Single-event-upset and alpha-particle emission rate measurement techniques

Single-event upsets (SEUs) are non-reproducible errors caused by radiation, such as an alpha-particle from packaging materials or the collision of a cosmic ray neutron with an atom in a semiconductor integrated circuit. Typical alpha-particle sources include C4 (controlled collapse chip connection) solder bumps, thorium and uranium impurities in underfill or molding compounds, as well as impurities introduced during wafer processing. Low-alpha-particle materials are used to minimize the SEU rate; a typical alpha-particle flux for processed wafers, low-alpha-particle underfill, and low-alpha-particle lead is <1 count/khr-cm², <2 counts/khr-cm², and <20 counts/khr-cm², respectively [1]. The expanding set of process materials can introduce new alpha-particle sources, such as ¹⁷⁴Hf in high-k gate dielectric materials. However, because of its long half-life (2 × 10¹⁵ years), the ¹⁷⁴Hf isotope is less of a concern than alpha-particle-emitting impurities in hafnium sputtering targets. The cosmic ray flux also cannot be ignored because it depends strongly on altitude and weakly on geomagnetic location and the solar cycle [2, 3]. In addition, building materials offer some neutron shielding.

In this paper, we discuss three topics related to SEUs: 1) detection methods used for measuring low levels of alpha-particle radioactivity in packaging materials, 2) system life testing, and 3) accelerated testing of devices. We review the alpha-particle detectors currently in use at IBM, as well as their calibration, background detection, and monitoring strategy. Some of the research activities related to fabricating new-generation, large-area, low-background alpha-particle counters are presented, as well as a review of system life-testing methods, including tests performed at high elevations (high neutron flux¹), at sea level, and in a deep mine shaft, with results compared to simulation models. Methods used for accelerated testing of device arrays using a variety of alpha-particle sources are presented. These sources include radioactive foils at the wafer level, a collimated beam of alpha-particles from a Tandem Van de Graaff accelerator, and hot underfill (HUF), where an alpha-particle emitter (²¹⁰Pb or a combination of ²¹⁰Pb and ²¹⁰Po) is mixed into the underfill and applied to a chip package. The testing strategy and techniques used are described, as well as the results in terms of the bit-fail cross-section, the failure-in-time (FIT) rate, and the critical charge.

In order to screen packaging materials for alpha-particle emission, alpha-particle flux measurements were conducted at the IBM Thomas J. Watson Research Center and at the IBM 300-mm and 200-mm fabrication facilities.² Three types of alpha-particle detectors were used: Gas proportional counters were used for large-area thin samples; ZnS scintillators were used for irregularly shaped or thick samples; and solid-state detectors were used to obtain energy spectra. Table 1 shows the advantages and disadvantages of each of these counters.

Large-area gas-filled proportional counters (from Alpha Sciences [4]) were generally used for thin samples, such as wafers. The counters we used include two models, with active areas of 825 cm² and 1,000 cm². The background counting rates of these counters (i.e., in the absence of a sample) are in the range of about 1.5–3.0 counts/hr. These detectors were generally used to sample lead (Pb) that had been plated or wafers with Pb-based C4s in order to monitor the activity of the plating process. Lead-plated samples from vendors were routinely examined before accepting orders of lead. Typically, samples were measured soon after they were plated and 90 days later so that the increase in activity at secular equilibrium could be accurately estimated [5]. A drawback to this detector design was the thin aluminized Mylar** window that the alpha-particles must pass through in order to ionize the gas inside. This thin window is delicate and sensitive to vibrations.

For thick or irregularly shaped samples, ZnS scintillators were used. Typical samples included Pb pellets that were used for plating, ceramic substrates, or bulk materials. Each detector had a scintillator that is 125 mm in diameter, approximately 100 μm thick, and viewed by a photomultiplier tube. The detector assemblies were in a light-tight environment and were filled with N₂ gas to displace radon. Typical backgrounds for these counters are on the order of 3 counts/hr.

Additionally, large-area (20-cm²) silicon detectors were used in a vacuum chamber to measure the activity and the energy distribution of the alpha-particles emitted on flat samples. Because of the limited area and the typical low count rates, we had to measure samples for weeks at a time in order to get statistically meaningful data.

For samples with low alpha-particle activity, the average count rate was measured both with the sample in the counter and with the sample out of the counter (background). The difference was attributed to the sample activity. Figure 1 presents the results of a calculation by showing the amount of time required to measure a sample as a function of the count rate (counts/hr) in the sample for several different background counting rates in the range of 0 counts/hr to 4.0 counts/hr, with σ = 0.5 (i.e., a statistical uncertainty of 50%) and a 100-hour background measurement. The black line shows combinations of counting time as a function of count rate such that a total of 4 counts were measured, giving a statistical uncertainty of 50% for the ideal case of no background counts.

Figure 1

As an example, with a background counting rate of 4.0 counts/hr and a sample counting rate of 0.6 counts/hr, it would take approximately 100 hours of measurement time to obtain σ = 50% on the measurement. On the basis of Figure 1, it would be nearly impossible to measure the activity of a sample with a counting rate of <0.2 counts/hr using a counter with a background count rate of >1.0 count/hr. Further modeling has shown that the measurement time can be reduced from about 200 hours to about 80 hours for a sample with an activity of 0.25 counts/hr if the background measurement time is increased to 400 hours and a 50% error is still maintained. Therefore, background in the counters must be reduced in order to measure low alpha-particle activity with few statistical errors in a reasonable amount of time. For most samples, 100-hour measurement times were used.

A thick ²³²Th source was used to monitor the detection efficiency of these detectors. The absolute activity of the source was measured in a silicon detector; the energy spectrum of this source is shown in Figure 2. Because the source was thick, there were no individual peaks from the decay chain. The same source was placed on the center of the sample trays in the gas proportional counters and on the center of the scintillators in the ZnS counters, and the count rates were measured in each. The lower-level discriminators on the amplifiers for the detectors were set at roughly 1 V, which was well above the noise level and, thus, effectively eliminated counting minimally ionizing beta-particles in the gas proportional counters. The ratio of the count rate from the ²³²Th source in these detectors to that measured in the silicon counter represents the efficiency, which was on the order of about 85%. The efficiency was used to correct the background-subtracted count rate as follows:

Figure 2

The alpha-particle energy spectrum from a proprietary ceramic material was measured on a 20-cm² silicon surface barrier detector as shown in Figure 3. Since the count rate was so low, the background had to be measured and subtracted out. In this case, in order to obtain statistically meaningful results, we had to collect the background and ceramic data for a period of 10⁶ seconds (11.5 days). The shape of the spectrum in Figure 3 is qualitatively similar to that of the thorium (Th) spectrum in Figure 2, that is, trace amounts of thorium are contained in the ceramic material.

Figure 3

Currently at IBM, two low-background alpha-particle counter development projects are underway: 1) an ionization counter and (2) a segmented silicon detector.

The ionization counter under development is cylindrical and is designed to accommodate 300-mm wafers, with guard rings that make the electric field uniform, thereby providing nearly constant detection efficiency from the center to the outer diameter of the wafer. Additionally, the counter is constructed from aluminum so that it can be evacuated to eliminate radon gas infiltration. Since the alpha-particle emissivity from aluminum is unacceptably high, the inside of the counter is passively shielded with an ultralow-emissivity material.

The development of a large-area, low-background, segmented silicon detector is also underway. For the prototype detectors, a 200-mm diameter, high-resistivity n-type substrate was segmented into 32 detectors of 6.25 cm² area. Individual detectors were tested for leakage current, current–voltage (I–V) characteristics, and functionality using an ²⁴¹Am source. Steps were taken to ensure that the manufacturing process would not add residual radioactivity to the detectors. The background from these detectors was comparable to that of commercially available silicon detectors. Figure 4 shows a prototype wafer with segmented detectors [6]; a mechanical probe was used to make the electrical connection to the detector of choice for individual testing.

Figure 4

In order to validate product SEU rate estimates based on accelerated testing using external particle sources and Monte Carlo theoretical simulations, SEU life tests were performed on SRAM product-like vehicles. The life test is a nonaccelerated (i.e., no external source) test in which the SRAM arrays are run under standard operating conditions. To differentiate the cosmic ray and alpha-particle SEU components, we performed tests at the following three locations: Burlington, Vermont, to provide reference SEU rate values near sea level (100 ft); Leadville, Colorado, to provide a high-altitude (10,200-ft) SEU in an enhanced cosmic ray-level environment; Kansas City, Missouri, in a quarry in order to isolate the alpha-particle component in an environment with greatly reduced cosmic flux. The neutron flux at Leadville, Colorado, was approximately 13 times greater than that at sea level at New York City [2, 3]. Consequently, life-test results confirm the results of both alpha-particle and cosmic ray SEU models. Furthermore, unexpected radiation sources can lead to high SEU rates in life tests. Figure 5 shows an example of results from an IBM product under life testing in which hundreds of modules were operational. The cumulative number of fails and the SEU rate are shown as a function of numbers of hours of testing. Because this is not an accelerated test, the failure rate is very low. Therefore, it can take more than 1,000 hours of life testing to obtain statistically meaningful data. Figure 5 shows that the SEU rate (right-hand scale) was larger than the asymptotic value when the data were collected over too short a period of time.

Figure 5

Additional data is shown in Figure 6 and is compared to model calculations for the modules under test. These data points were taken after many hours of life testing, so the results are statistically meaningful. The life tests in this example were run between 2,000 and 5,000 hours at each facility. The data is in reasonable agreement with the calculations. The SEU model was based on Monte Carlo simulations of the cosmic ray neutron, low-alpha-particle solder bumps, and low-alpha-particle underfill SEU components, using the SEMM-2 simulation tool [7–9]. SEMM-2 simulations were correlated with radioactive foil and high-energy proton beam tests, which are described below (see [10, 11]).

Figure 6

The IBM T. J. Watson Research Center has a 3-MV Tandem Van de Graaff accelerator that was used for SEU experiments and material analysis. The beam is collimated and monoenergetic, providing a different radiation environment for these studies compared to that of a radioactive source. The laboratory has seven beamlines, one of which is dedicated to SEU exposures [other lines are used for nuclear reaction analysis (NRA), Rutherford backscattering (RBS), hydrogen profiling using elastic recoil detection analysis (ERDA), and proton-induced x-ray emission (PIXE)]. The ion sources for the accelerator consist of a sputter source [Source of Negative Ions by Cesium Sputtering (SNICS)] [12] for ¹H, ²H, ¹²C, and other beams, as well as a radiofrequency source (alphatross) used to produce the alpha-particles (⁴He). The devices being irradiated must be wire-bonded (not fully packaged) because the beam is not very penetrating at the maximum energy used (E < 10 MeV). Within the SEU vacuum chamber, the chip was positioned on a goniometer, where the X/Y position and rotational orientation can be adjusted with respect to the defocused beam. The size of the defocused beam, at the plane of the device under test (DUT) was about 2 cm × 2 cm. In order to monitor the particle flux, a silicon detector, with a 0.5-mm Ta aperture, can be rotated into the beam before and after an SEU exposure. Typical particle fluxes are approximately 5 × 10⁶ ions/cm²-s. Figure 7 shows the beam flux monitor, and the DUT is shown rotated with respect to the beam and the ZnS viewing screen, at the rear of the photograph.

Figure 7

In typical experiments, several beam energies were used. The minimum energy was determined by the requirement that the beam penetrate through the back-end-of-line (BEOL) metal layers (consisting of the wiring levels and the various dielectric materials). The number of failures (bit flips) is recorded as a function of the device voltage at several angles between the beam and the chip. The angle at which the bit flips begin to occur is determined as a function of device voltage and beam energy. Given this critical angle, beam energy, and BEOL composition, the deposited charge in the sensitive collection volumes can be calculated and compared to the device model. On the basis of detailed simulations of particle transport through the BEOL materials, a simple method has been developed to extract the critical charge for silicon-on-insulator (SOI) technologies from the ion beam measurements [13]. Details of the modeling technique are described in References [14, 15]. Figure 8 shows bit-fail cross-section data as a function of the angle between the beam and the normal to the chip for an experimental IBM 65-nm SOI medium-power latch. The three sets of data correspond to three operating voltages. The bit-fail cross-section is defined as the number of upsets divided by the product of the number of bits and the incident particle fluence (the number of particles per unit surface area of the chip being tested).

Figure 8

Radioactive sources of alpha-particles were used to irradiate devices to determine their susceptibility to alpha-particle-induced SEUs. The sources used are thick ²³²Th foils and thin electroplated films of ²⁴¹Am. The ²³²Th source provided alpha-particles with a maximum energy of approximately 9 MeV, as shown in Figure 2, whereas the ²⁴¹Am source emitted alpha-particles predominately at 5.486 MeV.³ The flux from the ²³²Th source was about 70 alpha-particles/s-cm², and the activity of the ²⁴¹Am source was about 0.1 μCi (3 × 10³ alpha-particles/s).

The ²³²Th source was important for the SEU studies because thorium contamination is present in ceramics, underfill, and other materials used in packaging. The predominant alpha-particle energy from the ²⁴¹Am source was very close to that from ²¹⁰Po (5.5 MeV versus 5.3 MeV, respectively), which comes from the Pb in the C4 solder bumps.

Foil sources were used to validate a Monte Carlo code SEMM-2 [7–9], which was used to simulate SEU components of alpha-particle sources in product packaging solutions, such as solder bumps. The SEMM-2 code accounts for the SEU dependence on alpha-particle energy spectrum, source geometry, and metallization thickness. SEMM-2 was also used to model the alpha-particle emission from the thick ²³²Th source. SEMM-2 uses alpha-particle energy-range tables from SRIM (Stopping and Range of Ions in Matter) [16]. The agreement in the shape of the energy spectrum was nearly identical between the model and experimental data. This is important because it forms the basis for SEU modeling.

Fully packaged chips could not be tested with beams of low-energy alpha-particles because they will not penetrate the packaging materials. A technique developed over the last few years is to “spike” the underfill used for bonding the semiconductor device to a chip carrier with radioactive ²¹⁰Pb or a combination of ²¹⁰Pb and ²¹⁰Po [17]. As described in [17], it is possible to alter the half-life of the source by mixing the proper ratios of ²¹⁰Pb and ²¹⁰Po so that the alpha-particle emission rate remains nearly constant for the length of a typical experiment. The flux from recent formulations of HUF was approximately 1 × 10⁷ alpha-particles/khr-cm². Figure 9 shows a representative energy spectrum of the HUF alpha-particle emission. The endpoint energy was 5.3 MeV, and the spectrum shows the characteristic shape of emission from a thick source (thicker than the alpha-particle range in the material). This technique allowed for accelerated testing to occur on system chips without an external beam.

Figure 9

²¹⁰Po is the alpha-particle-emitting isotope in HUF and solder bumps, so HUF experiments can be correlated to the model bump SEU component. HUF accelerated the solder bump SEU rate through a flux factor and a geometric factor; the alpha-particle flux from HUF was more than five orders of magnitude higher than the alpha-particle flux of low-alpha-particle solder bumps. In addition, HUF filled the area between solder bumps and covered a larger fraction of the surface than the solder bumps; this led to a geometric acceleration factor equal to the ratio of surface area coverage:

In a recent experiment, several IBM POWER6* processor modules were filled with HUF as described above and the SEU rate was recorded over a period of time. Details of the POWER6 processor appear in Reference [18]. The SEU rates for the master and slave stages of flip-flops were recorded for both 0 → 1 and 1 → 0 transitions. The average SEU rate for each data transition was determined for the modules and in most cases was within a factor of 2 of the results of model calculations.

IBM test sites containing flip-flops and memory arrays were routinely tested for SEUs using high-energy protons from the Francis H. Burr Proton Therapy Center (formerly the Northeast Proton Therapy Center) at Massachusetts General Hospital in Boston [19]. Some of the equipment in the external user program at the hospital was originally used for SEU studies at the Harvard Cyclotron. The Cyclotron facility has two patient treatment gantries that rotate around the patient, an eye irradiation station, and a users' facility for SEU and irradiation studies.

The beam delivered from the Cyclotron arrived in the experiment room at 160 MeV. The energy of the beam was lowered and spread out laterally by passing through a series of Pb and/or Lucite** degraders. The length of these cylindrical 75-mm diameter degraders varies. In order to facilitate changing the beam energy rapidly and reduce the exposure of personnel to relatively high levels of residual radiation in the degraders after the beam has passed through them, a rotary mechanism has been developed that allows the degraders to be placed into the beam remotely [20]. A photograph of the rotary degrader is shown in Figure 10.

Figure 10

To check the beam properties (energy, fluence) and the functionality of the tester, the same “golden” part was irradiated for each visit to the Cyclotron facility and the energy dependence of the SEU rate was determined from 20 MeV to 148 MeV. The energy-dependent bit-failure cross-section was compared to that of previous experiments. The methodology used for making SEU measurements complies with the JEDEC (Joint Electron Device Engineering Council) specification [21].

In most experiments, the bit-failure rate at several operating voltages was measured at energies of 20 MeV to 148 MeV. Folding in the energy dependence of the bit-failure cross-section with the terrestrial neutron flux yielded an overall SEU rate for the device [21]. Figure 11 shows measured bit-failure cross-section data for 180-nm, 130-nm, and 90-nm SRAM SOI devices irradiated with 148-MeV protons as a function of operational voltage [10].

Figure 11

Alpha-particles will continue to be dominant SEU sources in modern circuits. As such, it will be increasingly important to have detectors available to measure very low levels of alpha-particle activity in solders, underfill, and molding compounds, as well as during wafer processing. Two detector types are under development to help meet this need.

Testing devices with radioactive foils, external beams of alpha-particles, radioactive underfill, and high-energy protons have been useful to validate our SEU model (developed by the SEMM-2 SEU simulation tool) and to determine the SEU rate and the range of critical charge necessary to flip bits in critical circuits.

We acknowledge valuable contributions from John Aitken, Carl Bohnenkamp, Robert Dennard, Michael Gaynes, AJ KleinOsowski, Nancy LaBianca, Scott McAllister, Chuck Montrose, Conal Murray, Tak Ning, and Henry Tang.

*Trademark, service mark, or registered trademark of International Business Machines Corporation in the United States, other countries, or both.
**Trademark, service mark, or registered trademark of E. I. du Pont de Nemours and Company or Sony Computer Entertainment, Inc., in the United States, other countries, or both.

¹A flux calculator is available at http://www.seutest.com/cgi-bin/FluxCalculator.cgi.
²E. Adams and P. Ronsheim, personal communication.
³An example of the ²⁴¹Am spectrum is shown by the vendor at http://www.ortec-online.com/detectors/chargedparticle/am1.htm.

Received July 3, 2007; accepted for publication November 6, 2007; Published online March 6, 2008.