Outline

1. A DFT Approach for Diagnosis and Process Variation-Aware Structural Test of Thermometer Coded Current Steering DAC's Rasit Onur Topaloglu and Alex Orailoglu{ rtopalog | alex }@cse.ucsd.eduUniversity of California, San DiegoComputer Science and Engineering Department 9500 Gilman Dr., La Jolla, CA, 92093

2. Outline • Current Steering Digital to Analog Converters 101 • A Process Variation-Aware Soft Fault Model • Process Variation Estimation • Reduction of Diagnosis Time Using Design for Testability Hardware • Experimental Results

3. Introduction • Higher precision applications drive Digital to Analog Converter (DAC) resolutions to higher bits day by day • Higher bit resolutions increase circuit complexity, hence increase test time and difficulty • In thermometer coded circuits, controllability is limited as each bit increment sums current of a new source with previous ones • Diagnosis of a fault or test is usually handled by exhaustively trying all input codes

4. Binary-Coded Current Steering Principle • Input digital code selects current sources to be added to analog output • Iout is the analog output • Current sources are in fact implemented by current mirrors using a common on-chip reference current 4I 2I I (110) input shown for an 8-bit binary CSDAC (Current Steering DAC)  Iout

5. Individual Current Distributions • Soft faults for exponentially valued current sources would contribute integral and differential non-linearity (INL and DNL) degradation during certain transitions e.g. transition from 2^n-1 to 2^n W1/L1 W2/L2 W1/L1 W2/L2 4I Iref I Iref 4I I

6. Impact of Binary Coding on INL and DNL Analog • Transitions from 2^n-1 to 2^n activate totally differents sets of current sources • Due to limited spatial correlation between these groups, DNL will tend to get larger, which implicitly tend to enlarge INL • In thermometer-coded (TC-CSDACs), in these transitions, one more current source is added only, and hence outputs of these two codes highly correlated due to the 2^n-1 common elements INL: max difference between overall real and ideal lines DNL: max of stepwise differences Digital 7 8

7. Thermometer-Coded CSDACs (TC-CSDACs) • (110) in binary is (0111111) in thermometer code • Equal weighting of current sources prevents significant impact for faulty sources • Error correction capability is another attractive reason for choosing thermometer code ex:0111011 not possible as 1’s should be consecutive I I I I I I I I  (0111111) input for an 8-bit thermometer coded CSDAC Iout

8. Diagnosis Restriction of TC-CSDACs • Current sources indexed with fixed bit positions • Fixed indexes imply a controllability restriction • Diagnosis time for a faulty current source exponentially increases as compared to binary coded CSDACS I I I I I I I I  (0111111) input for an 8-bit thermometer coded CSDAC Iout

9. 1 12 15 3 13 5 7 10 9 8 6 14 4 16 11 2 Design Considerations for TC-CSDACs • Current sources laid out in common-centroid layout style to minimize impact of process variations • A number of most significant bits (MSB’s) and least significant bits (LSB’s) are grouped within themselves to further reduce process variation impacts ex: In a 16-bit converter, current sources indexed by consecutive number laid out on separate corners

10. m MSB’s are interpolated by n LSB’s where B, total number of bits, is m+n Input to the TC-CSDAC is binary, hence binary to thermometer decoders used in the circuit Proposed fault model can be applied to MSB and LSB parts separately A Practical TC-CSDAC Architecture

11. Process-Aware Structural Soft Fault Model • Process variations should not be mistaken as faults • The proposed fault model: one current source might have an additional deviation from process variation effected value due to any modeled or un-modeled fault probability For each die, a current source will have a fixed value picked up from its probability density function, caused by process variations Isource I

12. Estimation of Process Variations • Current sources are systematically correlated due to their close locations on die • Current sources can be represented as a sum of independent components through a technique called Principal Component Analysis (PCA) • Principal components corresponding to largest eigenvalues account for most of the variation • Ratio of selected eigenvalues to all eigenvalues can be used to ensure a minimum variation I : normalized current source variables U : eigenvectors of correlation matrix C : principal components

13. Estimation of Process Variations • A reduced number of principal components, M<N, is equivalent to deleting some of the columns • Then, M of these equations can be chosen to obtain an M equation-M unknown system • The choice is made for consecutively indexed sources, as each source individually requires two measurements due to controllability restriction 

14. Process Variation-Aware Test Nominals • M sources are measured for each chip, U is calculated from correlation matrix, hence only C values are left to be determined • Once C values are calculated, unmeasured N-M source I values can be calculated • Hence, these steps provide process-variation aware nominal values for each current source using few measurements, as N>>M even for 98% variation 

15. Acquiring Principal Components On-Chip • Analog current is measured for up to principal component number of times; as low as ~ 6 measurements satisfactory to account for 98% variation • No additional hardware is required to take these measurements, for ex. 6 consequent input codes, (0..0000000),(0..0000001),(0..0000011),.., (0..0111111), can be used to get these measurements

16. Correlation Model • Output of a current source is spatially correlated to neighboring sources on layout as a result of silicon manufacturing steps • A spatial distance2 correlation model is used • According to the correlation model, the correlation starts from a number close to 1 and decreases towards 0 with distance between each pair of sources

17. Design for Test (DFT) Hardware • One more decoder and some combinational gates added to the original decoder • Similar modification done for row selection hardware • test_sel=0 : original mode • test_sel=1 : one column is selected using Ci inputs and setting row_sel=1

18. 1 12 15 3 13 5 7 10 9 8 6 14 4 16 11 2 Reduction of Diagnosis Time using DFT • Instead of exhaustively measuring current sources, particular groups of them are summed & measured • This reduces the diagnosis time from quadratic to linear • Process-variation aware nominal test points are used for each source to create variation aware nominals • One row selected such that the sum of current sources within are deviating from the average of remaining row sums; similarly for columns

19. Experimental Results • Even a minor 20% deviational soft fault around process variation estimated values can be caught with ~100% efficiency!

20. Error Rates for Process Estimation • Examination of normalized error in last column reveals that difference between real and estimated values are almost negligible using 6 principal components

21. Robustness for Increased Requirements • Increasing bit requirements indicate detection of lower deviational faults due to averaging of non-faulty sources approaching the population mean

22. Conclusions • A process variation aware DFT method is proposed • Even minor soft faults can be caught with the proposed technique due to accounting of process variations • A fast diagnosis procedure is proposed with reasonable addition of DFT HW • The proposed technique becomes more robust for increased bit requirements