Neutron Irradiation and Computer Memory: original experiments [longish]



Date view Thread view Subject view Author view

Peter B. Ladkin (ladkin(at)rvs.uni-bielefeld.de)
Mon, 12 Feb 2001 22:07:24 +0100


Folks, Safecrit people building avionics, or power MOSFETs for rail locomotives, might be somewhat interested in my continued probes on SEUs. But, as I have already pointed out, this is an O(10**-8) phenomenon which can be, and is, avoided by fault-tolerance engineering. I have traced back the literature to a 1979 paper that reported irradiation experiments of DRAMs with neutron beams. I'm copying the list because I like to finish what I have started (even though others may wish I wouldn't :-). Many thanks again to Jürgen Engels for his willingness to discuss this phenomenon with me. Unfortunately, I haven't been able to consult a frequently cited paper from 1984, because our library doesn't have the TNS from 1982 through 1986. However, this 1979 paper gives some food for thought. The paper is written by Guenzer, Wolicki and Allas from the (US) Naval Research Lab. Single Event Upset of Dynamic RAMS by Neutrons and Protons, IEEE TNS 26, December 1979, pp5048-5052. They mention that semiconductor manufacturers noticed SEU when they began to manufacture 4K and 16K DRAMs. The devices "were observed to upset spontaneously at intervals on the order of minutes." They cite two papers, also in IEEE TNS in 1979, which "demonstrated that" alpha particles from the spontaneous fission of U-238 and Th-230 were responsible for the upsets. "These radioactive impurities are found in part per million concentrations in the ceramic packaging materials typically used for ICs." Presumably these days are long gone, since the altitude experiments reported by Taber and Normand ten+ years later reported upsets of the frequency of O(10**-9) to O(10**-8) per bit per flight-hour. That's somewhat less than upsets on the order of minutes. So, first the conclusions, and then the commentary on the paper. The authors observed upsets generally at the rate of about O(1) per 10**8 n/sq.cm particle flux. However, the NEC chips exhibited two orders of magnitude more upsets than chips from other manufacturers (that is, if you'll pardon the explicit calculation, O(10**2) upsets per !0**8 n/sq.cm. particle flux), and one order of magnitude more 0-1 flips than 1-0 flips. The authors don't say how many actual upsets were observed, although they repeated one experiment 40 times in order to see how their results fit a Poisson distribution within a 50% confidence interval, and obtained 58 upsets (their data did not fit the 50% confidence interval, although it had roughly what one might call the same shape. This means that it had five points, went up between points 1 and 2, and was monotonically decreasing for points 2-5. The 50% confidence-interval Poisson does that too. This crude similarity seemed to content the authors). There are no (control) data provided on upsets without bombardment, so we don't know what the "background" level is. Their posited mechanism had 14 MeV neutrons engendering alpha particles. Well, they report that they got similar results even using a mean-6.4MeV beam which didn't generate any neutrons above 13MeV. Further, they say that their posited mechanism apparently cannot explain the upset from both states (i.e., 0-1 and 1-0 upsets equally). This paper appears to be the genesis of the hypothesis of neutron-caused single-event upsets. Subsequent papers claim that single-event upsets at altitude within the atmosphere are caused primarily by atmospheric neutrons. I would summarise the contents of the paper as follows. Bombarding 1979-fabricated memory chips with neutrons seems to cause an increase in the number of bit upsets. The mechanism is unknown; hypotheses are not substatiated by the experimental results. Furthermore, one can question the statistical significance of the number of actual events observed. Nevertheless, it seems reasonable to conclude that some effect has been demonstrated. The authors themselves say that the effect is more pronounced with proton bombardment. Jump from that to the claim that the increase in upsets with altitude is due to atmospheric neutron bombardment. One may wonder how the gaps may be filled in. Let's consider the evidence produced by proponents of the atmospheric-neutron hypothesis for altitude SEUs. First, there are only three data points: Altitude 0, altitude 29,000 ft and altitude 65,000 ft. The rate of upsets at altitude 0 (where the neutron flux is supposed to be 0) is about 3% of the measured rate at altitude 29,000 ft, which itself is about half the rate measured at 65,000ft. The neutron flux at 29,000ft is 0.3 times the flux at 65,000ft (all this data from Figure 7 of Normand, Single Event Effects in Avionics, IEEE TNS 43, pp461ff, 1996). That is, at ground level there are SEUs which are not explainable from the neutron flux (which is effectively 0), and the flux at 65,000ft is 3.3 times the flux at 29,000ft, but the SEU rate at 65,000ft is only twice that at 29,000ft. Compare that with the evidence of the original paper which demonstrated SEU with neutron bombardment, in which the authors' posited mechanism, by their own admission, couldn't explain their data in a number of respects. And compare with the complete lack of investigation of other possible mechanisms, such as EM field effects at altitude, or the incidence of other particles. I conclude that the supposition that SEU at altitude is caused primarily by atmospheric neutrons is based on reasoning which has not been explicitly formulated and for whose steps the evidence appears, well, meager at best, and contradictory in parts. That's the punch line. One is left still wondering how the gaps may be filled in. No change is required in my previous notes, except to the assertion that the effects of neutrons on SEU in DRAMs has not been observed. It has been, pretty much, providing one accepts that the NRL investigators weren't totally up the wall. Now to a detailed discussion of their paper. [begin paper discussion] Guenzer et al demonstrated the effects of protons and neutron beams by irradiating DRAM. They claim you need neutrons of 14MeV to obtain the interaction that they posit. They experimented with a computer connected to DRAM, half of which was irradiated. They used the NRL 75 MeV cyclotron with a 35MeV deuteron beam impinging on a Be target. This produces a neutron beam with a mean energy of 14MeV, which is what they wanted, but what they actually get is more like a Poisson distribution with a truncated lower part at 5MeV and half the peak intensity. The right tail goes to 30MeV at 1 unit yield; the peak is 15 units yield; the unit is 10**15 neutrons/sterad/MeV/Couloumb. The exact experimental results weren't included, and neither were the controls, but they have a table giving the fluences required to cause a single upset, for 15 different DRAM types. They measured 0-1 and 1-0 flips, because these correspond to different physical processes. They actually got some wildly different results. 18 of the results are in the range 0.7 units to 3 units, but they also got 5 results in the range 3-6 units, 4 results in the range 9-16 units, and one each at 89.3 and 149, in their total of 30 measurement/calculations. The big ones are not opposite big ones: the half-chip measuring 149 0-1s measured 10.4 1-0s. The half-chip measuring 89.3 0-1s measured 14.9 1-0s. Only two other chips were outside of a factor of three between 0-1 and 1-0 flips. Units are 10**8 particles/sq.cm. The summary, "almost all of the experimentally observed upset levels are" O(1) units, seems somewhat facile. 9 results out of 40 were not within this range. Curious observers would have wanted to know how and why NEC produces chips that apparently are two orders of magnitude more susceptible to SEU than Intel, Motorola, Texas Instruments and Hitachi chips. The experimenters couldn't be sure that the effects weren't due to higher-energy neutrons than those at the peak, so they also did some runs, with the four Motorola chips, with deuterons impacting a tritium target, which apparently produces nearly monoenergetic neutrons at 14MeV. These are, sort of, comparable (within about a factor of two) with the equivalent results from the deuteron-Be experiments. Don't ask me why they didn't just do everything with deuterons on tritium if they are looking for effects at a single energy level. It seems to me significant that both NEC chips produced two orders of magnitude more upsets than anyone else's. One Intel chip produced one order of magnitude more; otherwise, they are all roughly the same O(1), units being 10**8 particles/sq.cm. Based on their supposed explanation, they predicted that one alpha particle would be produced for every 2 x 10**6/sq.cm. incident neutrons or protons. If this would be correct, they got a 2% upset rate per alpha particle, and note that this is in "excellent, perhaps fortuitous, agreement" with experimental results from one of the previously cited papers that gave the alpha particles produced by the U-238/Th-230 interaction as the (main) cause of the SEUs. So apparently there is a one in fifty chance that an alpha particle engendered within the chip will cause a bit-flip. This seems to me just extraordinarily large an effect, but then again I don't know anything about particle physics. The researchers also say that the rates of bit-flips from 0-1 and 1-0 "are comparable", but that doesn't seem to be the case for the NEC chips, which record 149 against 10.4, a fourteen-fold difference, and 89.3 against 14.9, a six-fold difference. In discussing flip preferences, they focused on the Hitachi chip, for which they said "different mechanisms seem to govern upsets from the different states, since the energy dependence is so different." The Hitachi showed a four-fold preference for 0-1 against 1-0 upsets. They don't actually say what the ratio of the total 0 settings to the total 1 settings were, so one cannot infer anything about these different figures. One can only infer that NEC chips upset overall at two orders of magnitude the rate of other chips. In trying to find out the actual number of upsets observed, I had problems. There is one sentence in which they say "the mean number of upsets, based on 419 upsets, was (mu) = 1.82." There is a histogram, which showed, for 40 experiments, that 7 observed 0 upsets, 17 observed 1 upset, 8 observed 2 upsets, 7 observed 3 upsets, and 1 observed 4 upsets. They match the histogram against what is predictable from a 50% confidence interval based on a Poisson distribution. That is, from 58 upsets in 40 experiments. This particular histogram came from "an attempt to verify the random and statistical nature of the phenomenon". This seems to be rather meager data from which to present upset rates to two significant figures, in particular since this was the only part of the experiment repeated this number of times (in "an attempt....") Moving on to potential differences between 0-1 and 1-0 flips, they say "The previous explanation of single particle upset of dynamic RAMs is insufficient to explain the upset from both states." They provide a discussion of the differences between a 0-1 flip and a 1-0 flip. But the sentence appears to be unambiguous. Single particle upset cannot explain their data. Not only that, but their data appears to be relatively insensitive to neutron energies: "The 6.5 MeV average neutron energy spectrum has no neutrons above 13 MeV yet for most ships the upset fluence remains in the 10**8 n/sq.cm. range." In other words, they get similar results even when their posited explanation is invalid. Moving on, they say "The dynamic RAMs tested were even more sensitive to upset from protons than from neutrons." Their final sentence is: "Single particle upset, whether from residual radioactive activity [sic] or from cosmic rays, protons or neutrons in a hostile environment, is a problem that needs to be met in coming generations of improved performance devices." That conclusion is somewhat non-commital as regards the source of the problem. [end paper discussion] PBLContent-Type: text/plain X-Original-Content-Type: application/x-pkcs7-signature; name="smime.p7s" [The content of this part has been removed by the mailing list software]


Date view Thread view Subject view Author view