Efficient Reliability Testing of Emerging Memory ... - ScienceDirect

Available online at www.sciencedirect.com

ScienceDirect Physics Procedia 66 (2015) 568 – 575

C 23rd Conference on Application of Accelerators in Research and Industry, CAARI 2014

Efficient Reliability Testing of Emerging Memory Technologies Using Multiple Radiation Sources William G Bennett1*, Nicholas C Hooten1, Stephanie Weeden-Wright1, Ronald D Schrimpf1, Robert A Reed1, Michael L Alles1, En Xia Zhang1, Michael W McCurdy1, Dimitri Linten2, Malgorzata Jurzak2, Andrea Fantini2 (1)

Electrical Engineering and Computer Science, Vanderbilt University, USA (2) imec, Belgium

Abstract New semiconductor technologies can be difficult and costly to test for radiation reliability. Because the sample space for experiments may be large with new technologies, cost-effect use of specific testing methods can maximize the information obtained while reducing cost. Alternatives to particle accelerator testing are described that can inform later accelerator tests to maximize the use of the available accelerator beam time. Resistive Random Access Memories, a new emerging non-volatile memory, are used as a case study to guide this discussion. © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license © 2014 The Authors. Published by Elsevier B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/). Selection and peer-review under responsibility of the Organizing Committee of CAARI 2014. Selection and peer-review under responsibility of the Organizing Committee of CAARI 2014 Keywords: Type your keywords here, separated by semicolons ;

1. INTRODUCTION Research on radiation effects in microelectronics at the individual device level is often more about understanding the mechanisms that lead to reliability concerns than it is qualifying electronic parts for mission critical applications. When new devices, materials, or processes become available, their radiation response can differ ___________________ * Corresponding Author: Tel.: +1-615-430-5345. E-mail address: [email protected]

1875-3892 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Selection and peer-review under responsibility of the Organizing Committee of CAARI 2014 doi:10.1016/j.phpro.2015.05.076

William G. Bennett et al. / Physics Procedia 66 (2015) 568 – 575

569

drastically from previous generations. Familiar mechanisms can be manifested in new ways, or completely new mechanisms can be seen for the first time. The large sample space of possible mechanisms with a novel device can result in lengthy and costly testing if proper planning is not implemented from the start. This work presents several cost-effective alternatives to particle accelerator testing with the goal of minimizing the testing space during experiments at particle accelerator facilities. An efficient method for characterizing novel devices is described here to give insight into the possible challenges these devices present. An example case using Resistive Random Access Memories (RRAM) is presented to illustrate many of these challenges. RRAM is an emerging memory technology that stores data in switching oxides (as resistance) instead of electronic charge (e.g., in DRAMs). A diagram of one cell and read/write data are shown in Figure 1. The Hf/HfO2 based RRAM discussed here operates by creating oxygen defects (also called the conduction filament) in the HfO2 using strong electric fields across the thin oxide (Chen et al. 2013). When these defects create a path through the oxide, the effective resistance of the cell is lower than that in the original state. The cell can be switched between a High Resistance State (HRS) and Low Resistance State (LRS), allowing the circuit to store 1’s and 0’s, respectively. Here we define the HRS as greater than 100 kΩ and the LRS as less than 10 kΩ, but these values can change depending on system requirements.

2. Background

a) Data from lifecycle testing of a 1T1R RRAM cell. Solid line represent the HRS and LRS thresholds.

b) Drawing of the manufactured 1T1R RRAM resistive element (green) built on top of a 65 nm process NMOSFET Figure 1.

There are two main facets of radiation testing, long term, cumulative effects such as Total Ionizing Dose (TID) and Displacement Damage (DD), and immediate single-particle interaction effects often called Single Event Effects (SEE). TID occurs when incident particles (ions, energetic photons, etc.) generate charge in oxides that remains present long after the event has ended (Schwank et al. 2008). These trapped charges accumulate from many incident particles, and eventually lead to device degradation. Degradation in MOSFETs can result in a number of issues including leakage under isolation oxides, leakage through the transistor channel, or even channel mobility degradation. Other devices (BJTs, HEMTs, etc.) manifest the same degradation in different ways, but the primary source of the degradation is typically the trapped charge in oxides that accumulates over time. Displacement damage is the result of atomic displacements caused by incident particles colliding with lattice atoms in a device (Summers et al. 1993). These displacements typically reduce minority carrier lifetime, alter majority carrier densities, and reduce carrier mobility. In a laboratory setting, heavy ions and protons are often used to study the DD response of devices because they are very damaging. However, DD can also be accumulated from

570


highly energetic particles such as electrons and gamma rays. The rate at which these particles create damage is usually correlated to the incident particle’s Non-Ionizing Energy Loss (NIEL) (Summers et al. 1993). A large body of work exists characterizing the NIEL of a multitude of materials and incident particles, allowing experimenters to choose the appropriate particle for their test conditions. SEEs differ significantly from TID/DD both in their manifestation as well as their test methods. A Single Event Upset (SEU) is a change in a device’s state caused by a single incident particle (Massengill et al. 2007). SEUs occur because energetic particles generate electron-hole pairs as they pass through semiconductor devices. The generated charge transports through the affected device either by diffusion or drift. When this generated charge is collected at a particular node in the system, a brief transient current pulse is generated. This current can cause nodes to change state. For example, an SEU can cause the corruption of data in a memory cell, possibly causing faults in the larger system. The ability of different particles to generate differing amounts of charge in the system is often correlated to the particle’s Linear Energy Transfer (LET) or total deposited charge (Qdep). These factors are then used to compare the small selection of tested ions to a large number of ion species found in on-orbit conditions to predict reliability. Therefore, different particle accelerator facilities, each with its own subset of ions and energies, can be used to emulate on-orbit conditions of interest. 2. RADIATION TESTING SOURCES This section describes several commonly used radiation effects testing sources. This list is not intended to be all-inclusive. Rather, the goal is to discuss sources that are commonly used for novel device characterization. Included here are x-ray generators, radioactive decay button sources, particle accelerators, and lasers. X-ray generators are commonly used for TID testing because they can create trapped charge in oxides without causing atomic displacements (Fleetwood, Winokur, and Schwank 1988). The x-ray generator can be operated by a single experimenter, usually at a fixed photon energy (a commonly used energy is 10 keV), with a controllable flux. Using an x-ray generator is typically a cost-effective method to characterize the TID response of microelectronic devices; however, these setups are often limited in the amount of flux that can be delivered safely to the device, making them non-ideal for characterization at high X-ray fluences. Button and sealed radioisotope sources are an alternative to X-ray testing, and come in a variety of types. The generated particles and gamma rays depend on the elemental composition of the button source. A cesium source, for example, can generate a large amount of TID from gamma rays without creating significant displacements in the device. A Cf-252 source however, can generate a small amount of heavy fission fragments that can create DD in the device, while also emitting a large amount of alpha particles. The variety of particles that can be emitted by some button sources can make them difficult for tests where the incident particle species must be well-known to understand the physical mechanisms underlying the device response. The final sources commonly used for TID/DD damage testing are particle accelerators. Particle accelerators can produce a large flux of protons, neutrons, alpha particles, heavy ions, etc. A large particle flux can allow for devices to be tested at high fluences. Most of the particles will create TID and DD at the same time, which can complicate the interpretation of results. TID is usually tested first using an X-ray generator, and then corroborated using particle accelerators to help differentiate TID effects from DD effects. Particle accelerators are also used extensively for SEE testing (Dodd et al. 2007). The high flux allows for a relatively small target area (e.g. 2 x 2 nm) to be struck regularly by incident particles. The accelerator also allows for a large range of ion LET values for determining a part’s upset threshold. One issue with accelerators is coincidental occurrence of SEE, TID and DD, which can complicate the interpretation of results, as devices may fall out of specification by the end of testing.

571


Source X-ray Generator Particle Accelerator Decay Source Laser

Reliability Test TID TID/DD/SEE TID/DD/SEE SEE

Cost (per hr) $100-$200 $300-$5,000 $ $100-$200

Availability Moderate Difficult Moderate Moderate

Table 1.

Focused laser testing of microelectronics is a more recent test method where laser generated electron-hole pairs are used to emulate generation from an incident ion (McMorrow et al. 2003; McMorrow et al. 2002; Melinger et al. 1994). Lasers used for this work are non-destructive when used appropriately, and typically allow for a single device to be tested extensively without experiencing the device degradation typically encountered during accelerator testing. In some cases, the generated carriers can trigger a destructive response (e.g., laser-induced single event latchup). However, this is not the laser itself causing damage, but rather a destructive device response being triggered by the generated carrier density. Laser testing can be significantly more cost effective than accelerator testing, and, in some cases, can provide experimenters with information concerning the strike-location dependence of the device response, which would be difficult or impossible to obtain using traditional broadbeam particle accelerator testing. Lasers are typically tuned for Single or Two Photon Absorption (SPA/TPA), depending on the bandgap of the semiconductor material being illuminated and the wavelength of the laser light used. TPA will be the concentration here, because of its ability to focus through the backside of the silicon substrate. The advantage of backside testing is the ability to generate carriers under metal overlayers, which is not possible with topside laser generation. Table I summarizes some of the information in this section as a reference for the following sections.

3. TOTAL IONIZING DOSE AND DISPLACEMENT DAMAGE For the example case of radiation effects in RRAM cells, there was a body of existing work on radiation effects in the resistive element (Wang et al. 2010; Bi et al. 2013). The oxides had shown little sensitivity with TID and DD testing, but the literature had focused on the resistive element alone. This was verified for the Hf/HfO2 devices with similar results. When RRAM cells are manufactured they are usually attached to an access transistor that allows for read/write access to the cell from the control circuitry. When the RRAM is operated using the one transistor one resistor (1T1R) topology, it is done with nanosecond width pulses to allow for operation in the 100’s of MHz. Testing of the 1T1R cell for TID effects began with x-ray testing to see if operating the 1T1R cell with high-speed pulses instead of DC sweeps would change the radiation reliability response. The cells showed little degradation with doses in excess of 1 Mrad(SiO2) from 10 keV x-rays. This showed that the cell was resilient to TID and that the access transistor was not extremely sensitive (although it did show a small amount of STI leakage current).

572


a) Motorized stage in the vacuum end station of the Vanderbilt Pelletron with a device mounted in high speed package device.

b) 1.8 MeV proton irradiations are shown with solid markers on the left side, with HRS (green) and LRS (blue) mean values given. Annealing process during post-irradiation read/write operation are represented as open markers. Figure 2.

Promising TID results from the X-ray source informed 1.8 MeV proton irradiations to test for high fluence TID effects and/or DD effects. The end station at ion beam test facilities complicates microelectronic testing because of the need to perform in situ measurements without breaking vacuum or handling the device. In situ measurements are needed for several reasons including: device sensitivity to Electrostatic Discharge (ESD) when handled, time constraints for pumping down vacuum, and the ability to handle parts when they are above the background radiation levels. For the RRAM case, this requires high-frequency SMA connectors on the feed through to be able to send low noise, high-speed pulses to the device while in the end-station chamber. Initial tests were run in small fluence steps, with degradation similar to x-ray TID testing. When the fluence step size was increased, a complete state collapse occurred, shown in Figure 2, where the HRS and LRS states overlap. Figure 2 also shows the recovery of the device after several read/write cycles. This is an important outcome because it demonstrates the


573

importance of varying parameters like the fluence step size, and operating the device as close to the actual operating conditions as possible to understand its radiation response. Had the device been operated with DC sweeps (as it had been previously), this degradation would not have shown up. The mechanism for this failure has been verified to be a cumulative SEE (SEE are discussed in detail in the next section). Once the cell had upset from the HRS to the LRS, continuing strikes on the drain of the transistor forced the RRAM into an LRS failure mode. This failure is caused by forming more than one conductive path in the resistive element (opposed to a single filament created during normal electrical operation) that are subsequently difficult to remove. Once most of the filaments are removed (significant annealing required), the cell recovers back into normal operation. 4. SINGLE EVENT EFFECTS Single Event Effects posed an interesting challenge because resistive elements had be shown to be immune up to any feasible fluence. With the introduction of the 1T1R structure, the previous reliability had to be reconsidered because the access transistor was not immune to SEUs. Since the mechanism that would drive upsets in the RRAM cell had not been identified, backside Two Photon Absorption laser testing at Vanderbilt University was used to verify the cell’s susceptibility. TPA is the best choice due to its low cost and ability to provide strikelocation aware data (Bennett, W. G. et al.). For TPA laser testing to be useful for resistive memories, a beam shutter and chopper had to be installed. This slowed down the repetition rate, and allowed three or fewer pulse to pass through to the device. RRAMs do not have a restoration force like powered memories, so each laser pulse has an effect on the cell that is maintained until the cell is manually reset. If hundreds of pulses are allowed to strike the device, there would be no way to tell which pulse had an effect on the cell’s state. Limiting the number of incident pulses allowed for mechanisms to be seen that would otherwise be hidden in the data. All of the upset mechanisms in RRAM cells were initially seen using backside TPA. Laser data showed that the cell was only vulnerable when a bit line (BL) voltage in excess of 0.7 V was applied to the drain. The cell also upset only from the HRS to the LRS. Transitions from LRS to HRS were not possible even with large laser energies. These two factors are important for the RRAM’s overall reliability because they determine the cell’s ‘window of vulnerability’, because the cell would only be in these specific states a small fraction of its operation time. Another important result was also observed during TPA testing: the ability to accumulate state changes on the resistive element and eventually upset the cell. The upset was termed a Multiple Event Upset (MEU) because instead of taking a single incident ion (i.e., SEU) it required multiple ions to strike the cell to cause an upset. MEUs show up in data like the ion data shown in Fig. 4b. The cell transitions between the HRS and LRS thresholds, before finally transitioning below the LRS threshold. This was an unexpected result during TPA testing that needed

a) XYZ submicron resolution stage at the Vanderbilt TPA Laser test facility.

b) Number of TPA laser exposures required to upset the cell demonstrates the complexity of testing with NVMs that can maintain states between exposures. Figure 3.

574


to be verified using an ion source. TPA laser testing significantly reduces the cost of testing because it can inform heavy-ion tests that may be expensive. Heavy-ion data were collected at Lawrence Berkeley Nation Laboratory’s 88-inch cyclotron facility to verify the vulnerabilities seen during TPA laser SEE testing. The emphasis of the ion experiments was on the BL bias voltage and the LET of the incident particles. Since upsets were only seen to cause transitions in one direction, only a worst case test needed to be run to confirm this process with heavy-ions. Also, since the minimum voltage to upset had already been found to be 0.7 V, higher starting voltages could be used to determine the voltage upset threshold of the cell for different ions. Limiting the experimental sample space is vital to maximizing research output from the allotted beam time. While this may seem obvious, it is often over looked by new experimenters testing at accelerator facilities. MEUs were also seen in heavy-ion data, and, as expected, were a strong function of the incident particle LET. The larger the LET (or equivalent laser energy) the larger the duration of the generated voltage pulse. For the RRAM devices, longer voltage pulses cause more of a state change and make SEUs more likely. When lower LET ions are used, pulses shrink in duration and magnitude, and often require more than one ion to upset the cell. This is demonstrated in Fig.5 where the fraction of MEUs increases as the incident ion LET decreases.

a) Minimum voltage required on the BL of the device required to upset the RRAM cell for different ions at Lawrence Berkeley National Lab

b) Separate exposures during a measuring cycle under heavy-ion irradiation show the ability of the cell to maintain values between the HRS and LRS states. Figure 4.


575

5. CONCLUSIONS X-ray irradiations and TPA laser testing offer low-cost alternatives to particle accelerator testing. While they cannot replace ion-based test methods, the lower cost sources can inform ion testing to maximize research output at low cost. Lower cost accelerator facilities, like the Pelletron at Vanderbilt University, can be used in a similar manner to inform ion tests conducted at higher energy (and cost) facilities. For emerging microelectronics technologies, these facilities are crucial to properly characterizing these devices. Since many of these new technology’s reliability concerns may not be known, there is a large experimental space to cover with limited resources. This was shown using an emerging RRAM technology, but the process of informing expensive experiments with lower cost alternatives applies across the board and a facility prioritized test plan should always be considered when working with emerging technologies. 6. ACKNOWLEDGEMENTS Special thanks to J. A. Kozub for assistance with experiments, Synopsys Inc. for supplying the educational use of their device simulation software, Mike Johnson at Lawrence Berkeley National Lab, and Vanderbilt University for access to their ACCRE computing cluster, TPA laser setup, and ion facilities. Special assistance was also provided by imec, associated with device fabrication. This work was also supported by the Defense Threat Reduction Agency Basic Research Program, and by the Air Force HiREV program. 7. REFERENCES Bennett, W. G., Hooten, N. C., Schrimpf, R. D., Reed, R. A., Mendenhall, M. H., Alles, M. A., Bi, J., et al. “Single Event Induced Upsets in HfO 2 /Hf 1T1R RRAM.” IEEE Trans. Nucl. Sci. Bi, J.S., Z.S. Han, E.X. Zhang, M.W. McCurdy, R.A. Reed, R.D. Schrimpf, D.M. Fleetwood, et al. 2013. “The Impact of X-Ray and Proton Irradiation on $rm HfO_2/rm Hf$-Based Bipolar Resistive Memories.” IEEE Transactions on Nuclear Science 60 (6): 4540–46. doi:10.1109/TNS.2013.2289369. Chen, Yang Yin, R. Degraeve, B. Govoreanu, S. Clima, L. Goux, A. Fantini, G.S. Kar, D.J. Wouters, G. Groeseneken, and M. Jurczak. 2013. “Postcycling LRS Retention Analysis in RRAM 1T1R Device.” IEEE Electron Device Letters 34 (5): 626–28. doi:10.1109/LED.2013.2251857. Dodd, P.E., J.R. Schwank, M.R. Shaneyfelt, J.A. Felix, P. Paillet, V. Ferlet-Cavrois, J. Baggio, et al. 2007. “Impact of Heavy Ion Energy and Nuclear Interactions on Single-Event Upset and Latchup in Integrated Circuits.” IEEE Transactions on Nuclear Science 54 (6): 2303– 11. doi:10.1109/TNS.2007.909844. Fleetwood, D.M., P.S. Winokur, and J.R. Schwank. 1988. “Using Laboratory X-Ray and Cobalt-60 Irradiations to Predict CMOS Device Response in Strategic and Space Environments.” IEEE Transactions on Nuclear Science 35 (6): 1497–1505. doi:10.1109/23.25487. Massengill, L.W., O.A. Amusan, S. DasGupta, A. L. Sternberg, J.D. Black, A. F. Witulski, B.L. Bhuva, and M.L. Alles. 2007. “Soft-Error Charge-Sharing Mechanisms at Sub-100nm Technology Nodes.” In IEEE International Conference on Integrated Circuit Design and Technology, 1–4. doi:10.1109/ICICDT.2007.4299576. McMorrow, D., W.T. Lotshaw, J.S. Melinger, S. Buchner, Y. Boulghassoul, L.W. Massengill, and R.L. Pease. 2003. “Three-Dimensional Mapping of Single-Event Effects Using Two Photon Absorption.” IEEE Transactions on Nuclear Science 50 (6): 2199–2207. doi:10.1109/TNS.2003.820742. McMorrow, D., W.T. Lotshaw, J.S. Melinger, S. Buchner, and R.L. Pease. 2002. “Subbandgap Laser-Induced Single Event Effects: Carrier Generation via Two-Photon Absorption.” IEEE Transactions on Nuclear Science 49 (6): 3002–8. doi:10.1109/TNS.2002.805337. Melinger, J. S., S. Buchner, D. McMorrow, W. J. Stapor, T. R. Weatherford, A.B. Campbell, and H. Eisen. 1994. “Critical Evaluation of the Pulsed Laser Method for Single Event Effects Testing and Fundamental Studies.” IEEE Transactions on Nuclear Science 41 (6): 2574–84. doi:10.1109/23.340618. Schwank, J.R., M.R. Shaneyfelt, D.M. Fleetwood, J.A. Felix, P.E. Dodd, P. Paillet, and V. Ferlet-Cavrois. 2008. “Radiation Effects in MOS Oxides.” IEEE Transactions on Nuclear Science 55 (4): 1833–53. doi:10.1109/TNS.2008.2001040. Summers, G.P., E.A. Burke, P. Shapiro, S.R. Messenger, and R.J. Walters. 1993. “Damage Correlations in Semiconductors Exposed to Gamma, Electron and Proton Radiations.” IEEE Transactions on Nuclear Science 40 (6): 1372–79. doi:10.1109/23.273529. Wang, Yan, Hangbing Lv, Wei Wang, Qi Liu, Shibing Long, Qin Wang, Zongliang Huo, et al. 2010. “Highly Stable Radiation-Hardened Resistive-Switching Memory.” IEEE Electron Device Letters 31 (12): 1470–72. doi:10.1109/LED.2010.2081340.