l i a b l eh k C o m p u t i n gL a b o r a t o r y

Test Economics for Homogeneous Test Economics for Homogeneous Manycore SystemsManycore Systems

Lin Huang† and Qiang Xu†‡

†CUhk REliable computing laboratory (CURE)

Department of Computer Science & Engineering

The Chinese University of Hong Kong

‡CAS-CUHK Shenzhen Institute of Advanced Integration Technology

Observations on Manufacturing Test Observations on Manufacturing Test CostCost

Manufacturing test is responsible for achieving sufficient high defect coverage

As technology advances … Test patterns that target more kinds of errors become essential Accelerated testing methods (e.g., burn-in test) becomes difficult

Manufacturing test cost – a great share of production cost In particular, burn-in cost can range from 5-40% of production cost

If we are able to relax the coverage requirement, manufacturing If we are able to relax the coverage requirement, manufacturing cost can be dramatically reducedcost can be dramatically reduced

Manycore Processor Era Provides us An OpporManycore Processor Era Provides us An Opportunitytunity

The integration of a large number of cores on a single silicon die Increasingly popular in the industry

Traditional yield-driven redundantyield-driven redundant cores aims to improve the manufacturing yield

We propose to introduce a few test cost-driven redundanttest cost-driven redundant cores in addition to yield-driven spares for test cost reduction

If test cost reduction exceeds the manufacturing cost increment, If test cost reduction exceeds the manufacturing cost increment, the total production cost can be reducedthe total production cost can be reduced

Manycore Processor Era Provides us An OpporManycore Processor Era Provides us An Opportunitytunity

If test cost reduction exceeds the manufacturing cost increment, If test cost reduction exceeds the manufacturing cost increment, the total production cost can be reducedthe total production cost can be reduced

Consider a 16-core processor

To guarantee that all 16 cores work well provided they pass manufacturing test, we need …

– Very high defect coverage to identify killer defects– Sufficient burn-in to weed out chips with latent defects

Manufacturing test is responsible for 16 out of 20 cores (instead of all 20 cores) to work

– Defect coverage requirement can be lowered– Burn-in test can be reduced or eliminated– Manufacturing cost increases

AgendaAgenda

Background

Test Economics with Partial/No Burn-In

Test Economics with Partial Manufacturing Test

Experimental Results

Conclusion

Basics in Yield ModelingBasics in Yield Modeling

Defects on chip – Negative-binomial distribution

Defect type Killer defects Latent defects

Bathtub curve

Fai

lure

Rat

e

TimeInfant

MortalityUseful Life Wearout

Full Burn-InPartial Burn-In

Problem 1Problem 1[Partial Burn-In][Partial Burn-In]

Enable partial/no burn-in test only

Given defect coverage requirement, we consider to introduce redundant cores into manycore system that functions if no less than cores are defect-free We fabricate cores on a chip Chips with all cores pass test are sold out Eventually we need to guarantee cores are

defect-free at the end of infant morality

Determine the number of burn-in driven spares and burn-in time such that … The production cost per sold chip is minimized Product quality constraint is met

The Impact of Partial Burn-InThe Impact of Partial Burn-In

The reliability induced by latent defects follows Weibull distribution with decreasing failure rate

Assume that all latent defects reveal themselves after full burn-in time

Product Quality and Chip Test YieldProduct Quality and Chip Test Yield

Product quality requirement The probability that a sold chip actually functions at the end of infant

mortality should be higher than a threshold – no less than cores on a chip is defect-free at the end of infant

mortality – all cores on a chip pass manufacturing test after (partial) burn-in

Chip test yield

Product Quality and Chip Test YieldProduct Quality and Chip Test Yield

All Fabricated Cores

Defect-Free Cores Before Burn-In

Defect-Free Core After Partial Burn-In Time T

Defect-Free Cores at the End of Infant Mortality Period

u m t j

iIMi

All Fabricated Cores

Defect-Free Cores Before Burn-In

Defect-Free Core After Partial Burn-In Time T

Defect-Free Cores at the End of Infant Mortality Period

u m t j

iIMi

Product Quality and Chip Test YieldProduct Quality and Chip Test Yield

Define – -out-of- cores are initially defect-free – cores in that set maintain defect-free after burn-in time

We obtain

Cost ModelCost Model

Simple yet effective cost model – capture the key impact of introducing burn-in driven redundancy

Manufacturing cost – normalize to the case that manufacturing cost of each core for manycore chips without redundancy is 1 unit

ATE cost – ATE cost per fabricated core is unit

Burn-in cost – normalize the cost of fully burn-in process as unit and assume it is proportional to the burn-in time

Case Study on Partial/No Burn-InCase Study on Partial/No Burn-In

Homogeneous manycore system that functions with no less than 32 defect-free cores

Product quality requirement is set to 500DPPM

Problem 2Problem 2[Partial Burn-In & Relaxed Defect [Partial Burn-In & Relaxed Defect

Coverage]Coverage] Not only enable partial/no burn-in test but also relax the defect

coverage for core tests

We introduce test cost-driven spares and yield-driven ones We have totally identical

cores on chip Chips containing no less than pass-

test cores are shipped out Eventually we need to guarantee cores

are defect-free at the end of infant morality

Problem 2Problem 2[Partial Burn-In & Relaxed Defect [Partial Burn-In & Relaxed Defect

Coverage]Coverage] Determine the number of test cost-driven spares, number of

yield-driven spares, defect coverage for core test, and burn-in time such that … The production cost per sold chip is minimized Product quality constraint is met

The Impact of Test Decision CriterionThe Impact of Test Decision Criterion

Ideally a prefect manufacturing test is able to reject all bad cores while accept all defect-free ones and

In reality …

Test escapes

False rejects

Product Quality with False RejectsProduct Quality with False Rejects

Redefine – no less than cores on a chip is defect-free at the end of infant

mortality – no less than cores among all cores on a chip pass

manufacturing test after (partial) burn-in

Similarly, we have

Product Quality with False RejectsProduct Quality with False Rejects

Notations – -out-of- cores are initially defect-free – cores in that set maintain defect-free after burn-in time – among good cores on a chip, pass the test – among bad cores, pass the test

We have

Cost ModelCost Model

Total production cost

ATE cost depends on defect coverage

Experimental SetupExperimental Setup

Homogeneous manycore system that functions with no less than 32 defect-free cores (i.e., )

The best , and combination in terms of production cost is determined by exploring solution space

System parameters , , , , , , Product quality requirement is set to 500DPPM

Tradeoff between Burn-In Cost and ATE Tradeoff between Burn-In Cost and ATE Cost under Product Quality ConstraintCost under Product Quality Constraint

High defect density Low defect density

Comparison between Traditional and Comparison between Traditional and Proposed StrategyProposed Strategy

22.28%

Comparison between Traditional and Comparison between Traditional and Proposed StrategyProposed Strategy

25.26%

Comparison between Traditional and Comparison between Traditional and Proposed StrategyProposed Strategy

ConclusionConclusion

We propose to introduce spare cores into manycore system Burn-in test time can be shorten Defect coverage requirement can be relaxed Without sacrificing quality of the shipped products

We develop novel analytical models to verify the effectiveness of the proposed strategy

Test Economics for Homogeneous Manycore STest Economics for Homogeneous Manycore Systemsystems

Thank you for your attention !Thank you for your attention !