Integrated passive devices (IPDs) have been widely used in advanced packaging of semiconductor chips, to improve their power integrity and impedance matching. There is a growing demand in guaranteeing signal and power integrity for the chips used in safety-critical products, such as those used in automotive, aviation, industrial, and defense systems, where IPDs help improve quality and reliability of the chips. Therefore, IPD testing and screening itself is essential. Note that the cost of replacing failed IPDs is much higher than the cost of manufacturing them, so screening bad IPDs before mounting is also crucial. In this work, we propose a machine learning (ML) based screening methodology to identifying the IPDs that have potential reliability issues. Based on the parametric data of 360,000 IPDs collected from the wafer probing test, the proposed Semiconductor Quality Net (SQnet) is trained to predict the IPDs which have low breakdown voltage, i.e., low reliability. Keeping the overkill rate below 10%, our method can screen out 6 to 15X more bad dies than the existing industrial methods, i.e., DPAT and GDBC.

This paper introduces an anomaly detection methodology with machine learning for Circuit Probing (CP) using Integrated Passive Device (IPD) as example devices. The IPD can improve the power integrity, performance, and package dimensions of the Integrated Fan-Out Package on Package (InFO-PoP), which is more cost-effective than 3D Integrated Circuits (3DIC) to achieve “More than Moore's law” for mobile devices. Because a defective IPD can invalidate the entire package, the previous test methods are dedicated to very high-end screening for the underkill/failure-escape of high quality and reliable devices. On the other hand, the overkill issues are not concerned yet, which periodically impact the yield and cost. In this paper, we propose a new flow with machine learning methodologies to detect previously ignored anomalies on site-aware wafer-maps for predictive maintenance. The proposed flow covers the overkill and re-test issues to complete the high-quality and cost-effective test methodology with test defense.

Integrated passive devices (IPDs) have been widely used in advanced packaging of semiconductor chips, to improve their power integrity and impedance matching. There is a growing demand in guaranteeing signal and power integrity for the chips used in safety-critical products, such as those used in automotive, aviation, industrial, and defense systems, where IPDs help improve quality and reliability of the chips. Therefore, IPD testing and screening itself is essential. Note that the cost of replacing failed IPDs is much higher than the cost of manufacturing them, so screening bad IPDs before mounting is also crucial. In this work, we propose a machine learning (ML) based screening methodology to identifying the IPDs that have potential reliability issues. Based on the parametric data of 360,000 IPDs collected from the wafer probing test, the proposed Semiconductor Quality Net (SQnet) is trained to predict the IPDs which have low breakdown voltage, i.e., low reliability. Keeping the overkill rate below 10%, our method can screen out 6 to 15X more bad dies than the existing industrial methods, i.e., DPAT and GDBC.

This paper introduces a high quality screening methodology, using Integrated Passive Device (IPD) as an example device, which has the capability to increase power integrity of Integrated Fan Out Wafer Level Chip Scale Package (InFO WLCSP). InFO WLCSP is a more cost effective solution to achieve “More than Moore's law” for mobile devices than 3D integrated circuits (3DIC). Ultra-thin IPD with high capacitance density is capable to further shrinking InFO package dimensions and boosting its performance. Although the cost percentage of an IPD in the whole InFO package is negligible, a defective IPD can cause the whole InFO package malfunctioned. Traditional test methods such as good die in a bad cluster (GDBC) and dynamic part averaging testing (DPAT) are used screening IPD devices, however it is insufficient for high quality screen standards. In this paper, we propose a novel methodology that leads to an effective screening method to guarantee high quality and reliable IPD devices for InFO packages. We will show results on some industrial cases to validate our claims.

Redundancy repair is a commonly used technique for memory yield improvement. In order to ensure high repair rate and final product yield, it is necessary to develop a repair scheme for the coming three-dimensional (3D) architecture of stacked DRAM. According to the JEDEC mobile memory technology roadmap, the interface of 3D DRAM, including the Wide I/O and High-Bandwidth Memory (HBM), is mainly classified as channel-based memories. In this paper, we propose a built-off self-test(BOSR) scheme at the controller level for channel-based 3D memory to enhance final product yield after the bonding of a memory cube to its corresponding logic die. The logic die contains the Channel controller, in which the BOSR circuit resides. Experimental results show that the repair rate is high with higher cluster failure ratio due to the flexible algorithm we choose. The area overhead is low and it decreases significantly when the memory size or channel count increases. The performance penalty is also low due to the parallel execution of address comparison and repair. Moreover, the manufacture cost is lower than conventional DRAM architecture due to allocator-based redundancies. Finally, the proposed scheme can easily be applied to other channel-based 3D memories.

With the increasing demand of super high scale of integration and small form factor in advanced semiconductor products, especially those that integrate DRAM and logic dies, 3D IC and Wafer-Level Chip-Scale Packaging (WLCSP) are considered promising approaches. In Integrated Fan-Out (InFO) WLCSP, a large number of fine-pitch pads, where neighboring pads cannot be probed simultaneously due to insufficient pitch, are used as the contact interfaces of inter-die interconnections. If the fine-pitch pads cannot be probed, the interconnections between the pads and boundary scan cells (BSCs) cannot be tested, which can lead to higher defect level. From industrial investigation, untested fine-pitch pads lead to 1-2% test coverage loss. To improve the overall test coverage, in this paper, we propose a pre-bond probing methodology for fine-pitch pads of InFO WLCSP. By the proposed probing schemes, open/short faults on the interconnects between the fine-pitch pads and BSCs can be all tested by the ATE. Moreover, for short faults that only occur between adjacent pads (interconnects), we propose a grouping method to determine the test patterns at each probing stage, which can minimize the test time. We also show that our method can achieve 100% test coverage of open/short faults.

Due to more aggressive design rules adopted by memories than logic circuits, memories have been considered as the major technology driver of advanced logic circuits, so far as CMOS process technology is concerned. Memory failure pattern identification therefore is important, and is traditionally considered a key task that can help improve the efficiency of memory diagnosis and failure analysis. Critical failure patterns (that are the yield killers), however, may change in different memory designs and process technologies. It is difficult to identify critical failure patterns from high-volume memory failure bitmaps if they are not predefined. To solve this problem, we propose a local parallel search algorithm for efficient memory failure pattern identification. In addition, the proposed system integrates the defect-spectrum-based and coordinate-distance-based methods to identify critical memory failure patterns from a large amount of memory failure bitmaps automatically, even if they are not defined in advance. In our experiment for 132,488 4-MB memory failure bitmaps, the proposed system can automatically identify six critical yet undefined failure patterns in minutes, in addition to all known patterns. In comparison, the state-of-the-art commercial tools need manual inspection of the memory failure bitmaps to identify the same failure patterns.

The three-dimensional integrated circuit (3D IC) is considered a promising approach that can obtain high data band-width and low power consumption for future electronic systems that require high integration level. One of the popular drivers for 3D IC is the integration of a memory stack and a logic die. Because the yield of a 3D IC is the product of respective yields of the mounted dies, the yields of the memory dies and logic die must be high enough, or the 3D IC will be too expensive to be manufactured. To obtain a high yield of 3D ICs, efficient test and repair methodologies for memories are necessary. In this paper, we target the channel-based 3D dynamic random access memory (DRAM) and propose two 3D redundancy architectures, i.e., Cubical Redundancy Architectures 1 and 2 (CRA1 and CRA2). We use Wide-IO DRAM as an example for discussion. In CRA1, spares are associated with each DRAM die as in a conventional 2D architecture. In CRA2, we use a static random access memory (SRAM) on the logic die as spares. Experimental results show that the CRA1 can achieve up to 18% higher stack yield than traditional redundancy architecture with the same area overhead. On the other hand, the CRA2 can achieve the same yield as the CRA1 with 40% less spares, but 1.3% higher area overhead.

Redundancy repair is a commonly-used technique for memory yield improvement. In order to ensure high repair efficiency and final product yield, it is necessary to explore and develop the memory redundancy architecture carefully. However, due to the different failure distributions of memory arrays and various design constraints of memory architectures, it is difficult to explore the efficiency of the memory architecture thoroughly. In this paper, we propose a redundancy architecture exploration methodology to find the redundancy architecture with highest repair rate under redundancy constraints. Given a set of design constraints, failure distributions, and memory architectures, our methodology can explore at least 3(log2M* log2N* log2S) redundancy architectures systematically, where M, N, and S are the address sizes of memory row and column in a die, and the number of slices in the memory cube, respectively. In our experiments, the repair rates of 10 different 3D redundancy architectures with 3 different redundancy analysis algorithms in a given failure pattern distribution are simulated. The experimental result shows that the difference of the repair rates between the most efficient and least efficient memory redundancy architectures is up to 49.42%.

Error correction code (ECC) and built-in self-repair (BISR) schemes have been wildly used for improving the yield and reliability of memories. Many built-in redundancy-analysis (BIRA) algorithms and ECC schemes have been reported before. However, most of them focus on either BIRA algorithms or ECC schemes. In this paper, we propose an ECC-Enhanced Memory Repair (EEMR) scheme for yield improvement. Many modern memories are equipped with ECC in addition to BISR. We evaluate the back-end flow that combines both ECC and BIRA to determine whether yield can be improved by proper sequencing of the two steps. We also collect and identify important failure patterns and their distributions from over 100,000 sample memory instances, which are used to enhance the EEMR scheme that incorporates ECC. As ECC is failure pattern sensitive, careful evaluation from realistic failure bitmaps is necessary. We also verify the feasibility of implementing the proposed EEMR scheme by real test data. Experimental results from industrial 4Mb memory instances show that the proposed EEMR scheme gains over 2% instance yield on average, as compared with the traditional scheme. We also investigate the reliability of the EEMR scheme with different ECC specifications and BIRA algorithms.

Three-dimensional (3D) integration has recently become a popular technology for integrated circuits (IC). 3D IC with the passive silicon interposer is currently the main trend in the industry, especially for processor-memory integration. Evaluating the economic efficiency of test operations in the interposer-based 3D IC thus is important. We propose a cost model for the Die-to-Wafer (D2W) and Die-to-Die (D2D) stacking, including manufacturing cost and test cost. A tool which is based on the proposed cost model is developed. We use this tool for cost analysis and for finding the most cost effective test flow. The results show that, in some applications, test flows including the iterative known-good stack (KGS) test and the pre-bond interposer test significantly reduce the cost, when the KGS test yield is lower than 98.2% and the pre-bond interposer test yield is lower than 99.38%. A Shmoo plot is depicted to show the lower bound of the yield of the final package level test, given the number of stacked dies and the final yield. For different applications, the proposed model evaluates the critical yield or cost values, which helps the designers to determine the most cost effective test flow and the system architecture.

As VLSI technology advances and memories occupy more and more area in a typical SOC, memory diagnosis has become an important issue. In this paper, we propose the Memory Failure Pattern Analyzer (MFPA), which is developed for different memories and technologies that are currently used in the industry. The MFPA can locate weak regions of the memory array, i.e., those with high failure rate. It can also be used to analyze faulty-cell/defect distributions automatically. We also propose a new defect distribution model which has 1–12 times higher accuracy than other theoretical models. Based on this model, we propose a defect-spectrum-based methodology to identify critical failure patterns from failure bitmaps. These failure patterns can further be translated to corresponding defects by our memory fault simulator (RAMSES) and physical-level failure analysis tool (FAME). In an industrial case, the MFPA fits the defect distribution with the proposed model, which has 12 times higher accuracy than the Poisson distribution. With our model, it further identifies two special failure patterns from 132,488 faulty 4-Mb macros in 1.2 minutes.

There is growing need for embedded memory built-in self-repair (MBISR) due to the introduction of more and more system-on-chip (SoC) and other highly integrated products, for which the chip yield is being dominated by the yield of on-chip memories, and repairing embedded memories by conventional off-chip schemes is expensive. Therefore, we propose an MBISR generator called BRAINS+, which automatically generates register transfer level MBISR circuits for SoC designers. The MBISR circuit is based on a redundancy analysis (RA) algorithm that enhances the essential spare pivoting algorithm, with a more flexible spare architecture, which can configure the same spare to a row, a column, or a rectangle to fit failure patterns more efficiently. The proposed MBISR circuit is small, and it supports at-speed test without timing-penalty during normal operation, e.g., with a typical 0.13 μm complementary metal-oxide-semiconductor technology, it can run at 333 MHz for a 512 Kb memory with four spare elements (rows and/or columns), and the MBISR area overhead is only 0.36%. With its low area overhead and zero test-time penalty, the MBISR can easily be applied to multiple memories with a distributed RA scheme. Compared with recent studies, the proposed scheme is better in not only test-time but also area overhead.