Analysis

Thank you!

www.asic-reliability.com franksill@ufmg.br

Unintrusive Aging Analysis based on Offline Learning

11 Sill Torres - Aging Analysis

Activity Sensor

• [7] R. Baranowski, et al., "On-line prediction of NBTI-induced

aging rates," in DATE 2015, pp. 589-592.

• Monitoring of

switching activity

• f the circuit’s

primary inputs (PI)

• r pseudo-primary

inputs (PPI)

12 Sill Torres - Aging Analysis

Aging

Altera, RELIABILITY REPORT 56, 2013

20 40 60 80 130 nm 90 nm 65 nm 40 nm 25 nm Stratix Stratix II Stratix III Stratix IV Stratix V

FIT (Failures in 109 h)