Non-Linear Neural Classifier for Testing Analog/RF Circuits

1. A Non-Linear Neural Classifier and its Applications in Testing Analog/RF Circuits* Haralampos Stratigopoulos & Yiorgos Makris Departments of Electrical Engineering & Computer Science YALE UNIVERSITY * Partially supported by NSF, SRC, IBM & Intel

2. Test and Reliability @ YALE http://eng.yale.edu/trela

3. Reliability • Concurrent error detection / correction methods (FSMs) • RTL Concurrent error detection in microprocessor controllers • Design transformations for improved soft-error immunity (logic) • Asynchronous circuits • Fault simulation, ATPG, test compaction for SI circuits • Test methods for high-speed pipelines (e.g. Mousetrap) • Concurrent error detection and soft-error mitigation Current Research Areas • Analog/RF circuits • Machine learning-based approach for test cost reduction • Design of on-chip checkers and on-line test methods • Neuromorphic built-in self-test (BIST)

4. Outline • Testing analog/RF circuits • Current practice • Limitations and challenges • Machine learning-based testing • General idea • Regression vs. classification • Testing via a non-linear neural classifier • Construction and training • Selection of measurements • Test quality vs. test time trade-off • Experimental results • Analog/RF specification test compaction • Neuromorphic BIST • Summary

5. Current practice is specification testing ATE performanceparameters designspecifications test configurations compare chip pass/fail Analog/RF IC Test – Current Industry Practice • Post-silicon production flow Chip Automatic Test Equipment (ATE) Wafer Interface Board

6. Test Cost: • Expensive ATE (multi-million dollar equipment) • Specialized circuitry for stimuli generation and response measurement Test Time: • Multiple measurements and test configurations • Switching and settling times Alternatives? • Fault-model based test – Never really caught on • Machine learning-based (a.k.a. “alternate”) testing • Regression (Variyam et al., TCAD’02) • Classification(Pan et al., TCAS-II’99, Lindermeir et al., TCAD’99) Limitations

7. How does it work? • Infer whether the specifications are violated through a few simpler/cheaper measurements and information that is “learned” from a set of fully tested chips Underlying assumption: • Since chips are produced by the same manufacturing process, the relation between measurements and performance parameters can be statistically learned Machine Learning-Based Testing General idea: • Determine if a chip meets its specifications without explicitly computing the performance parameters and without assuming a prescribed fault model

8. Use machine-learning to approximate these functions unknown,complex,non-linear functions(no closed-form) alternate tests(x1, … , xn) Regression: Classification: • Explicitly learn these functions (i.e. approximate f:x →) • Implicitly learn these functions(i.e. approximate f:x→Y,Y={pass/fail}) Regression vs. Classification Problem Definition: specificationtests(T1, … , Tk) simple functions performanceparameters  designspecifications compare pass/fail

9. training set of chips simpler measurements acquisition of measurement patterns specification tests projection on measurement space pass/fail labels LEARNING learn boundary trained classifier pass/fail label TESTING new (untested) chip measurement pattern Overview of Classification Approach

10. x2 x1 Order of Separation Boundary • The boundary that best separates the nominal from the faulty classes in the space of the simple measurements can be highly non-linear

11. x2 x1 Previous Work: Linear Methods(Lindermeir et al., TCAD’99, Pan et al., TCAS-II’99, Variyam et al., TCAD’00) b1 b2 b5 b6 b3 b4 •  i n, allocate a hyper-plane bi that optimally separates the patterns when labeled with regards to specification i • Acceptance region is bounded by intersection • Decomposition into n two-class problems solved individually

12. Our Solution: Non-Linear Neural Classifier • Allocates a single boundary of arbitrary order • No prior knowledge of boundary order is required • Constructed using linear perceptronsonly • The topology is not fixed, but it grows (ontogeny) until it matches the intrinsic complexity of the separation problem

13. perceptron output geometric interpretation threshold activation function training w adjusts i j to minimize error Linear Perceptron perceptron connectivity synapses

14. y i w i , y 1 w y i , y 1 w - i 2 w i , y y i - y i 1 0 - i 2 - i 1 = x 1 o w i y 1 1 y w y - i 3 i d - i 2 w y i 2 = x 1 + d 1 • Every newly added layer assumes the role of the network output o y 1 = x 1 o w w i d i w 0 i 1 d = x 1 å = 2 x x x o x + d 1 i d 1 = i 1 Topology of Neural Classifier • Pyramid structure • First perceptron receives the pattern x  Rd • Successive perceptrons also receive inputs from preceding perceptrons and a parabolic pattern xd+1=åxi2, i=1…d • Non-linear boundary by training a sequence of linear perceptrons

15. Training and Outcome • Weights of added layer are adjusted through the thermal perceptron training algorithm • Weights of preceding layers do not change • Each perceptron separates its input space linearly • Allocated boundary is non-linear in the original space x  Rd Theorem: • The sequence of boundaries allocated by the neurons converges to a boundary that perfectly separates the two populations in the training set

16. Boundary Evolution Example (layer 0)

17. Boundary Evolution Example (layer 1)

18. Boundary Evolution Example (layer 2)

19. Boundary Evolution Example (output layer)

20. ) % ( e t a r Finding The Trade-Off Point (Early Stopping) • Monitor classification on validation set • Prune down network to layer that achieves the best generalization on validation set • Evaluate generalization on test set Matching the Inherent Boundary Order 100 95 Is Higher Order Always Better? 90 • No! The goal is to generalize • Inflexible and over-fitting boundaries 85 training set validation set 80 75 70 0 2 4 6 8 10 12 14 number of layers

21. x2 x2 x1 x1 LinearlyDependent Are All Measurements Useful? Non-Discriminatory

22. New patterns • Several possible boundaries exist – choice is random • Random label assignment to new patterns Curse of Dimensionality • By increasing the dimensionality we may reach a point where the distributions are very sparse

23. Genetic Measurement Selection • Encode measurements in a bit string, with the k-th bit denoting the presence (1) or the absence (0) of the k-th measurement Fitness function: NSGA II: Genetic algorithm with multi-objective fitness function reporting pareto-front for error rate (gr) and number of re-tested circuits (nr)

24. Inexpensive Machine Learning Test Expensive SpecificationTest MeasurementPattern in Guard-band Few Chips Specification Test Measurements DesignSpecs High-Cost Tester Compare Highly Accurate Pass/Fail Decision Two-Tier Test Strategy Highly Accurate Pass/Fail Decision Simple Alternative Measurements Most Chips Low-Cost Tester Neural Classifier All Chips

25. Reasons for Non-Zero Classification Error • Continuity of analog signals and parametric drifts • Finite training set

26. Guard-bands • Introduce a trichotomy in the measurement space in order to assess the confidence of the decision

27. Testing Using Guard-bands Re-test via specification test Classify chip to dominant class • Examine measurement pattern with respect to the guard-bands • Identify circuits that need to be re-tested via specification tests

28. D D D • Identify the overlapping regions • Clear the overlapping regions • The ontogenic classifier allocates the guard-band Guard-band Allocation Nominal patterns Faulty patterns • Guard-bands are allocated separately

29. Guard-band Allocation Nominal patterns Faulty patterns D D D D • The guard-banded region can been varied by controlling D

30. Experiment 1: Switch-Capacitor Filter Specifications considered • Ripples in stop- and pass-bands • Gain errors • Group delay • Phase response • THD Experiment Setup • N=2000 instances • N/2 assigned to the training set, N/4 to the validation set and N/4 to the test set

31. White-Noise Test Stimulus

32. Test Time vs. Test Accuracy Trade-Off guard-bands test time=15ms specification test time >0.5s

33. Experiment 2: RFMD2411 • 541 devices (training set 341, validation set 100, test set 100) • Considered 13 specs @850MHz (7 different test configurations)

34. Stimulus DUT Response Test Configuration 28 tones noise floor

35. Test Time vs. Test Accuracy Trade-Off guard-bands test time=5ms specification test time ~1s

36. Case-Study: • RFCMOS zero-IF down-converter for cell-phones • Fabricated at IBM, currently in production • Data from 944 devices (870 pass – 74 fail) • 136 performances (65 non-RF – 71 RF) • Non-RF measurements assumed much cheaper than RF, aim at minimizing number of measurements Analog/RF Specification Test Compaction Non-disruptive application: • Instead of “other, alternate measurements” use existing inexpensive specification tests to predict pass/fail for the chip and eliminate expensive ones

37. Setup and Questions Asked Setup: • Split 944 devices into training and test set (472 devices each, chosen uniformly at random. • Repeat 200 times and average to obtain statistically significant results Question #1: • How accurately can we predict pass/fail of a device using only a subset of non-RF measurements? Question #2: • How does this accuracy improve by selectively adding a few RF measurements to this subset?

38. Adding RF performances: • By adding to these 15 non-RF performances 12 RF performances, error drops to 0.38% (i.e. 1 out of 472 devices mispredicted) Results Only non-RF performances: • 15 measurement suffice to predict correctly over 99% of the devices (in our case 4 out of 472 are on average mispredicted

39. What’s Next? Neuromorphic BIST BIST Input Mixed- Signal / RFCircuit Output On-ChipStimulusGenerator ProgrammableResponseAcquisition AnalogMux Input BIST Control TrainableOn-Chip NeuralClassifier BIST Output A stand-alone BIST method for mixed-signal/RF circuits: • On-chip generation of simple test stimulus • On-chip acquisition of simple measurements • On-chip implementation of trainable neural classifier

40. Summary Contribution: • A non-linear neural classifier supporting a novel paradigm for testing analog/RF circuits • Achieves significant reduction in test cost without sacrificing accuracy of specification test • Enables exploration of trade-off between test cost and test accuracy via guard-bands Applications: • Non-disruptive: Specification test compaction • Disruptive: Machine learning-based test • Futuristic: Neuromorphic BIST

41. More Information Publications: • H-G. D. Stratigopoulos, Y. Makris, “Constructive Derivation of Analog Specification Test Criteria,” Proceedings of the IEEE VLSI Test Symposium (VTS), pp. 245-251, 2005 • H-G. D. Stratigopoulos, Y. Makris, “Non-Linear Decision Boundaries for Testing Analog Circuits,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (T.CAD), pp. 1360-1373, Nov. 2005 • H-G. D. Stratigopoulos, Y. Makris, “Bridging the Accuracy of Functional and Machine-Learning-Based Mixed-Signal Testing,” Proceedings of the IEEE VLSI Test Symposium (VTS), pp. 406-411, 2006 • H-G. D. Stratigopoulos, P. Drineas, M. Slamani, Y. Makris, “Non-RF to RF Test Correlation Using Learning Machines: A Case Study,” Proceedings of the IEEE VLSI Test Symposium (VTS), pp. 9-14, 2007 • H-G. D. Stratigopoulos, Y. Makris, “Error Moderation in Low-Cost Machine Learning-Based Analog/RF Testing ,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (T.CAD), pp. 339-351, Feb. 2008 Contact: • E-mail: yiorgos.makris@yale.edu, haralampos.stratigopoulos@imag.fr