Testing RRAM and Computation-in-Memory Devices: Defects, Fault Models, and Test Solutions

Resistive random access memory (RRAM) is a promising emerging memory technology that offers dense, non-volatile memories that do not consume any static power. Furthermore, RRAMdevices can be written and read out in nanoseconds, and it is possible to use them to performcomputation-in-memory (CIM). These benefits make this technology a potential replacement for Flash or even dynamic random access memory (DRAM). This is also clearly seen by the community; both universities and companies are prototyping RRAMs, and there are already some commercial RRAMs available. In order to deliver high-quality products, the RRAMs need to be tested properly so that a manufacturer can guarantee the quality. This dissertation focuses on test development for RRAMs. Traditionally, production defects in memories, such as DRAM, are modeled as linear resistors in or between two nodes of the circuit. In literature, many researchers have applied a similar approach for RRAMs. However, we demonstrate that this method of modeling defects is inappropriate, because the models fail to describe the defective behavior of the RRAM device. Instead, those models describe defects in the interconnections that surround the RRAM device. To overcome this, we propose the Device-Aware Test (DAT) approach that consists of three steps. First, the approach models the actual physics of defective devices and thus leads to realistic defectmodels. Second, the defect models are used to performaccurate faultmodeling and analysis. Third, the results from this step are used to develop high-quality RRAM tests. We do this by first characterizing the defect. We analyze the complete production process of a RRAM. During this analysis, we identify what can go wrong in every step, and in what kind of defects this may result. All identified defects need to be properly modeled, so that a high-quality test can be developed.Next,we analyze howit affects the performance of a defect-free device, and incorporate the resulting defective behavior in a compact defect model. This model is calibrated and accurately describes the effects of the defect. Second, we apply this defect model in a RRAM circuit to perform fault modeling and analysis. We systematically define the complete space of all faults that could occur, and then apply an analysis methodology to validate which faults actually occur in the circuit. Third, we develop a test for the validated faults. This test only needs to detect faults that are actually sensitized, and thus is shorter than generic tests, while it also has a better fault coverage. We apply the DAT approach to RRAM forming defects and RRAM intermittent undefined state faults. The results show that these two defect models sensitize different faults than the traditional defect models do. Since the DAT defect models describe the actual physics of the defects, we can conclude that the traditional approach will lead to lowquality tests that generate test escapes and reduce the production yield. Furthermore, we demonstrate that the faults cannot easily be detected by existing test algorithms, and that special tests need to be developed to detect them. We also apply the DAT approach to a RRAM-based computation-in-memory (CIM) architecture to develop a test for it. We shows that a CIM device needs to be tested both in its memory and computation configuration, as there are unique faults in both configurations. We define the complete fault space for CIM faults and validate it using the DAT approach. Subsequently, we develop a test that detects the faults in both configurations. Furthermore, we study how process, voltage and temperature variations affect the performance of the CIM architecture. We demonstrate that certain operations are more susceptible to these variations than other ones.