A GEANT-Based Model for Single Event Upsets in ... - IEEE Xplore

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 53, NO. 4, AUGUST 2006

2353

A GEANT-Based Model for Single Event Upsets in SRAM FPGAs for Use in On-Detector Electronics Steve Skutnik and John Lajoie

Abstract—A model is developed to calculate expected Single Event Upset rates in Xilinx’s line of radiation-hardened field programmable gate array (FPGA) offerings for the radiation environment at the PHENIX experiment on the Relativistic Heavy Ion Collider at Brookhaven National Laboratory. The results of this model are compared to an experiment carried out at PHENIX, where actual upset data was obtained. Specific attention is given to unique features of the model, including major sources of “secondary” radiation flux. Index Terms—Field programmable gate arrays (FPGA), radiation effects, single event effects, single event upset modeling.

I. INTRODUCTION

G

IVEN the significant advantages of reprogrammable logic devices in terms of cost and design flexibility, considerable interest has arisen in the nuclear physics community regarding their feasibility for implementation in on-detector electronics. However, given the fact that such devices carry the drawback of being susceptible to radiation-induced bit upsets and other failures, study is warranted as to their expected failure rate when placed in the radiation-hostile environments of particle physics experiments. The study of soft errors in reprogammable logic has been a subject of considerable research (cf. [1], [2]), particularly given the radiation-hostile environments of aerospace applications. However, unlike traditional aerospace applications where the radiation flux is well-understood and from a relatively uniform source (i.e., atmospheric neutrons, protons and heavy ions in space), the radiation environment created from heavy ion collisions (such as the Au Au collisions at RHIC) is far less homogeneous. Rather, it is marked by large fluxes of many different particle types, both from well-characterized upset sources (such as protons and neutrons) and far less well-documented types, such as pions and kaons. As we shall demonstrate in this paper, a simple GEANTbased model in which the upset cross sections are calculated in accordance with known proton and neutron differential SEU cross sections (cf. [3], [4]) proves highly successful at providing a first-order estimate of expected upset rates. This may ultimately serve as a “proof of principle” for future experiments to evaluate the applicability of rad-hard logic solutions for on-detector electronics. Manuscript received March 11, 2006; revised May 20, 2006. The authors are with the Department of Physics and Astronomy, Iowa State University, Ames, IA 50011-3160 USA. (e-mail: [email protected]; [email protected]). Digital Object Identifier 10.1109/TNS.2006.878282

II. EXPERIMENTAL SETUP A. Device Under Test For this test, a prototype version of a Xilinx radiation-hardened FPGA (model XQR2V6000) was used with a Virtex II 1152 proto board, encased in an Aluminum enclosure. A standard JTAG interface was connected to a 100-foot parallel cable, which allowed for the testing equipment to remain outside the interaction region (IR) in the experimental assembly hall. To ensure proper device communication, the individual JTAG headers were soldered directly to the corresponding pins upon the proto board. The self-contained box was attached to a frame for a separate prototype GEM detector in the IR March 3, 2004 at the “far” position of approximately 250 cm below the beam collision point and raised up into the near position (approximately 53 cm below the collision point) from March 17th to March 19th, after which it was lowered back into its original far position. Data collection began on March 5th when development on the data acquisition software stabilized and ended on March 24th with the end of 200 GeV running at RHIC [5]. B. Data Acquisition and Monitoring To check for SEUs, Xilinx provided us with their Fault Inject and Verification Tool (FIVIT) software [6]. This software is capable of checking a programmable logic device for SEUs by comparing the current logic configuration with a mask file containing the chip configuration, in essence checking each individual configuration bit in the downloaded design to ensure the correct state for each. The FIVIT software is also capable of determining whether an upset occurs in the block RAM (BRAM) or the configurable logic blocks (CLB). the configuration logic of the chip. It should be emphasized that both of these are measured in terms of the used blocks only, which is not the same as the actual configuration and block RAM capacity of the chip. Throughout the experiment, only errors in the configurable logic blocks were observed, namely due to the fact that little design logic was stored in the block RAM sections. Every 2 minutes, the integrated luminosity (the total number of beam collision interactions, given by (1)) was calculated by polling the current beam interaction rate via a Perl script and multiplying by the time elapsed between measurements.

(1) The beam interaction rate is measured in collision events per unit time, which is then integrated over the

0018-9499/$20.00 © 2006 IEEE

2354

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 53, NO. 4, AUGUST 2006

TABLE I CHRONOLOGICAL SUMMARY OF SEUS (r = 53 CM) WITH 16394483 CLB BITS AND 2414592 BRAM BITS

TABLE II CHRONOLOGICAL SUMMARY OF SEUS (r = 250 CM) WITH 16394483 CLB BITS AND 2414592 BRAM BITS

elapsed time to give a total number of interaction events ( ). The interaction rate is calculated from the beam luminosity ( ), scaled by the interaction cross section for Au Au collisions . Thus the units of integrated luminosity are simply “events”. Every 20 minutes, the chip configuration was checked by FIVIT, with GUI operations being automated through an AutoIt automation script [1]. A Perl script then extracted results from the FIVIT log into a useful format, which was uploaded to the web for online monitoring. Confirmed upsets were distinguished from occasional, spurious JTAG communication errors by verifying that the errors were consistent upon a second and third readback. In addition, SEUs occurred in only a few bits at a time, whereas problems in both the power supply and JTAG communication resulted in errors which were several orders of magnitude greater than “genuine” SEUs. Upsets were correlated to the closest measured integrated luminosity count in order to give a proper account of the time interval that could be expected between upsets. After a confirmed SEU occurrence, the chip was reconfigured via FIVIT and the integrated luminosity count in the log reset.

form, i.e., in order of occurrence. After each upset, the integrated luminosity count was reset. It should also be noted that the luminosity values listed in Tables I and II are for the observed upsets, and do not account for the total time in the active detector environment (measured in integrated luminosity). This marks the difference between the calculation of “observed mean” time between upsets and the “aggregate mean,” calculated as the total number of upsets observed within the entire exposure period.

C. Dosimetry Measurements Over the course of the experiment, two dosimetry badges were used to measure the ionizing radiation dose the chip was exposed to over the course of running. The first badge was attached to the board while the setup remained in the far position and was removed prior to the board being elevated to the near position. The second badge was attached to the board when it was raised into the near position, however it was not able to be removed from the board prior to the end of the run due to the lack of access to the beam. Nonetheless, the dose from the board’s exposure in the near position to the beam can be corrected analytically using the dosage measured from the first badge. This is due to the fact that the predicted dose close to the beam was an order of magnitude greater than the far position, and the luminosity of the following 64 GeV run was order of magnitude lower than that of the 200 GeV run. Given that the integrated luminosity of the “near” dose was known, the dose from the badge’s exposure in the far position can be subtracted off from the estimated dose rate of the first badge, thus giving an estimated dose in the “near” position. D. Experimental Results Tables I and II give a summary of SEUs observed. Note that the units are measured in terms of integrated luminosity (i.e., “interaction events”). Upsets are presented in chronological

III. OVERVIEW OF MODEL A. Model Interface The PISA software package [8] is a custom interface developed by the PHENIX collaboration [9] to the popular GEANT particle simulation software [10], a package generally regarded as a standard for purposes of detector simulations. The GEANT3-based PISA interface was designed specifically for use with the PHENIX collaboration to simulate detector behavior, having several standard PHENIX detector components designed for simulation in the GEANT tracking engine. Users can choose what detectors to “activate” in the IR and select among physical effects to model—for example Cherenkov radiation, Bremsstrahlung, etc., as well as specifying “sensitivity” of detectors to particular particle types, such as neutral hadrons. Tracking threshold energies can also be modified in the various detector media to balance simulation performance with resource consumption, ultimately providing a highly flexible interface which can be used to “fine-tune” detector responses. For the specific purposes of this model, the tracking threshold energies for certain materials such as the iron magnet cores were lowered from their default values of 500 MeV to 50 keV. This was done in order to properly account for the radiation dose generated by low-energy “secondary” neutrons created in interactions with the magnet media, which account for a significant portion of the calculated SEU cross section. Users can select from several event-generation interfaces, from single particle events generated by the user to more sophisticated packages such as HIJING, the event generator used for purposes of this simulation. The HIJING package allows for a completely customizable simulation of nuclear collisions; users can control the atomic mass (A), the atomic number (Z), the RMS values of the collision vertex, and the laboratory momentum, as well as the collision centrality. For purposes of this simulation, the HIJING event generator was configured as to simulate the experimental conditions as closely as possible, with an RMS value of the z-vertex set at 22.0 cm, laboratory GeV, with minimum-bias collisions of momentum Au Au nuclei.

SKUTNIK AND LAJOIE: A GEANT-BASED MODEL FOR SINGLE EVENT UPSETS IN SRAM FPGAS

B. Estimation of a Weibull Curve for Neutrons Based Upon Integrated Cross-Section Data 1) Introduction and Motivation: In this section, we shall present an estimated differential SEU cross-section curve for neutrons. The motivation for this comes from the fact that SEU cross sections for neutrons as measured by Xilinx at the Los Alamos National Laboratory’s Atmospheric Neutron Science Center (LANSCE) test facility [11] in addition to independent testing by iRoC [12] estimate the SEU cross section from neutrons in the following form:

2355

TABLE III OBSERVED AND CALCULATED SEU CROSS-SECTION VALUES FOR A VIRTEX II-FAMILY DEVICE FROM A MINIMIZED WEIBULL-FORM FIT TO XILINX AND IROC DATA, AS COMPARED TO MEASURED VALUES, WITH 2 DEGREES OF ~ ! 0. CROSS SECTIONS IN [cm ] FREEDOM AND 

(2) where the fluence is taken as the integral of the flux distribution as a function of energy, and the fluence rate is taken to be a constant as a function of time. While such a form provides a rough means of estimating the SEU rate due to conditions such as atmospheric neutron flux (of particular interest for aerospace applications), it does not provide an actual differential neutron cross section as a function of energy. Rather, while such a “cross section” carries the appropriate units, it is in fact a convolution of an assumed spectrum (taken to simulate the Hess spectrum of atmospheric neutrons) and the neutron SEU cross section of the device - hence it is not a “true” cross section. What is therefore needed is a differential cross-section profile for the device as a function of energy. Such a function would be independent of the assumed neutron energy spectrum and thus able to used with any neutron energy distribution. The goal of this section will thus be to present a means of estimating a differential SEU cross-section distribution for a Xilinx Virtex-II family device. Chief among the assumptions in this shall be that the SEU cross-section distribution for neutrons will be of the form of a Weibull curve, the functional form as that found for protons. This assumption shall be tested along with alternative functional forms. 2) Estimation of the Differential Cross Section: In independent tests at LANSCE, both Xilinx and iRoC tested the upset rate of a Virtex-II family FPGA to evaluate the susceptibility of FPGAs to upsets due to atmospheric neutrons. In addition to the LANSCE data, iRoC also made a measurement of the SEU sensitivity of a Virtex-II device based upon a mono-energetic beam of 14 MeV neutrons. This data was used in addition to the other data points to test the calculated Weibull curve at a specific energy point. To obtain the desired Weibull fit, the general Weibull form was taken in ROOT’s1 TMinuit minimization class [13], minimized by a standard chi-squared test. Thus, by allowing the Weibull parameters to freely float and minimizing the chi-square MeV and MeV bins and the value for the integrated 14 MeV bin, the resulting equation was found to be:

(3) 1ROOT is an object-oriented data analysis framework commonly used in high-energy and nuclear physics. For more information, see [5].

Fig. 1. Minimized Weibull SEU cross-section distribution fit (dotted), compared to a linear fit (dot-dashed), plotted with 14 MeV neutron cross section as measured by iRoC [12]. Cross section is given in [cm 1 bit ].

where the cross section is measured in units of and E in MeV. 3) Comparison of the Proposed Fit to Observed Neutron Data: Table III gives a comparison between the predicted and measured values for an Xilinx XQR2V6000 FPGA, with the corresponding errors. One will notice the good adherence to all of the data points provided by the minimized Weibull curve. While the number of data points used for the fit is limited (and hence the actual values of the parameters would likely depend upon more data), the Weibull form appears to be capable of accurately fitting the available data. To test the uniqueness of such a fit, other fit forms were attempted, including linear, polynomial, and logarithmic fits. Only a linear fit provided any meaningful adherence to the provided data points, however such a fit is of significantly poorer quality than the Weibull form, in particular at matching the lower-energy regime of the neutron SEU spectrum. This would indicate that aspects to the Weibull form such as the “saturation value” appear to be an essential physical component of the cross-section distribution. Fig. 1 gives a graphical comparison of the minimized Weibull fit to the available SEU data, plotted with the SEU cross section from iRoC’s test with a mono-energetic beam of 14 MeV neutrons [12]. C. Calculation of SEU Cross-Section Contribution for Other Particle Species While the cross section of protons is known as a function of incident particle energy [4], this is not true for other more

2356

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 53, NO. 4, AUGUST 2006

Fig. 2. Comparison of Weibull distribution for protons from Xilinx [4] (solid), as compared to the estimated cross section for anti-protons (dashed), estimated from the nuclear interaction cross-section ratio to protons.

“exotic” particle types which are prevalent in collision products, including pions and kaons. While some studies (particularly [14]) have made attempts to characterize the differential SEU cross section of pions, no pion cross-section data existed at the time of the experiment for the device under test. However, unlike the experiment conducted by Gelderloos et al. [14], we did not have access to a pion beam source for the purposes of this experiment. Thus, a method of calculating the differential pion cross section will be proposed in this section. The assumption was made that the SEU cross sections of such particles directly scaled to the interaction cross section of such particles with protons at the same kinetic energy [7]. Thus, for each type of particle, a scale factor was determined by taking the ratio of the measured interaction cross section over the proton–proton interaction cross section at the same kinetic energy. This in turn accounts for both the lower mean interaction cross section as well as the several resonance energies for particles such as pions and kaons that a fixed scale factor could not account for. This process was carried out for all particle types with a significant proton interaction cross section, including anti-protons, , and , shown in Figs. 2–4. For anti-neutrons, the special case of anti-neutrons, it was assumed that the antineutron cross section scaled to the neutron cross section approximately as the same ratio as anti-protons to protons [16]. An example of these “scaling factors” (in this case, for pions) is shown as a function of particle energy relative to the proton Weibull distribution in Fig. 3. It should be noted that the apparent “discontinuities” in Figs. 2–4 are artifacts due to particular resonance energies of the particle cross sections. These artifacts are thus “folded” into the calculated cross section, resulting in the apparent “jumps.” IV. ANALYSIS OF RESULTS A. Comparison of Predicted Doses to Measured TLD Values In the following sections, we shall present the human-equivalent rem doses as calculated in the prior sections to the values as

Fig. 3. Comparison of Weibull distribution for protons from Xilinx [4] (solid), as compared to the estimated cross section for  (dotted) and  (dashed), estimated from the nuclear interaction cross-section ratio to protons.

Fig. 4. Comparison of Weibull distribution for protons from Xilinx [4] (solid), as compared to the estimated cross section for K (dotted) and K (dashed), estimated from the nuclear interaction cross-section ratio to protons.

measured by the TLD badge attached to the test box in the IR, comparing these to the values predicted by the PISA model as a objective measure of the model’s accuracy at simulating particle fluences. 1) Dose as Measured by TLD Badge: The badge results returned by BNL personnel monitoring were in human-equivalent rem dose. Five categories of doses were measured, corresponding to the three human tissue depths (“shallow”, corresponding to an epidermal dose; “lens”, corresponding to the lens of the eye; and “deep”, corresponding to a deep-tissue dose), and two neutron doses, one from LiF chip upon the badge (measuring “slow” neutrons with a kinetic energy less than 0.5 MeV) and a CR39 film for “fast” neutrons (of energy greater than 0.5 MeV) [17]. To evaluate these doses in terms of energy absorbed rather than the human-equivalent dose (i.e., radiation absorbed dose or rad), said measurements need to be re-scaled by the quality factors used in originally calculating the rem doses. Given the fact that the conversion proves to be inexact at best, particularly for the energy-dependent neutron quality factors, the calculated doses in the PISA model were made in the human-equivalent

SKUTNIK AND LAJOIE: A GEANT-BASED MODEL FOR SINGLE EVENT UPSETS IN SRAM FPGAS

TABLE IV COMPARISON OF MEASURED RADIATION DOSE BY TLD BADGES TO DOSE CALCULATED BY PISA FOR r = 53 CM, GIVEN IN [(rem)=(event)]

TABLE V COMPARISON OF MEASURED RADIATION DOSE BY TLD BADGES TO DOSE CALCULATED BY PISA FOR r = 250 CM, GIVEN IN [(rem)=(event)]

rem dose. Hence, the quality of the PISA model shall be evaluated by comparing the measured dose from the TLD badges, converted into a per-event rem dose to a per-event human dose rate as calculated by PISA (Table IV). 2) Radiation Doses as Predicted by PISA: Tables IV and V give a comparison of calculated dosages to measured dosages, in rem/event (an “event” being defined as one integrated luminosity count), broken down by radiation source. As previously noted, the doses calculated by PISA are in human-equivalent rem form, rescaled to a per-event dose rate. 3) Comparison of Radiation Doses as an Evaluation of the Pisa Model: Clearly evident is the reasonable agreement between the PISA model’s dosage calculations and the measured dosages for the far ( cm) position. With the exception of % of the measlow neutrons, the predicted dose is within sured dose. Meanwhile, the predicted slow neutron dose was still within an order of magnitude of the measured dose. Doses for the “near” position prove to provide a more interesting challenge, given the fact that the dose is at best an estimate. This is due to the fact that the second badge was also exposed to a measured amount of integrated luminosity in the far position at 200 GeV and then an unspecified integrated luminosity at 64 GeV. Hence the larger degree of underestimation for the near dose may largely be due to the uncertainty in the exact “dose” received while in this position. The case of slow neutrons presents a particularly difficult case for simulations. To properly estimate the dose of low-energy neutrons requires setting the tracking thresholds of detector components (such as the magnet media) down significantly, with the very real expense of computation time and stability to consider. While as a compromise the tracking threshold was set down to 50 keV, at least some further accuracy in representing the low-energy neutron spectrum may be recovered by further lowering this threshold, as well as that of other critical detector components. Nonetheless, the peak energy of neutrons created in the detector environment (in particular through interactions with the magnet media) is several orders of magnitude higher than

2357

TABLE VI SEU STATISTICS FOR OBSERVED EVENTS AND THEIR PREDICTED VALUES FROM PISA (INTEGRATED LUMINOSITY IN COUNTS)

the thermal neutron flux, This may indicate that the accuracy thermal neutron spectrum produced by HIJING/PISA may be limited no matter what. Decreasing these tracking thresholds thus ultimately proves to be a matter of diminishing returns. While lowering the tracking thresholds may deliver a somewhat more accurate low-energy neutron dose, this comes at the expense of exponentially increasing the simulation time by increasing processing and memory requirements and in turn decreasing the stability of simulations. Despite the failure of such a model to completely account for the thermal neutron doses, the model is overall quite successful at approximating doses from gammas and high-energy neutrons to first order, giving a sound model for calculating the approximate absorbed dose for rad-hard FPGA devices and hence useful lifetime estimates. Likewise, it demonstrates that the estimated SEU rates, which shall be presented in the following section, are based upon a sound physical model. While the residual inaccuracy in the thermal neutron doses is unfortunate, it should be noted that the peak in the energy spectrum for neutrons appears to occur at much higher energies. Thus, while the thermal neutron dose is likely underestimated, the thermal neutron flux appears to be orders of magnitude lower than that of higher-energy neutrons. From this, one may infer that the corresponding cross-section contribution is consequently lower as well. B. Comparison of Observed SEU Rates to Predicted Rates Table VI gives a summary of both observed SEUs from experiment along with the observed mean (defined as the average integrated luminosity observed per SEU event), aggregated mean (defined as the total integrated luminosity divided by the number of SEU events), and the standard deviation of the observed mean, compared to the predicted SEU rates in PISA. For purposes of statistics, the observed upset distribution was taken to follow a Poisson distribution. Given this form, the mean value of upsets was taken as the aggregate observed mean (i.e., the number of upsets over the total period of observation). Hence the sigma of such a distribution is taken as the square root of the mean , i.e.,:

(4) Thus for these purposes, the bounds define a 68% confidence interval, i.e., given the mean value of upsets , it is 68% likely that any new measurement would fall within the bounds . (Likewise, it is 95% likely that a value would fall within of .)

2358

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 53, NO. 4, AUGUST 2006

TABLE VII STATISTICAL EVALUATION OF UPSETS IN THE EXPERIMENT AND MODEL, ASSUMING A POISSON DISTRIBUTION, FOR EACH ORIENTATION

TABLE VIII FRACTIONAL SEU CROSS-SECTION CONTRIBUTION BY PARTICLE TYPE FOR THE NEAR AND FAR POSITIONS IN PISA

As a measure of how well the observed data then compares to the model, the probability of the observed number of upsets occurring in bounds of the predictions is given from the Poisson distribution was calculated, where such a probability is given as: (5) For reference, the values of in addition to the predicted number of upsets over the observed interval, along with the probability of the observed number of upsets occurring based upon the model, are given in Table VII. From Table VII, it should be noted that the experimentally observed number of upsets fall well within the sigma bounds of the PISA model, assuming a Poisson distribution form. Likewise, the probability of exactly three upsets occurring within the experimentally observed time period, based upon the predicted mean values of upsets, is approximately 21% and 22%, respectively for the near and far positions, indicating a good agreement between the PISA model and experiment. Based on these results, the PISA model can be used to make reasonably accurate predictions to the SEU rate to first order, so long as the tracking thresholds are modified within the detector and magnet media in the IR such that an accurate neutron flux can be estimated. A caveat should be made for both cases however that the observed SEU rates were based off of a very small data set, due to the limitations of the experimental setup (as to minimize interference with other experiments in the detector) and the relative infrequency of upsets. Hence the “observed” SEU rates should be interpreted accordingly: with few data points, the trend is still subject to a high degree of statistical error. Thus deviations from this “observed” rate are based upon an estimate—what is useful in this case is their agreement to within an order of magnitude. C. Relative SEU Cross-Section Contribution for Various Particle Types in Pisa Model Table VIII gives the relative cross-section contribution for various particle types in the PISA model for each position. One will observe that while protons contribute the smallest fraction of the cross section in the near position, these become the chief source of SEUs in the far position. This can easily be explained by the fact that the decays of kaons and pions

Fig. 5. Origin coordinates of neutrons incident upon the near test annular region, centered about r = 53 cm, z = 0 cm.

significantly lowers the flux from such particles in the far position. Likewise, one will notice that pions alone vastly contribute to the SEU rate in the near position and still present a significant fraction of the SEU cross-section contribution in the far position. This proves remarkably different than the concerns of traditional space and terrestrial applications, which deal primarily with faults induced by heavy ion fragments and atmospheric neutrons induced by cosmic ray showers, respectively—both of which are well-understood. Thus the introduction of actual pion data for the device under test, similar to [14], would greatly enhance the predictive power of the model. In particular, the proposed scaling relationship may fail to account for resonances in the SEU cross section for pions near 150 and 400 MeV [14]. These resonance peaks can result in SEU cross sections several times that of protons at equivalent energy. This appears to be consistent with the relative quality of predictions between the near and far positions. In particular, the far position predictions appear to be grow more accurate with decreasing dependence upon an unknown factor (i.e., cross sections for pions and kaons). Likewise, the use of pion-Silicon interaction data (as opposed to the simplified model of pion-proton interactions) for the SEU scaling relationship would likely provide a more accurate model for upsets in devices with a known proton/neutron upset crosssection profile. D. Radiation Contribution From Secondary Sources One of the key features of the proposed model is its ability to map out secondary sources of radiation (e.g., secondary particles generated through interactions with the detector media). The most critical source of secondary radiation in this particular experiment showed to be neutrons, which were predominantly generated within the iron magnet cores. Figs. 5 and 6 show the distribution of neutrons incident upon the test annuli in the PISA simulation. One will notice that the neutrons are predominantly created in the magnet media (locm). cated around

SKUTNIK AND LAJOIE: A GEANT-BASED MODEL FOR SINGLE EVENT UPSETS IN SRAM FPGAS

Fig. 6. Origin coordinates of neutrons incident upon the far annular test region, cm, z cm. centered about r

= 250

=0

Fig. 7. Momentum distribution of neutrons originating from the magnet media cm. incident upon the near annular test region, r

= 53

2359

Fig. 8. Momentum distribution of neutrons originating from the magnet media incident upon the far annular test region, r cm.

= 250

Fig. 9. Mean SEUs per bit, per device as a function of fill time for Q-Pro Virtex II device with 16394483 CLB bits in a RHIC I Au Au fill with initial interaction rate of 14 kHz, near position (r cm). Solid: Mean (aggregate), Dashed:  , Filled: Predicted (with systematic errors).

= 53

+

Figs. 7 and 8 show the momentum spectrum of neutrons incident upon the annular region created in the magnet media area. In particular, one will notice an exceptionally strong interaction peak around 50 MeV. It should be noted that to fully “observe” the neutron interaction peak required setting the material tracking threshold for the magnet iron to 50 keV, down from its default value of 500 MeV. This comes with the consequence of significantly increasing computation time in the simulation, however it proves necessary in order to fully capture the secondary neutron spectrum (and hence gain an accurate cross-section contribution). V. CONCLUSION A. Estimating a Mean SEU Rate Per Bit as a Function of Fill Time In order to characterize an average failure rate in a device due to neutron interactions in a useful fashion, the observed upset rate in terms of integrated luminosity was folded into a distribution of the number of events per Au Au fill2 as a function 2A

“fill” is defined as a single running session when the beam is ramped up to full energy (200 GeV) with ions for collisions—in this case, Au Au.

+

Fig. 10. Mean SEUs per bit, per device as a function of fill time for Q-Pro Virtex II device with 16394483 CLB bits in a RHIC I Au Au fill with initial cm). Solid: Mean (aggregate) interaction rate of 14 KHz, far position (r Dashed:  ; Filled: Predicted (with systematic errors).

= 250

+

of time. From this, one can obtain a mean number of SEUs per fill Figs. 9 and 10 as a function of time for a standard RHIC fill. give a distribution of the mean SEUs per bit for the device under

2360

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 53, NO. 4, AUGUST 2006

test in each of the positions tested for a standard RHIC fill. One will observe that the predicted SEU rate (solid) with systematic errors (filled) fall within the first bounds (dashed) for both the near and far positions, indicating a good agreement between the model’s predictions and the experimental observations.

REFERENCES [1] M. Ceschia, M. Violante, M. Sonza Reorda, A. Paccagnella, P. Bernardi, M. Rebaudengo, D. Bortolato, M. Bellato, P. Zambolin, and A. Candelori, “Identification and classification of single-event upsets in the configuration memory of SRAM-based FPGAs,” IEEE Trans. Nucl. Sci., vol. 50, no. 6, pp. 2088–2094, Dec. 2003. [2] R. Katz, K. LaBel, J. J. Wang, B. Cronquist, K. Koga, S. Penzin, and G. Swift, “Radiation effects on current field programmable technologies,” IEEE Trans. Nucl. Sci., vol. 44, no. 6, pp. 1945–1956, Dec. 1997. [3] M. Alderighi, A. Candelori, F. Casini, S. D’Angelo, M. Mancini, A. Paccagnella, S. Pastore, and G. R. Sechi, “SEU sensitivity of Virtex configuration logic,” IEEE Trans. Nucl. Sci., vol. 52, no. 6, pp. 2462–2467, Dec. 2005. [4] S. E. E. Consortium, Xilinx Single Event Effects Consortium Report: Virtex II Static SEU Characterization Transl.:Single Event Effects Consortium. Tech. Rep., 2003. [5] S. E. Skutnik, “A Scalable Analytic Model for Single Event Upsets in Radiation-Hardened Field Programmable Gate Arrays in the PHENIX Interaction Region,” Master’s Thesis, Iowa State Univ., Ames, IA, Mar. 2005 [Online]. Available: http://shepody.physics.iastate.edu/skutnik/thesis.pdf [6] M. Alderighi, F. Casini, S. D’Angelo, M. Mancini, A. Marmo, S. Pa-store, and G. R. Sechi, “A tool for injecting SEU-like faults into the configuration control mechanism of Xilinx Virtex FPGAs,” in Proc. 18th IEEE Int. Symp. Defect and Fault Tolerance in VLSI Systems, Nov. 2003, pp. 71–78. [7] HiddenSoft. Autoit scripting software, [Online]. Available: http://www. autoitscript.com/ [8] PHENIX Integrated Simulation Application (PISA) Transl.:PHENIX collaboration, Brookhaven National Laboratory, [Online]. Available: http://vpacl7.phy.vanderbilt.edu/simulation/PISA/pisa.html [9] PHENIX Collaboration, [Online]. Available: http://phenix.bnl.gov [10] GEANT detector description and simulation tool Transl.:CERN, [Online]. Available: http://wwwasd.web.cern.ch/wwwasd/geant/ [11] J. J. Fabula and A. Lesea, “The NSEU response of static latch based FPGAs,” in Proc. Military and Aerospace Programmable Logic Devices Int. Conf. (MAPLD) Transl.:, Washington, DC., Sep. 2003, paper C5. [12] M. Olmos, Radiation Results of the SER Test of Actel, Xilinx and Altera FPGA Instances iRoC Technologies, Tech. Rep., 2003 [Online]. Available: http://www.actel.com/documents/RadResultsIROCreport.pdf [13] R. Brun, Root Analysis Package [Online]. Available: http://root.cern.ch [14] C. Gelderloos, R. Peterson, M. Nelson, and J. Ziegler, “Pion-induced soft upsets in 16 mbit DRAM chips,” IEEE Trans. Nucl. Sci., vol. 44, no. 6, pp. 2237–2242, Dec. 1997. [15] S. Eidelman, “Review of particle physics,” Phys. Lett. B, vol. 592, pp. 1+–, 2004 [Online]. Available: http://pdg.lbl.gov [16] K. Johansson, P. Dyreklev, B. Granbom, N. Olsson, J. Blomgren, and P. Renberg, “Energy-resolved neutron SEU measurements from 22 to 160 MeV,” IEEE Trans. Nucl. Sci., vol. 45, no. 6, pp. 2519–2526, Dec 1998. [17] R. Reciniello, Oct. 2004, private communication.

B. Summary In this paper, we have proposed a means of estimating the average SEU rate for an SRAM-based FPGA device with a known SEU cross section through a GEANT-based interface. In particular, we have presented a means of estimating the SEU cross section for neutrons, as well as more exotic particle species such as pions and kaons. Finally, we have presented an evaluation of major sources of radiation from secondary particle sources, such as neutrons created by the detector media, and the adjustments to such a model required to account for such. Upon comparison with the experimental results, it was found that while the model was highly successful at predicting accurate upset rates to first order in both observed positions, its predictions appeared to be of higher quality farther from the interaction region. This appears to correlate with better knowledge of the dominant particle cross sections in this region. One can thus conclude that such a model could conceivably be applied in similar detector environments in order to calculate the expected failure rate in a given SRAM-based device.

ACKNOWLEDGMENT The authors would like to thank Xilinx, who provided both the XQR2V6000 rad-hard FPGA prototype as well as the FIVIT software used for SEU monitoring, along with the accompanying training. Particularly, they would like to thank C. Carmichael and B. Bridgford of Xilinx; much of the success of this experiment was due directly to their assistance. Likewise, the authors wish to acknowledge the support of the BNL Collider-Accelerator Division and the Department of Energy for making all of this possible.