Tools / Single events / NAND bit flips / Mechanism in 3D

What does one particle do to flip a bit in NAND flash?

It takes charge away. A programmed cell holds electrons on a floating gate, or in a nitride layer, that raise its threshold voltage above a read reference. For the nanoseconds an ion track lives, the dielectrics around that charge conduct along the track, and part of the charge leaves. The threshold falls; if it falls below a reference, the cell reads wrong, and in a multi-level cell the Gray code decides which page takes the error. The view below walks through the strike, the conductive path, the loss, the read and the accumulation over fluence and dose for a planar floating-gate string and a 3D charge-trap string, by three routes to the same flip: one heavy ion, many light ions whose losses add, and dose from electrons, protons and photons that shifts every cell a little. The numbers are those of the JPL charge-loss model as fitted to the Samsung 8 Gb SLC part.

3Dfloating gate and 3D charge trapSLC, MLC, TLCone heavy ion, many light ions, or doseJPL charge-loss model, Samsung 8 Gb SLC fitruns in your browserNAND bit-flip mission estimatorEstimator helpDRAM stuck bits in 3DMemory EDAC tool

Drag to turn, scroll or pinch to zoom. Step through the six stages with the buttons at the bottom left; the panel at the right chooses the cell type, the number of levels, the programmed level, the route to the flip (one heavy ion, many light ions, or dose), the particle and its LET, where the track passes, the fluence and the dose since the last program, and the parameter set, and reads out the shift, the probabilities and the expected errors.About 0.7 MB, no external libraries.

The six stages of the view

1

Programmed, or erased

Electrons tunneled onto the floating gate, or into the nitride, hold the threshold above the read reference. SLC uses two levels and one reference; MLC four levels and three references; TLC eight and seven, each level with a smaller margin. The erased cell has given up its electrons and many more: NAND cells are erased well past neutrality by design, so its gate is net positive, drawn in orange, with the oxide field reversed, and it reads all ones.

2

Strike, or exposure

A heavy ion crosses the stack in picoseconds and leaves a column of electron-hole pairs. A proton leaves too little ionization to matter directly; it flips a bit only through the rare nuclear reaction whose recoil acts like a heavy ion. On the dose route, electrons, protons and photons pass in enormous numbers, each too light to matter alone.

3

Transient conductive path

For nanoseconds the tunnel oxide and the inter-poly dielectric conduct along the track. Electrons leave the floating gate through that path; a track beside the gate charges the oxides nearby and shifts the threshold by a much smaller amount. Many light ions open many paths, and the losses add; dose takes charge a little at a time. On an erased cell the path lets electrons onto the net positive gate instead: the threshold moves up toward neutral, never across a reference, and the state does not change.

4

Threshold shift

The threshold falls by an amount set by the LET. If it crosses a reference the cell reads as the level below; Gray coding turns a one-level drop into a one-bit error on one page, and a larger drop into errors on more pages.

5

Accumulation

Losses add over the fluence since the last program, and dose adds a loss to every cell. The model sums Poisson statistics of strong and weak hits against the distribution of critical loss over the cells and gives the error probability per cell.

6

Erase and program

The block erase drives the electrons of the floating gate out through the tunnel oxide into the p-well, past neutrality, until the gate is net positive; in a 3D charge-trap string the trapped electrons de-trap into the channel first and holes from the channel are then trapped in the nitride. Program tunnels electrons back from the channel until the cell verifies at its level. The model counts losses since the last program; the published record adds partial annealing, leakage of hit cells after rewrite, and retention loss after dose.

Where the charge goes

The stored charge of a floating-gate cell sits at the surfaces of an isolated conductor between two dielectrics: the tunnel oxide below, about 8 nm, through which the cell is programmed and erased, and the inter-poly dielectric above, through which the control gate couples to it. Cellere and colleagues showed that a single ion through the tunnel oxide of a floating-gate cell removes a part of the stored charge within the time the ionization column exists, and that the loss grows with LET: the dense column makes the oxide briefly conductive along the track, a transient conductive path, and the field of the stored charge drives electrons out through it [1, 2]. Because the gate is a conductor, charge from the whole gate can reach the path; a direct hit empties a fraction of the gate that depends on the LET, generally linear at normal incidence for ions of similar energy and sub-linear across a cyclotron cocktail [2, 6]. The same group found that ions passing beside a cell shift its threshold by much less, through charge trapped in the surrounding oxides, and that the strong-hit losses are broadly distributed with extreme tails [4, 7].

A 3D charge-trap cell stores its electrons in a silicon nitride tube that lines the channel hole, at the height of each word-line plate. The nitride is an insulator, so the charge of a cell stays near where it was injected rather than spreading over a conductor. A track through the tunnel oxide drains only the charge within its reach, a small part of the ring, which is the picture behind the smaller per-hit shifts and the continuous, exponential post-strike tails measured on 3D parts [16, 17]. The geometry adds its own feature: an ion running along the channel hole crosses every plate of the string and can touch one cell per plate, so one particle produces a column of errors through the stack, mapped by the interleaving to several pages [18]. The planar strong and weak classes of the model have no demonstrated counterpart in charge-trap cells, and the recent X-ray and alpha data on replacement-gate cells show a different coupling of dose and particle losses than planar cells [19].

Levels, references and which bits flip

A cell reads by comparing its threshold with the references between the levels. SLC has one reference; a programmed cell fails when its threshold falls below it. MLC places four levels in the window, TLC eight, so each programmed level has a fraction of the SLC margin to the reference below it, and the same loss that an SLC cell survives flips a TLC cell. The levels carry Gray-coded values, adjacent levels differing in one bit: in the view, MLC reads 11, 10, 00, 01 from the erased level upward, with the lower page in the least significant bit. A one-level drop is therefore a single-bit error on one page, and a drop across two references errs on two pages. Which page the erased-adjacent level belongs to is a vendor choice; the view states its convention.

Charge loss moves thresholds down only. A cell can cross a reference by loss only if its neutral level, the threshold with no stored charge, lies below that reference. Erased SLC blocks of a 90 nm part showed no errors to 1 Mrad(SiO2) for that reason [13], while erased-cell errors in sub-40 nm parts have been attributed to a neutral distribution straddling the read level or to data scrambling [14]. Upsets of erased cells have also been reported with high-energy protons and heavy ions, for which simple charge loss with the natural spread of the neutral threshold does not account, at least for SLC, and whose mechanism remains open [25]. In NAND the neutral threshold lies between the states, near the read level: cells are erased to about −3 V by design, past neutrality, since over-erase is no concern in a NAND string, and the programmed and erased thresholds of a cell both relax toward the neutral value on baking [27, 28]. The erased gate is therefore net positive, with the oxide field reversed, and what radiation can do to it is let electrons in, which moves the threshold up toward neutral; the textbook picture of an erased gate with very little charge [26] is the NOR one, where erase past neutrality is avoided by compaction [28]. The view places the neutral threshold 0.2 V below the lowest read reference of each scheme (an assumption: the JPL papers do not give the window of the part), draws the erased cell with its positive charge, and shows a strike on it moving the threshold toward neutral and never across a reference. The model has no term for erased cells and the published tests of the part loaded every cell with 0, the charged state [11]; the view applies the model's losses to the positive charge, capped at neutral, as an illustration.

The model that puts numbers on it

The quantitative layer of the view, and of the estimator that goes with it, is the method of JPL Publication 16-9 [8] as fitted to the Samsung 8 Gb SLC NAND flash [9, 10]. It treats charge loss as a compound Poisson process with two classes of interaction. A strong interaction, a track through the floating gate, occurs with cross section σS per cell and removes a loss LET/b1; a weak interaction, a track nearby, occurs with cross section σW and removes LET/b2. Losses are measured in units of the Weibull scale of the critical loss, the loss a cell can take before it reads wrong, which is distributed over the cells with shape k. Over a fluence H since the last program, the probability that a programmed cell has suffered a charge-loss error is

pCLE = 1 − Σm,n P(m; HσS) P(n; HσW) exp[−(m·LET/b1 + n·LET/b2 + xD)k],

with P the Poisson distribution and xD = μ·DEF·D·(σS/b1 + σW/b2) the deterministic loss that a dose D adds to every cell, μ = 6.25×104 MeV/(mg rad) converting dose to LET times fluence and DEF an empirical factor. At small fluence this reduces to a single-hit cross section σS[1 − exp(−(LET/b1)k)] + σW[1 − exp(−(LET/b2)k)], a Weibull curve in LET.

Parameter setb1b2σSσWkDEF
Samsung 8 Gb SLC, conservative [10, Table IV]23.211551.80×10-101.00×10-85.0571.65
Samsung 8 Gb SLC, Device A [10, Table III]23.211556.00×10-111.00×10-85.0571.65
Samsung 8 Gb SLC, Device B [10, Table III]22.822717.74×10-112.00×10-97.4191.65
Example device of the publication [8, Table II]49.038111.10×10-107.21×10-97.6431.0

b in MeV cm2/mg (LET in SiO2), σ in cm2 per cell. The conservative set raises σS of Device A to the saturation cross section measured after 150 krad(Si), and DEF was set from one 40Ar point after that dose [10].

What the arithmetic says for the conservative set: a strong hit at an LET of 9.0 MeV cm2/mg (40Ar) removes 0.39 of the scale and flips 0.83 percent of the cells it hits; a weak hit removes 0.008 of the scale and matters only by accumulation; at 107 ions per cm2 a cell expects 0.0018 strong and 0.10 weak interactions, and the die of 8.6×109 cells expects about 1.3×105 errors; the same arithmetic with the Device A parameters gives the 4×104 of the large-fluence counts the fit was made to reproduce [9, 10]. The publication's worked mission, three years in interplanetary space behind 100 mils of aluminum with one worst-week event and no refresh, gives 6.7×10-10 per floating gate for its example device, about one cell in 2.2×109 [8]. Measured proton cross sections of the Samsung part are far below what the heavy-ion curve would suggest, 5.1×10-20 cm2 per bit at 105 MeV and 1.2×10-19 at 480 MeV, and electron exposure produced errors that rose with dose rather than with fluence [11, 12]. The view's MLC and TLC levels rescale the SLC fit by each level's fraction of the margin; that is an illustration of why the levels matter, not a fit, since the published parameters are for an SLC part.

Three routes to the same flip

The view offers the same cell three ways to lose its margin, and the model prices each. One heavy ion: a xenon track at an LET of 54 MeV cm2/mg removes 2.3 times the scale in one strong hit, more than the margin of any programmed cell, so the single-hit cross section is the saturation value σS. Many light ions: a neon track at 2.9 removes 0.125 of the scale, and the median SLC cell needs eight such hits; the view forces them onto one cell to show that the losses add, but in the model a cell collects strong hits at random, Poisson with mean HσS, which at 108 ions per cm2 is 0.018, so eight hits on one cell has a probability of order 10-19. The low-LET errors seen at large fluence come instead from the cells with the smallest critical loss, a fraction 1 − exp(−0.1255.057) = 2.7×10-5 of those hit, from the weak interactions that every ion contributes, and from the dose the beam itself deposits [9, 10]. Dose: the deterministic loss xD grows by 1.7×10-3 of the scale per krad(Si) with the conservative set, so the median SLC cell crosses its reference at about 590 krad(Si), the top TLC level at about 80 krad(Si), and the fraction of SLC cells reading wrong from dose alone at 150 krad(Si) is 9.7×10-4. Measured on the same part, that fraction was 1 to 2×10-6 unbiased and about 0.05 biased [10], which is why the dose term is the least constrained part of the model and why bias during exposure must be stated with any dose number. Electrons, protons and photons enter the model alike, through the dose and the empirical factor DEF; their physical dose enhancement in the cell stack differs, from 4.7 at 50 keV to 0.56 at 60 MeV for electrons in this part [12].

What this means in test and in flight

The model's structure carries assumptions that the planar floating-gate record qualifies, and the estimator lists each as an assumption of the run rather than hide it. Measured losses scale with the charge still stored above the neutral level, so a second hit removes less than the first and the sum of dose and ion shifts overestimates by 5 to 20 percent [1, 2, 7, 15]; the additive form is generally conservative for accumulation. Strong-hit losses are broad with extreme tails, so the single Weibull absorbs them into k, and the low-LET and low-dose predictions are tail extrapolations whose two same-date-code fits differ by two to three orders of magnitude at an LET of 1 to 3 [4, 10]. A part of the errors anneals, from a few percent to most of them over hours to months, hit cells can leak after rewrite above an LET of about 30 to 50 MeV cm2/mg, and dose shortens retention [3, 5, 20]. The cosine law generally fails: the loss follows the LET at the cell while the number of cells crossed grows with angle, and an omnidirectional cross section was about twice the normal-incidence value in one Samsung stack and below it in a Micron part [6, 21]. Dose effects depend on bias, photon or electron energy, write history and page, by orders of magnitude on the same part number [10, 12, 22]. Several parts show cross sections that fall with fluence, by 20 percent to two orders of magnitude, which the published parameter range cannot produce and which points to a depletable vulnerable subpopulation [17, 21, 23]. For missions with error correction and scrubbing, functional interrupts, functional dose failure and destructive events generally dominate the data-loss risk over cell errors, so the cell model is the right tool for long unscrubbed storage, high dose, multi-level operation and weak codes, and a bounding comparison of all mechanisms belongs in front of it.

The NAND bit-flip mission estimator that goes with this page takes the model as published with its published fits: a parameter set (or five parameters of your own), a CREME96 ion spectrum or a pasted one, the storage interval since the last program or refresh, the dose at the die, solar events, the measured proton cross sections and the code, and returns deprogrammed cells against the interval in the same way the DRAM stuck-bit estimator does. The additions the October review recommended, field-proportional losses in a log-headroom coordinate (the form in which Gerardin and colleagues modeled proton upsets in 41 nm cells [24]), measured margin distributions, a measured dose term, an angular response and a separate 3D branch, are not in it; the estimator states each as a limit instead.

The model in the view

  • Geometry: a generic floating-gate string of the 50 nm class, three cells along the bit line with the word lines across, and a generic charge-trap string of eight plates, drawn to explain and not to the scale of any product.
  • Threshold window: erased at −2.5 V; SLC programmed at 2.5 V with the reference at 0; MLC at 0.8, 2.3 and 3.8 V; TLC at 0.5 to 4.5 V in steps of 0.67 V; references midway. The neutral threshold, with no net charge on the gate, is placed 0.2 V below the lowest read reference of each scheme (−0.2 V for SLC, −1.05 V for MLC, −1.2 V for TLC); a programmed level holds electrons in proportion to its height above it, the erased cell carries a positive charge in proportion to its depth below it, and a loss can bring a cell to neutral but not past it, from either side. The Weibull scale of the model is drawn as the SLC margin of 2.5 V, so a loss of x of the scale is a shift of 2.5x volts.
  • Stored charge, one point per elementary charge: cyan for an electron, orange for a positive charge. The floating gate is a degenerately doped polysilicon conductor, so its excess charge sits at the surfaces, within a fraction of a nanometer, and divides among the faces in proportion to their capacitance to what they face: the floor to the channel through about 8 nm of tunnel oxide, the top and the two sidewalls along the word line to the control gate through about 12 nm equivalent of inter-poly dielectric, since the control gate wraps the gate along the word line. That division, about 30 percent to the channel and 70 percent to the control gate in the drawn geometry, is the coupling ratio. On a conductor the charge also crowds toward the edges and corners, where the field is strongest; the view draws about half of it within a band of the dielectric thickness along the edges, which is what gives the programmed gate its bright frame. The count follows from the threshold above neutral and the control-gate capacitance, drawn at 550 electrons per volt, the published figure for a 90 nm cell (50 electrons per 100 mV at 90 nm, 39 at 72 nm, 20 at 50 nm [26]); the drawn gate, 60 nm on a 120 nm pitch, would hold about 280 per volt, so the sheet is about twice as dense as the published scale, for legibility. That gives about 1500 electrons at the SLC programmed level and 3100 at the top TLC level; the erased gate carries the same kind of distribution with the opposite sign, about 1300 positive charges for the SLC erased level 2.3 V below neutral (the published count of about 700 electrons for a 3 V window at the 60 nm node [27] agrees with the 280 per volt above). In the charge-trap string the electrons occupy deep traps of the nitride at random sites under each plate, about 1000 per cell at the SLC level, near 1013 per cm2 over the ring, densest toward the tunnel oxide through which they were injected, with a few spread into the gap between plates; the erased cell holds trapped holes at the same kind of sites.
  • One strike: a track through the floating gate is a strong interaction with loss LET/b1; beside it, a weak one with LET/b2; far away, nothing. In the charge-trap string the loss is the share of the ring within the reach of the track, a slider from 2 to 20 nm, and the plates crossed by the track count as cells hit.
  • Many light ions: the chosen number of tracks through the gate, one after another, each with its own path and loss LET/b1; the readout gives the number of such hits the median cell can take. Dose: a shower of electrons, protons and photons while the loss grows as xD, the charge leaving by the two mechanisms of the published picture, electrons emitted over the tunnel-oxide barrier and carriers generated in the oxides swept into the gate [26]; the plot shifts the whole level and shades the part below its reference, the dose-only error fraction 1 − exp(−(xD/m)k) for a level of margin m.
  • Accumulation: the cell draws its own strong and weak hits from the Poisson means HσS and HσW, and the dose shift is applied continuously; the readout gives the population result from the equation above.
  • Protons: the strike shows a nuclear recoil with an LET of 10 MeV cm2/mg; the readout uses the measured proton cross section of the part and the dose the proton fluence deposits at 0.0073 MeV cm2/mg.
  • Erase and program: the threshold ramps from where the cell is down to the erased level and back up to its level, and the points follow it. Floating gate: the electrons tunnel out to the p-well, and once the gate is neutral each further electron that leaves adds a positive charge; program brings electrons from the channel, which neutralize the gate and then accumulate. Charge trap: the trapped electrons de-trap into the channel first, then holes from the channel are trapped in the nitride [29]; program brings electrons that neutralize the holes and then accumulate. Voltages of order 20 V for erase and 18 V for program are typical figures, not those of the part.
  • Shown, not computed: the number and paths of the carriers, the radius of the column, the timing of the stages.

This page is an explainer. Its numbers come from the model and parameter sets listed above and are not a qualification result for any part or mission; see the site disclaimer.

References

  1. G. Cellere et al., "Anomalous charge loss from floating-gate memory cells due to heavy ions irradiation," IEEE Transactions on Nuclear Science, vol. 49, no. 6, pp. 3051 to 3058, 2002.
  2. G. Cellere et al., "Transient conductive path induced by a single ion in 10 nm SiO2 layers," IEEE Transactions on Nuclear Science, vol. 51, no. 6, pp. 3304 to 3311, 2004.
  3. G. Cellere et al., "Radiation induced leakage current in floating gate memory cells," IEEE Transactions on Nuclear Science, vol. 52, no. 6, pp. 2144 to 2152, 2005.
  4. G. Cellere et al., "Secondary effects of single ions on floating gate memory cells," IEEE Transactions on Nuclear Science, vol. 53, no. 6, pp. 3291 to 3297, 2006.
  5. M. Bagatin et al., "Annealing of heavy-ion induced floating gate errors: LET and feature size dependence," IEEE Transactions on Nuclear Science, vol. 57, no. 4, pp. 1835 to 1841, 2010.
  6. G. Cellere et al., "Angular dependence of heavy ion effects in floating gate memory arrays," IEEE Transactions on Nuclear Science, vol. 54, no. 6, pp. 2371 to 2378, 2007; and G. Cellere et al., "Effect of ion energy on charge loss from floating gate memories," IEEE Transactions on Nuclear Science, vol. 55, no. 4, pp. 2042 to 2047, 2008.
  7. M. Bagatin et al., "Sensitive volume and extreme shifts in floating gate cells irradiated with heavy ions," IEEE Transactions on Nuclear Science, vol. 62, no. 6, pp. 2815 to 2821, 2015; and S. Gerardin et al., "Heavy-ion induced threshold voltage tails in floating gate arrays," IEEE Transactions on Nuclear Science, vol. 57, no. 6, pp. 3199 to 3205, 2010.
  8. L. D. Edmonds, "A method for estimating the probability of floating gate prompt charge loss in a radiation environment," JPL Publication 16-9, Jet Propulsion Laboratory, California Institute of Technology, 2016.
  9. L. D. Edmonds, F. Irom and G. R. Allen, "Fitting a model to floating gate prompt charge loss test data for the Samsung 8 Gb SLC NAND flash memory," RADECS, 2016.
  10. L. D. Edmonds, F. Irom and G. R. Allen, "Total ionizing dose influence on the single event effect sensitivity in Samsung 8Gb NAND flash memories," IEEE Transactions on Nuclear Science, vol. 64, no. 8, pp. 2046 to 2053, 2017.
  11. F. Irom, G. R. Allen and D. J. Sheldon, "Heavy ion, proton and electron single-event effect measurements of a commercial Samsung NAND flash memory," IEEE Radiation Effects Data Workshop, 2016.
  12. F. Irom, L. D. Edmonds, G. R. Allen, W. Kim and S. Vartanian, "Electron irradiation of Samsung 8-Gb NAND flash memory," IEEE Transactions on Nuclear Science, vol. 65, no. 1, pp. 27 to 33, 2018.
  13. M. Bagatin et al., "TID sensitivity of NAND flash memory building blocks," IEEE Transactions on Nuclear Science, vol. 56, no. 4, pp. 1909 to 1913, 2009.
  14. S. Gerardin et al., "Degradation of sub 40-nm NAND flash memories under total dose irradiation," IEEE Transactions on Nuclear Science, vol. 59, no. 6, pp. 2952 to 2958, 2012.
  15. M. Bagatin et al., "Increase in the heavy-ion upset cross section of floating gate cells previously exposed to TID," IEEE Transactions on Nuclear Science, vol. 57, no. 6, pp. 3407 to 3413, 2010.
  16. Y. Jiao et al., "Heavy ion-induced SEUs and layer dependence in dual-deck 3-D NAND flash memories," IEEE Transactions on Nuclear Science, vol. 72, no. 8, pp. 2689 to 2696, 2025.
  17. D. Chen et al., "Heavy ion and proton-induced single event upset characteristics of a 3-D NAND flash memory," IEEE Transactions on Nuclear Science, vol. 65, no. 1, pp. 19 to 26, 2018.
  18. E. P. Wilcox and M. J. Campola, "A TID and SEE characterization of multi-terabit COTS 3D NAND flash," IEEE Radiation Effects Data Workshop, 2019; and E. P. Wilcox, M. B. Joplin and M. D. Berg, "Single-event effects response of 96- and 176-layer 3D NAND flash memories," IEEE Radiation Effects Data Workshop, 2023.
  19. M. Bagatin, S. Beltrami, A. Paccagnella and S. Gerardin, "X-ray effects on alpha-induced soft error rate in 3-D NAND replacement gate flash memory cells," IEEE Transactions on Nuclear Science, vol. 72, no. 8, pp. 2536 to 2541, 2025.
  20. T. R. Oldham et al., "Retention characteristics of commercial NAND flash memory after radiation exposure," IEEE Transactions on Nuclear Science, vol. 59, no. 6, pp. 3011 to 3015, 2012; and M. Bagatin, S. Gerardin and A. Paccagnella, "Retention errors in 65-nm floating gate cells after exposure to heavy ions," IEEE Transactions on Nuclear Science, vol. 59, no. 6, pp. 2785 to 2790, 2012.
  21. K. Gruermann et al., "SEU and MBU angular dependence of Samsung and Micron 8-Gbit SLC NAND-flash memories under heavy-ion irradiation," RADECS, 2011.
  22. M. J. Kay et al., "Effect of accumulated charge on the total ionizing dose response of a NAND flash memory," IEEE Transactions on Nuclear Science, vol. 59, no. 6, pp. 2945 to 2951, 2012.
  23. D. Chen et al., "Heavy ion irradiation fluence dependence for single-event upsets in a NAND flash memory," IEEE Transactions on Nuclear Science, vol. 64, no. 1, pp. 332 to 337, 2017.
  24. S. Gerardin et al., "Proton-induced upsets in 41-nm NAND floating gate cells," IEEE Transactions on Nuclear Science, vol. 59, no. 4, pp. 838 to 844, 2012.
  25. M. R. Shaneyfelt and N. A. Dodds, "Upsets in erased floating gate cells with high-energy protons," IEEE Transactions on Nuclear Science, vol. 64, no. 1, pp. 421 to 426, 2017.
  26. M. J. Marinella, "Radiation effects in advanced and emerging nonvolatile memories," IEEE Transactions on Nuclear Science, vol. 68, no. 5, pp. 546 to 572, 2021.
  27. S. Aritome, NAND Flash Memory Technologies, IEEE Press and Wiley, 2016 (erase to −3 V, section 2.2; the neutral threshold, section 6.2.3; electrons per bit versus node, Fig. 5.27).
  28. J. E. Brewer and M. Gill, eds., Nonvolatile Memory Technologies with Emphasis on Flash, IEEE Press and Wiley, 2008 (the NAND erased state as overerased, section 3.3.1; compaction in NOR, section 3.2.3).
  29. R. Micheloni, ed., 3D Flash Memories, Springer, 2016 (charge-trap erase by electron de-trapping then hole injection, section 2.4; hole supply by gate-induced drain leakage or from the substrate, sections 4.6 and 7.5.3).

To the NAND bit-flip mission estimator · Estimator help and sources · The DRAM stuck-bit mechanism in 3D · Memory EDAC tool