We present a solution for reducing the number of defective analog/RF integrated circuits (ICs) that escape detection during manufacturing testing. Also known as underkill, these ICs may fail when deployed in their target application and eventually become customer returns, casting doubt on the effectiveness of the employed test solution and affecting the bottom line. To ameliorate this problem, we introduce an adaptive outlier detection solution that identifies ICs which are suspect of becoming customer returns and proactively bins them as failing. The outlier detection boundary used by our method is dynamically computed based on the performance distribution of devices on each wafer and the underlying model is updated when new ICs are returned from customers and failure analysis confirms that they are indeed defective devices. The effectiveness of our method in reducing underkill while minimizing the incurred yield loss is evaluated using an industrial dataset from Texas Instruments.

As quality expectations of integrated circuits (ICs) continue to rise, contemporary semiconductor manufacturing and test solutions experience increased pressure to prevent any defective parts from being shipped, even if this comes at the cost of sacrificing yield. Known as “overkill”, this lost yield is essentially the result of overly conservative decisions made to compensate for imperfect silicon, imperfect test, as well as uncertainties related to the application wherein a fabricated IC will be eventually deployed. Such decisions are often driven by auxiliary production characterization or quality control tests and processes, which are not directly related to the specifications of a product but, rather, mainly reflect the test environment. Nevertheless, based on these tests and in an effort to err on the side of caution, industry often scraps a small yet not insignificant percentage of perfectly good devices. To address this problem and judiciously recover a portion of the yield that is left on the table without increasing risk, we introduce a machine-learning based solution which exploits the correlation between specification tests and auxiliary tests in order to independently assess confidence in the validity and significance of the latter, for which limits are empirically defined. Effectiveness of our method is evaluated using an industrial dataset provided by Texas Instruments.

To combat the effects of process variation in modern, high-performance integrated Circuits (ICs), various post-manufacturing calibrations are typically performed. These calibrations aim to bring each device within its specification limits and ensure that it abides by current technology standards. Moreover, with the increasing popularity of mobile devices that usually depend on finite energy sources, power consumption has been introduced as an additional constraint. As a result, post-silicon calibration is often performed to identify the optimal operating voltage (Vmin) of a given Integrated Circuit. This calibration is time-consuming, as it requires the device to be tested in a wide range of voltage inputs across a large number of tests. In this work, we propose a machine learning-based methodology for reducing the cost of performing the Vmincalibration search, by identifying the optimal wafer-level search parameters. The effectiveness of the proposed methodology is demonstrated on an industrial dataset.

This special session will discuss how machine learning can transform test. The first talk will review the key challenges and will argue whether contemporary tools, such as deep learning, could offer any advantages over traditional methods. The second talk will focus on extracting useful information from big test data. The third talk will view adaptive test as running machine learning algorithms in real time on big data.

Post silicon trimming is extensively used to counter the effects of manufacturing process variation on certain critical electrical parameters of an integrated circuit (IC). Usually, trimming is performed iteratively by adjusting the resistance value of a trim circuit to specific discrete values. Test programs represent those values by codes and apply common search algorithms in order to find a code which makes a device (optimally) compliant to its design specifications. Consequently, manufacturing yield is increased significantly, yet at the expense of added test time and complexity. In this work, we introduce a novel methodology wherein a trained multivariate model is used to predict, adaptively for each wafer, the optimal starting point of the algorithm that searches for the trim code. Thereby, we seek to minimize the number of code changes that the search algorithm has to perform and, by extension, the overall trim time. In order to provide this prediction prior to wafer sort, so that simplicity of test-floor logistics does not get compromised, the predictive model is built using electrical test (e-test) measurements, which are available before wafer sort, and is trained through measurements from a set of early wafers. Effectiveness of the proposed method in reducing trim time is demonstrated on 370 wafers of an high performance device manufactured by Texas Instruments.

Yield estimation is an indispensable piece of information at the onset of high-volume production of a device, as it can inform timely process and design refinements in order to achieve high yield, rapid ramp-up, and fast time-to-market. To date, yield estimation is generally performed through simulation-based methods. However, such methods are not only very time-consuming for certain circuit classes, but also limited by the accuracy of the statistical models provided in the process design kits (PDKs). In contrast, herein we introduce yield estimation solutions which rely exclusively on silicon measurements and we apply them toward predicting yield during: 1) production migration from one fabrication facility to another and 2) transition from one design generation to the next. These solutions are applicable to any circuit, regardless of PDK accuracy and transistor-level simulation complexity, and range from rather straightforward to more sophisticated ones, capable of leveraging additional sources of silicon data. Effectiveness of the proposed yield forecasting methods is evaluated using actual high-volume production data from two 65-nm RF transceiver devices.

We introduce a machine learning approach for distinguishing between integrated circuits fabricated in a ratified facility and circuits originating from an unknown or undesired source based on parametric measurements. Unlike earlier approaches, which seek to achieve the same objective in a general, design-independent manner, the proposed method leverages the interaction between the idiosyncrasies of the fabrication facility and a specific design, in order to create a customized fab-of-origin membership test for the circuit in question. Effectiveness of the proposed method is demonstrated using two large industrial datasets from a 65nm Texas Instruments RF transceiver manufactured in two different fabrication facilities.

Yield estimation is an indispensable piece of information at the onset of high-volume manufacturing (HVM) of a device. The increasing demand for faster time-to-market and for designs with growing quality requirements and complexity, requires a quick and successful yield estimation prior to HVM. Prior to commencing HVM, a few early silicon wafers are typically produced and subjected to thorough characterization. One of the objectives of such characterization is yield estimation with better accuracy than what pre-silicon Monte Carlo simulation may offer. In this work, we propose predicting yield of a device using information from a similar previous-generation device, which is manufactured in the same technology node and in the same fabrication facility. For this purpose, we rely on the Bayesian Model Fusion (BMF) technique. The effectiveness of the proposed methodology is evaluated using sizable industrial data from two RF devices in a 65nm technology.

We propose a methodology for dynamically selecting an optimal probe-test flow which reduces test cost without jeopardizing test quality. The granularity of this decision is at the wafer-level and is made before the wafer reaches the probe station, based on an e-test signature which reflects how process variations have affected this particular wafer. The proposed method offers flexibility by optimizing test flow per process signature and its implementation is simple and compatible with most commonly used Automatic Test Equipment. Furthermore, unlike static test elimination approaches, whose agility is limited by the relative importance of the permanently dropped tests, the proposed method is capable of exploring test cost reduction solutions which achieve very low test escape rates. Decisions are made by an intelligent system which maps every point in the e-test signature space to the most appropriate probe-test flow. Training of the system seeks to optimize the test flow of each process signature in order to maximize test cost reduction for a given target of test escapes, thereby enabling exploration of the trade-off between test cost reduction and test quality. The proposed method is demonstrated on an industrial dataset of a million devices from a 65nm Texas Instruments RF transceiver.

Since 2004, Texas Instruments and Portland State University have collaborated to develop and deploy test data analytical methods for use in a variety of applications, including quality screening, burn-in minimization, high cost test replacement and/or removal, and operations monitoring. In this paper, key findings amassed during this time are summarized.

Location Averaging is a nearest electrical neighbor method of Statistical Outlier Screening (SOS) for screening outliers from probe test data used at Texas Instruments (TI). Location Averaging is sensitive to variances, not only from die parametrics, but also the test solution. Variance in outlier counts specific to Location Averaging that present either a strong commonality in the test solution, or test pattern on the wafer map, or both, can be indicative of a problem in the test solution. Analysis of outlier counts provides a method of monitoring and maintaining the test solution through data-driven techniques to optimize execution through the reduction of downtime and excursions.

We introduce a methodology for dynamically selecting whether to subject a wafer to a complete or a reduced probe-test flow, while ensuring that the concomitant test cost savings do not compromise test quality. The granularity of this decision is at the wafer-level and is made before the wafer reaches the probe station, based on an e-test signature which reflects how process variations have affected this particular wafer. While the proposed method may offer less flexibility than approaches that dynamically adapt the test flow on a per-die basis, its implementation is simpler and more compatible with most commonly used Automatic Test Equipment. Furthermore, unlike static test elimination approaches, whose agility is limited by the relative importance of the dropped tests, the proposed method is capable of exploring test cost reduction solutions which maintain very low test escape rates. Decisions are made by an intelligent system which maps every point in the e-test signature space to either the complete or the reduced test flow. Training of the system seeks to maximize the number of wafers subjected to the reduced flow for a given target of test escapes, thereby enabling exploration of the trade-off between test cost reduction and test quality. The proposed method is demonstrated on an industrial dataset of a few million devices from a Texas Instruments RF transceiver.

Yield estimation is an indispensable piece of information at the onset of high-volume production of a device. It can be used to refine the process/design in time so as to guarantee high production yield. In the case of migration of production of a specific device from a source fab to a target fab, yield estimation in the target fab can be accelerated by employing information from the source fab, assuming that the process parameter distributions in the two fabs are similar, but not necessarily the same. In this paper, we employ the Bayesian Model Fusion (BMF) technique for efficient yield prediction of a device in the target fab. BMF adopts prior knowledge from the source fab and combines it intelligently with information from a limited number of early silicon wafers from the target fab. Thus, BMF allows us to obtain quick and accurate yield estimates at the onset of production in the target fab. The proposed methodology is demonstrated on an industrial RF transceiver.

We investigate the utility of correlations between e-test and probe test measurements in predicting yield. Specifically, we first examine whether statistical methods can accurately predict parametric probe test yield as a function of e-test measurements within the same fab. Then, we investigate whether the e-test profile of a destination fab, in conjunction with the e-test and probe test profiles of a source fab, suffice for accurate yield prognosis during fab-to-fab product migration. Results using an industrial dataset of ~3.5M devices from a 65nm Texas Instruments RF transceiver design fabricated in two different fabs reveal that (i) within-fab yield prediction error is in the range of a few tenths of a percentile point, and (ii) fab-to-fab yield prediction error is in the range of half a percentile point.

Laser trimming is used extensively to ensure accurate values of on-chip precision resistors in the presence of process variations. Such laser resistor trimming is slow and expensive, typically performed in a closed-loop, where the laser is iteratively fired and some circuit parameter (i.e. current) is monitored until a target condition is satisfied. Toward reducing this cost, we introduce a novel methodology for predicting the laser trim length, thereby eliminating the closed-loop control and speeding up the process. Predictions are obtained from waferlevel spatial correlation models, learned from a sparse sample of die on which traditional trimming is performed. Effectiveness is demonstrated on an actual wafer of laser-trimmed ICs.

This paper introduces an adaptive test method to dynamically control test flow and test contents with continuous per die updates of test fail rates. The method employs Bayesian statistics to model a separate fail rate for each test. Test reordering and elimination is based on statistics of these predicted fail rates and is naturally monitored by a wafer based reset. Wafer sort test response data for two 65nm integrated circuit products is used to demonstrate this method. Test time reductions of about 30% are achieved with quality levels within industry expectations.

Adaptive testing is a generic term for a number of techniques which aim at improving the test quality and/or reducing the test application costs. In adaptive tests, the test content or pass/fail limits are not fixed as in conventional tests, but dependent on other test results of the currently or previously tested chips. Part-average testing, outlier detection, and neighborhood screening are just a few examples of adaptive testing. With this Embedded Tutorial, we are offering an introduction to this topic, which is hot in the test community, to the wider DATE audience.

Quality improvement and cost reduction in the overall IC manufacturing and test processes are being continuously sought. Outlier screening methods can address both of these needs. As technology scales, it has become increasingly difficult to screen outliers without excessive Type I or II errors. Hundreds of parameters are collected at wafer probe, but there lacks a systematic way of selecting outlier screens. In this paper we describe a statistical approach to both identify outliers and select beneficial screening parameters more effectively. Results on a 90 nm design to reduce the burn-in fails are described.

A burn-in reduction screen is presented based on a subset of measurements from the wafer sort data. The subset of wafer sort data is identified by a principal component analysis (PCA) using die passing all the tests. This paper addresses the curse of dimensionality and shows the importance of reducing the dimensions in segregating the potential reliability fails for additional screening from the normal healthy population. The method presented is demonstrated using 90nm volume data and compared to 8 hour burn-in experiments