A Nonvolatile Memory Overview

By Jitu J. Makwana, Dr. Dieter K. Schroder*
A Nonvolatile Memory Overview
Jitu J. Makwana, Dr. Dieter K. Schroder*
E-mail: j.makwana@worldnet.att.net, *schroder@asu.edu
This paper presents a basic nonvolatile memory (NVM) overview. Section I begins with the introduction including a brief background of NVM's and the common terms used in the memory industry. The description and explanation of how an NVM is programmed (adding electrons) using hot-carrier injection is covered in section II. Section III covers the erasing or removing of electrons from floating gates of NVM's. A brief mechanism of Fowler-Nordheim tunneling is covered. Section IV introduces the model that can be used to predict the NVM programming characteristics. The hot-carrier injection model addressed is the "Lucky Electron" model. Section V covers the reliability aspects of NVM's. The common reliability issues an NVM encounters are the data retention, endurance, and disturbs.
KEY WORDS: nonvolatile, memory, hot-carrier injection, tunneling, reliability, retention, memory disturbs, EPROM, Flash EEPROM
I. INTRODUCTION
Memory can be split into two main categories: volatile and nonvolatile. Volatile memory loses any data as soon as the system is turned off; it requires constant power to remain viable. Most types of random access memory (RAM) fall into this category. Nonvolatile memory does not lose its data when the system or device is turned off. A nonvolatile memory (NVM) device is a MOS transistor that has a source, a drain, an access or a control gate, and a floating gate. It is structurally different from a standard MOSFET in its floating gate, which is electrically isolated, or "floating". Nonvolatile memories are subdivided into two main classes: floating gate and charge-trapping. Kahng and Sze proposed the first floating gate device in 1967 [1]. In this memory, electrons were transferred from the floating gate to the substrate by tunneling through a 3 nm thin silicon dioxide (SiO2) layer. Tunneling is the process by which an NVM can be either erased or programmed and is usually dominant in thin oxides of thicknesses less than 12 nm. Storage of the charge on the floating gate allows the threshold voltage (VT) to be electrically altered between a low and a high value to represent logic 0 and 1, respectively.
In floating gate memory devices, charge or data is stored in the floating gate and is retained when the power is removed. All floating gate memories have the same generic cell structure. They consist of a stacked gate MOS transistor as shown in figure 1. The first gate is the floating gate that is buried within the gate oxide and the inter-polysilicon dielectric (IPD) beneath the control gate. The IPD isolates the floating gate and can be oxide or oxide-nitride-oxide, ONO. The SiO2 dielectric surrounding the transistor serves as a protective layer from scratches and defects. The second gate is the control gate which is the external gate of the memory transistor. Floating gate devices are typically used in EPROM (Electrically Programmable Read Only Memory) and EEPROM's (Electrically Erasable and Programmable Read Only Memory).
Figure 1. A typical floating gate memory structure.

Charge-trapping devices were invented in 1967 [2] and were the first electrically alterable semiconductor devices. In charge-trapping memory devices, charge or data is stored in the discrete nitride traps and is also retained when the power is removed. Charge-trapping devices are typically used in MNOS (Metal Nitride Oxide Silicon) [3], [4], SNOS (Silicon Nitride Oxide Semiconductor) [5], and SONOS (Silicon Oxide Nitride Oxide Semiconductor) [6]. Figure 2 shows a typical MNOS charge-trapping memory structure.
Figure 2. An MNOS memory cell structure.

The charges in MNOS memories are injected from the channel region into the nitride by
quantum mechanical tunneling through an ultra-thin oxide (UTO) which is typically 1.5
- 3 nm.
The first EPROM, a floating gate device, was developed using a heavily doped polysilicon (poly-Si) as the floating gate material known as the floating gate avalanche-injection MOS memory (FAMOS) [7]. The gate oxide thickness was of the order of 100 nm to prevent weak spot or shorting path between the floating gate and the substrate. Charging of the EPROM was achieved by biasing the drain junction to avalanche breakdown where the electrons in the avalanche plasma were injected from the drain region into the floating gate. The FAMOS could only be erased by ultraviolet (UV) or x-ray. The EPROM was perceived as a tool for system prototyping before a design was committed to Read Only Memory (ROM). Today, one can obtain EPROM's in either a ceramic package with a quartz window that allows for UV exposure or a plastic package without a quartz window. These memories are known as one-time-programmable (OTP) EPROM's. OTP's are inexpensive, however, additional testing after assembly is not possible. EPROM's in ceramic packages with a quartz window are expensive but do allow additional testing since the memory can be erased using UV light.
Although the 1970's saw the UV-erasable, electrically programmable memories become commercially successful, there was an ever-present attraction toward making the EPROM's electrically erasable, EEPROM. H. Iizuka et. al [8], proposed the first electrical erasing NVM known as the stacked gate avalanche-injection MOS (SAMOS) memory. SAMOS memory consisted of double poly-Si gates with an external control gate. The external control gate made electrical erasability possible and as a result improved the erasing efficiency. The EEPROM's basic approach with electrical means of restoring the charged floating gate to its original uncharged status replaced UV emission approach. Cheaper packaging and a greater ease of use were the first advantages of EEPROM's over their UV-erasable counterparts. The disadvantage of EEPROM's was the cell size that was two to three times the size of an EPROM cell that resulted in a larger die size. EEPROM cells consist of two transistors, one, a floating gate transistor and the other, a select gate transistor, as shown in figure 3. The select gate transistor is used to select or deselect floating gate transistors for programming or erasing. Die size was further increased to incorporate error correction circuitry or redundancy circuits.
Figure 3. An EEPROM with select gate transistor.

During the 1980's, a novel nonvolatile memory product was introduced, referred to as Flash EEPROM [9]. The first products were merely the result of adapting EPROM's in such a way that the cell could be erased electrically as well. These devices used hot-electron injection for programming and tunneling for erasing. This new genre of Flash EEPROM's could not be erased by bytes but could only be erased by the entire chip or large sections of the chip. Since the need to erase by bytes as in EEPROM's was no longer needed in Flash EEPROM's, the select transistor was removed from the cell structure. Thus the Flash EEPROM's were two to three times smaller than earlier EEPROM cells. The generic cell structure of a Flash EEPROM is similar to a generic cell structure shown in figure 1.
Below is a nomenclature or a list of memory terms (not exhaustive) used in the literature, industry, and education fields:
Bit - The basic unit of memory, "1" or "0".
Byte - A group of 8 bits.
Cell - The physical semiconductor structure that stores one bit of data.
Array - Repetition of memory cell in a two-dimensional matrix.
RAM - Random access memory is fast, temporary storage for your computer
ROM - Read-only memory is fast, permanent storage for your computer.
Program - The operation of adding or removing electrons from the storage
medium** of a memory cell. Sometimes called "Write". Charge is altered in the storage medium and thus the threshold voltage.
Erase - The operation of adding or removing electrons from the storage
medium** of a memory cell. Charge is altered in the storage medium and thus the threshold voltage.
** Floating gate (1st Polysilicon) or discrete nitride traps.
Read - The process of determining the state of the bit cell.
Endurance - Write/erase cycles a memory can endure before failure, typical
guarantee is 10 K cycles.
Injection - Common terms HCI: Hot-carrier Injection, HEI: Hot-electron
Injection, CHEI: Channel Hot-electron Injection. Process of adding charge using high fields.
Tunneling - Process of adding/removing charge but does not require high fields.
Gate oxides are thin < 12 nm.
Data Retention - Typically a time value pertaining to a memory cell's ability to
retain data.
Disturb - Charge gain or loss in a memory cell, gate disturb, read disturb,
program disturb, etc.
EPROM - Electrically programmable read-only memory.
EEPROM - Electrically programmable and erasable read-only memory.
Flash - Term used to describe erasing of a memory in large sections.
II. BASIC PROGRAMMING MECHANISMS
In both the floating gate and charge-trapping memories, the charge needed to program the device has to be injected into the floating gate or into the nitride layer respectively. In order to change the charge or data content of NVM';s, two major mechanisms have been shown to be viable: FN tunneling (F-N) through thin oxides (< 12 nm) [10] and channel hot-electron injection (CHE) [11].
IIa. Fowler-Nordheim (FN) Tunneling
One of the most important injection mechanisms used in NVM's is FN tunneling. When a large voltage Vcg is applied at the control gate during programming, its energy band structure will be influenced as shown in figure 4.
Figure 4. Energy band diagram of a floating gate memory during programming by FN tunneling.

In the figure, ec and ev are the conduction and valence bands respectively, Eg is the energy band gap (1.1 eV for silicon), fb is the Si-SiO2 energy barrier (fb is 3.2 eV for electrons and 4.7 eV for holes). The applied Vcg creates the electric field resulting in a potential barrier. This barrier provides a path for the electrons in the substrate to tunnel through the thin gate oxide (typically less than 12 nm) and eventually be collected in the n+ poly-Si floating gate. The bending of the energy bands of the IPD and the gate oxide are different due to the thickness differences between them. The IPD ranges from 25 nm to 45 nm while the gate oxide ranges from 5 nm to 12 nm. The electrons collected at the floating gate leads to a tunneling current density and is given by [12].
(1)
with
(2)
and
(3)
where,
h = Planck's constant
= Energy barrier at the injecting surface (3.2 eV for Si-SiO2)
q = Charge of a single electron (1.6x10-19 C)
m = Mass of a free electron (9.1x10-31 kg)
m* = Effective mass of an electron in the band gap of SiO2 (0.42 m)
= h/2
inj = Electric field at the injecting surface =(V/cm)
Vapp = Voltage applied across the tunnel oxide (V)
Vfb = Flat band voltage (V)
tox = Tunnel oxide thickness (cm)
Equation 1 shows that tunneling current density is exponentially dependent on the applied voltage, Vapp, which influences the electric field, ?inj, across the gate oxide.
Figure 5 shows a cross-section of an NVM with electrons tunneling uniformly with Vcg at positive potential while the source (Vs), the drain (Vd), and the substrate (Vsub) are at ground potential.
Figure 5. Uniform tunneling to program Flash EEPROM.

An optional method that can be used to program Flash EEPROM's is given in figure 6 which is called drain-side tunneling. Drain-side tunneling is sometimes preferred over the uniform tunneling due to the programming speed as a result of higher tunneling current density due to smaller injecting area.
Figure 6. Drain-side tunneling to program Flash EEPROM.

IIb. Hot-carrier Injection (HCI)
NVM's can also be programmed by hot-carrier injection. The method of programming is by hot-electron injection for n-type NVM's built on p-substrates and by hot hole injection for p-type NVM's built on n-substrates. Hot-hole injection is very slow due to the hole mass as well as the Si-SiO2 energy barrier of 4.7 eV for holes, which is why all NVM's manufactured today are n-type on p-substrates.
The memory cell is programmed by charging the floating gate via the injection of hot-electrons from the drains pinch-off region. The hot-electrons get their energy from the voltage applied to the drain (Vd) of the memory cell. They are accelerated by the lateral electric field (Elat) along the channel into even higher fields surrounding the drain depletion region. Once these electrons gain sufficient energy they surmount the energy barrier of 3.2 eV between the silicon substrate and the silicon dielectric layer or gate oxide.
With positive Vd and channel voltages, electrons injected into the oxide of an n-channel memory cells return to the substrate unless a high positive Vcg is applied to pull the electrons toward the floating gate. The energy band structure for NVM programming by hot-electron injection is shown in figure 7.
Figure 7. Energy band diagram of a floating gate memory during programming by hot-electron injection.

As the floating gate becomes fully charged, the gate current (Ig) is reduced to almost zero
because the oxide electric field (Eox) (in the beginning of the injection process Eox was
attractive to the electrons) is now repulsive to the electrons. In general, to the first order,
Vcg increases the charge on the floating gate while Vd affects the programming speed.
Figure 8 shows a cross-section of an NVM with hot-electron injection programming. Vcg and Vd are at positive potential of 15 V and 10 V respectively while Vs and Vsub are at ground potential. The p-well is also shown, as it is the process needed to separate n-channel and p-channel MOS transistors from NVM's devices.
Figure 8. Hot-electron injection mechanism for programming in NVM's.

III. BASIC ERASING MECHANISMS
Section II covered the two programming schemes, namely, FN tunneling and hot-electron injection. In order to reprogram an NVM, it first has to be erased. This section will cover the erasing schemes commonly employed in the industry.
The electrons that are injected into the floating gate are trapped by the high gate to oxide
energy barrier of 3.2 eV. Since the potential-energy barrier at the oxide-silicon interface is greater than 3.0 eV, the rate of spontaneous emission of electrons from the oxide over this barrier is negligibly small. The net negative charge which remains on the floating gate shifts the VT to a positive value.
There are two methods of erasing or removing charge:
UV emission.
FN tunneling.
IIIa. UV Emission
Referring to figure 9, electrons gain enough energy acquired from the UV radiation to surmount the energy barrier from the floating gate to either the control gate or to the substrate, which reduces the VT. The typical time it takes to change the VT from programmed state to neutral or erased state is 10 minutes.
Figure 9. Energy band diagram of UV erase of an NVM.

IIIb. FN Tunneling
FN tunneling can also be used to erase an NVM. One of the methods is by applying a large negative voltage at the control gate. The energy band structure will be influenced as shown
in figure 10. The applied Vcg creates the electric field resulting in a potential barrier. This barrier provides a path for the electrons to tunnel from the floating gate to the substrate through the thin gate oxide.
Figure 10. Energy band diagram of a floating gate memory during erasing by FN tunneling.

Figure 11a and 11b shows two choices to erase a Flash EEPROM. For uniform tunneling, a large negative Vcg is applied while for drain-side tunneling method, both a negative Vcg and a positive Vd are applied.
Figure 11a. Uniform tunneling to erase Flash EEPROM.

Figure 11b. Drain-side tunneling to erase Flash EEPROM.

In general, uniform tunneling is slower that drain-side tunneling, but, drain-side tunneling tends to cause reliability issues. The reliability issue is the gate oxide damage that occurs near the drain since a small area is bombarded by electrons and that the tunneling current density as a result of small area is higher.
IV. HOT-CARRIER INJECTION MODEL
One of the methods of programming an EPROM or a Flash EEPROM is by channel hot-electron injection (CHE) where hot-electrons are generated in the high field region between the pinched-off channel and drain. Electrons with sufficient energy are injected across the oxide to the floating gate, thereby programming the device (increasing the threshold voltage to positive value, VT). This process of programming is slow due to the injection efficiency, which is dependent on three probability events. The hot-carrier injection mechanism gives rise to impact ionization at the drain, by which both minority (electrons) and majority (holes) carriers are generated. The highly energetic holes are normally collected at the substrate contact and form the substrate current (Isub) while the minority carriers are collected at the drain and forms the drain current (Ids). If the oxide electric field (Eox) favors injection, these carriers are injected over the energy barrier (fb) of the gate oxide and gives rise to hot-carrier injection gate current (Ig). In the case of floating gate memories, these electrons change the charge content of the floating gate.
There are two models that can be used to describe the gate current due to hot-electron injection. The two models are the lucky-electron model [13] and the effective electron temperature model [14].
IVa. Lucky-electron Model and Threshold Programmed VT
The lucky-electron approach of modeling the hot-electron distribution was originated by Shockley [15]. Conceptually, the lucky-electron model can be described as follows. In order for hot-electrons to reach the gate, the hot-electrons must gain sufficient kinetic energy from the lateral channel field (Elat) and have its momentum redirected towards the Si-SiO2 interface in order to surmount the SiO2 energy barrier (fb). Figure 12 shows the concepts involved in the lucky-electron model. The three events involved in the lucky-electron model are:
A - B Event: A channel electron has to gain energy from the Elat and become "hot". The hot-electron momentum has to be re-directed towards the Si-SiO2 interface. The probability associated with this process isand is defined as the probability of an electron having enough normal momentum to surmount the Si-SiO2 potential barrier height.
Figure 12. The three processes in the lucky-electron injection model.

B - C Event: Once the hot-electron is re-directed, it must not suffer any energy robbing collisions. The probability associated with this event is PSEMI. PSEMI is defined as the probability of an electron traveling to the Si-SiO2 interface without suffering any collisions.
C - D Event: While in transit from the Si-SiO2 interface to the floating gate, the electron must not suffer any collisions in the oxide image potential well. The probability associated with this event is Pinsul and is defined as the probability of an electron suffering no collision in the oxide image potential well.
Since these three probabilities are statistically independent, the resultant probability is the
product of the probability for each individual event. The gate current is given by
(4)
where,
lr = Momentum re-direction scattering mean free path = 92 nm
Leff = Effective channel length of the floating gate transistor (cm)
Ids = Drain-source current (A)
The charge on the floating gate changes the threshold voltage (VT) of the floating gate transistor by
(5)
where,
DVT = VT (Programmed) - VT (Initial)
DQfg = Qfg (Programmed) - Qfg (Initial) = Change in floating gate charge.
The floating gate charge changes according to
(6)
where,
Dt is the programming time (s)
The change in floating gate charge shifts VT from the initial or natural threshold voltage VTi by
(7)
where,
Cfg = Floating gate to control gate capacitance (F)
Figure 13 shows a typical transfer characteristic for the programmed and initial or erased states. The figure shows that the Ids-Vcg curves are parallel to each other. The shift of the Ids-Vcg curve from initial or erased state to programmed state is equal to DQfg/ Cfg.
Figure 13. Ids-Vcg transfer characteristics.

V. NONVOLATILE MEMORY RELIABILITY
Nonvolatile memory cells have some important functional characteristics, which are used to evaluate the performance of the cell. These characteristics are divided into two main classes, namely endurance and retention. In order to understand endurance and retention characteristics, it is imperative to know some of the fundamentals associated with gate oxide and interpolysilicon dielectric, IPD, integrity. Although traps are the storage sites in MNOS, SNOS, and SONOS memories, they constitute the very means that lead to reliability failures in EPROM's, EEPROM's, and Flash EEPROM's. The gate oxide and IPD quality can affect endurance and retention.
The primary failure mechanism of the gate oxide pertains to oxide breakdown and trap-up due to high injection electric field stressing during hot-electron injection or FN tunneling. It was suggested that oxide defects and broken Si-O bonds serve as trapping centers [16] for positive (holes) charge. Oxide breakdown occurs after a fixed amount of charge per unit area (Qbd) has been injected and has been shown to be a function of applied electric field [17]. Qbd is an industry standard electrical test used to measure the quality of the oxide with higher Qbd (good oxide quality) as the desired goal. Trap-up is defined as the trapping of electrons in the oxide during programming/writing operations. These trapped charges change the injection fields and thus, the amount of charge transferred to and from the floating gate during programming.
As described earlier, one of the components of a nonvolatile memory cell structure is the IPD. In nonvolatile memories, IPD is used to isolate the floating gate from other electrodes (control gate, source, drain, and the substrate) and hence, should be defect-free to prevent charge leakage from the floating gate. Since the floating gate is a poly-Si layer, it is commonly oxidized during the IPD growth process. The oxidation of poly-Si layer modifies the surface topology due to enhanced oxidation at the grain boundaries of the poly-Si, forming interface protuberances and inclusions [18]. The surface nonuniformities causes electric field enhancement resulting in higher leakage currents, which is a drawback for an insulating IPD. Figure 14 shows a circuit diagram of the effects of nonuniformities with respect to electric field enhancement. Other factors that influences the IPD quality are the doping of the floating gate poly-Si layer and the temperature of both the poly-Si deposition and the oxidation [19]. Multiple dielectric stacks such as the oxide-nitride-oxide (ONO) are now commonly used as IPD's for lower leakage due to lower defect densities and higher electric field properties [20]. The lower leakage currents are achieved due to the fact that the electrons that have leaked from the floating gate gets trapped in the oxide-nitride interface which builds an electric field that opposes further charge loss [21]. Typical thicknesses of the ONO stack are 5 - 10 nm, 20 nm, and 3 nm for the bottom oxide, nitride, and the top oxide respectively. The bottom oxide is the oxide above the floating gate while the top oxide is the oxide beneath the control gate.
Figure 14. IPD leakage current due to nonuniformities.

Va. Endurance Characteristics
The endurance characteristics give the memory threshold voltage window, which is the difference between the threshold voltages in the programmed/written state and the erased states, as a function of the number of programming cycles, as shown in figure 15. Nonvolatile memories can be programmed and erased frequently at the expense of introducing permanent gate oxide damage such as oxide breakdown and trap-up. This implies that the total number of program operations is limited; for example, most commercially available EEPROM products are guaranteed to withstand 106 programming cycles. The damaging of the memory cell during cycling is normally referred to as "degradation" and the number of cycles the memory can withstand is called "endurance". Threshold voltage window closure occurs when the threshold voltage difference between the programmed and erased states cannot be distinguished. The phenomenon of window closure has been attributed to trapping of injected electrons in the gate oxide due to pre-existing electron traps. There is also evidence of trap generation during programming and erasing due to high electric field stress (Einj). It is thus important to have high quality gate oxides that can endure constant electron stress during hot-electron injection or FN tunneling.
Figure 15. Typical EEPROM cell threshold voltage window versus log cycles.

Vb. Retention Characteristics
When a nonvolatile memory cell can no longer hold the charge in the floating gate, it is said to have affected its retention capability. Retention is a measure of the time that a nonvolatile memory cell can retain the charge whether it is powered or unpowered. In floating gate memories, the stored charge can leak away from the floating gate through the gate oxide or through the IPD. This leakage caused by mobile ions and oxide defects, result in a shift of the threshold voltage of the memory cell. Different charge loss mechanisms have been described [22], [23], namely, charge loss due to thermionic emission, charge loss due to electron detrapping, and charge loss due to contamination such as positive mobile ions. To improve the retention characteristics of the memory cell, various improvements to the quality of the gate oxide and IPD becomes very important.
Retention can be quantified by measuring or estimating the time it takes for the floating gate to discharge when it is intended to keep the information stored. When charge loss occurs, a shift in the VT of the memory cell occurs according to equation 8.
(8)
where dQFG, CFG, and dVT are the floating gate charge loss, floating gate capacitance, and the floating gate VT shift respectively. Equation 9 shows the evaluation of the number of electrons lost and equation 10 shows how the number of electrons lost is related to leakage current, ILeakage and retention time, dt.
(9)
(10)
For a typical CFG of 30 fF and a VT shift of 3 V, the # of electrons lost from the floating gate to control gate is 5.6x105. Table 1 shows the retention time, dt, for various ILeakage associated with a loss of 5.6x105 electrons or 3 V VT shift.
Leakage Current, ILeakage (A)
Retention Time, dt (Years)
1x10-20
0.28
5x10-21
0.56
1x10-21
2.84
5x10-22
5.68
2.85x10-22
10
Table 1. Retention time as function of ILeakage.
Table 1 shows that it would take 10 years for an NVM to loose charge equivalent to 3 V shift for a leakage current of 2.85x10-22 A.
Vc. Memory Disturbs
Widespread use of nonvolatile memories in production systems requires data retention
for ten years or more. A typical memory array undergoes stresses that arise during programming and erasing commonly referred to as disturbs. The four principal memory cell disturbs that can occur in an array are dc erase, dc program, program disturb, and read disturb [24]. The two common disturbs that can impact memory cells during programming are dc program and program disturb. Memory cell disturbs can also occur during erasing and is called dc erase. Finally, during "read" operation when the memory is read to determine the state (logic 0 or 1), memory cell disturb can also occur and is called read disturb. Figure 16 shows the schematic description of a memory array and will be used to describe the memory cell disturbs.
Figure 16. Schematic description of a memory array with disturbs.

In figure 16, the memory array has columns connected to the drain of each memory cell
(COL 1, COL 2, and COL 3) that represent the "bitline" and are used to select or deselect the memory cell for programming or erasing. The memory array also has rows that are connected to the control gate of each memory cell (ROW 0 and ROW 1) that represent the "wordline" and are also used to select or deselect the memory cell for programming or erasing. From previous discussions on hot-carrier injection, both the bitline and the wordline have to have voltages applied in order for the memory cell to program. If either the bitline or the wordline is not connected to a high potential, then hot-carrier injection will not occur.
The four types of memory disturbs are:
1) DC Erase: This type of disturb occurs on cells that are already programmed (Cell A). These programmed cells are on the same wordline (ROW 1) as the cell being programmed (COL 2 and ROW 1). During the programming operation ROW 1 is brought to a high voltage of 15 V. The electric field due to the high voltage appears across the IPD, which may be high enough to cause conduction of electrons across it from the floating gate (also known as poly-Si to poly-Si erase mode mechanism) to the control gate. The result, loss of charge and consequent reduction of the programmed threshold voltage, which in severe cases causes complete loss of data.
2) DC Program: Also known as gate-disturb, occurs on cells that are unprogrammed or in erased state (Cell B). These unprogrammed cells are on the same wordline (ROW 1) as the cell being programmed (COL 2 and ROW 1). These unprogrammed cells have a few electrons on their floating gates and thus low threshold voltages, VT. When ROW 1 is raised to 15 V, the electric field across the tunneling gate dielectric may be high enough to cause electron tunneling into the floating gate from the substrate and increase the VT. In severe cases, the cell is programmed unintentionally and is called soft-write.
3) Program Disturb: Also known as drain-disturb occurs on programmed cells. A programmed cell (Cell C) sharing a column (COL 2) with the cell being programmed (COL 2 and ROW 1) will also experience high electric fields between the floating gate and the drain. This may cause electrons to tunnel from the floating gate to the drain, and lead to a reduced VT.
4) Read Disturb: This disturb mechanism occurs on erased bits that share a wordline with a bit that is being read. The common wordline places the control gate of the erased cells at 5 V. The selected device's drain is driven to about 1 V. The unselected bits have their source, drain, and substrate at 0 V.
VI. CONCLUSION
An NVM overview was given which began with a brief background of nonvolatile memories and the terminology's used in the NVM industry. The two common methods of programming are the hot-carrier injection and FN tunneling. In FN tunneling, gate oxides thicknesses have to be less than 12 nm whereas in the case of hot-carrier injection, the thickness of the gate oxide is not critical. The two common methods employed for erasing are by UV emission and FN tunneling. UV emission is commonly used for UV EPROM's while FN tunneling is commonly employed in EEPROM's and Flash EEPROM's, provided the thicknesses are less than 12 nm. UV emission takes a long time to erase relative to FN tunneling. Typical UV emission erase time is 10 minutes while the FN tunneling takes anywhere from 1 ms to 10 ms depending on the applied potentials at the control gate and the drain. The fastest method of programming is via hot-carrier injection that can typically take up to 100 ms. To understand the physics of hot-carrier injection, a "lucky-electron" model was covered. The model shows how the gate current is evaluated based on various probability events. Although the injection efficiency is low, the process of adding electrons onto the floating gate is fast due to high electric fields. As for every device, reliability issues are encountered. Endurance and data retention are the two common reliability issues an NVM undergoes. In addition to endurance and data retention, memory disturbs also occur due to memory stressing during programming and erasing.
References
[1] Kahng, D. and Sze, S. M. (1967) A floating gate and its application to memory devices. Bell Systems Technical Journal. 46, 1283.
[2] Wegener, H. A. R, Lincoln, A. J., Pao, H. C., O'Connell, M. R., and Oleksiak, R. E. (1967) The variable threshold transistor, a new electrically alterable, non-destructive read-only storage device. IEEE IEDM Technical Digest. 1.
[3] Wegener, H. A. R, Lincoln, A. J., Pao, H. C., O'Connell, M. R., and Oleksiak, R. E. (1967) Metal-insulator-semiconductor transistor as a non-volatile storage element. International Electron Devices Meeting. (Abstracts). 58
[4] Frohman-Bentchkowsky, D. (1970) The metal-nitride-oxide-silicon (MNOS) transistor-characteristics and applications. Proceedings of IEEE. 58, 1207.
[5] Yatsuda, Y., Hagiwara, T., Kondo, R., Minami, S., and Itoh, Y. (1979) N-channel Si-gate MNOS device for high speed EAROM. Proceedings 10th Conference in Solid State Devices. 11.
[6] Suzuki, E., Hiraishi, H., Ishi, K., and Hayashi, Y. (1983) A low voltage alterable EEPROM with metal-oxide-nitride-oxide-semiconductor (MONOS) structure. IEEE Transactions on Electron Devices. ED-30, 122.
[7] Frohman-Bentchkowsky, D. (1974) FAMOS-A new semiconductor charge storage device. Solid State Electronics. 17, 517.
[8] Iizuka, H., Masuoka, F., Sato, T., and Ishikawa, M. (1976) Electrically alterable avalanche-injection type MOS read-only memory with stacked-gate structures. IEEE Transactions on Electron Devices. ED-23, 379.
[9] Masuoka, F., Asano, M., Iwahashi, H., and Komuro, T. (1984) A new flash EEPROM cell using triple poly-Si technology. IEEE IEDM Technical Digest. 464.
[10] Yeargain, J. and Kuo, K. (1981) A high density floating gate EEPROM cell. IEEE IEDM Technical Digest. 24.
[11] Guterman, D., Rimawi, I., Chiu, T., Halvorson, R., and McElroy, D. (1979) An electrically alterable nonvolatile memory cell using a floating gate structure. IEEE Transactions on Electron Devices. ED-26, 576.
[12] Lezlinger, M. and Snow, E. H. (1969) Fowler-Nordheim tunneling in thermally grown SiO2. Journal of Applied Physics. 40, 278.
[13] Tam, S., Ko, P., and Hu, C. (1984) Lucky-electron model of channel hot-electron injection in MOSFET's. IEEE Transactions on Electron Devices. ED-31, 1116.
[14] Takeda, E., Kume, H., Toyabe, T., and Asai, S. (1982) Submicrometer MOSFET structure for minimizing hot-carrier generation. IEEE Transactions on Electron Devices. ED-29, 611.
[15] Shockley, W. (1961) Problems related to p-n junctions in silicon. Solid-State Electronics. 2, 35.
[16] Harari, E. (1978) Dielectric breakdown in electrically stressed thin films of thermal SiO2. Journal of Applied Physics. 49, 2478.
[17] Modelli, A. and Ricco, B. (1984) Electric field and current dependence of SiO2 intrinsic breakdown. IEEE IEDM Technical Digest. 148.
[18] DiMaria, D. J. and Kerr, D. R. (1975) Interface effects and high conductivity in oxides grown from polycrystalline silicon. Applied Physics Letters. 27, 505.
[19] Faraone, L. (1986) Thermal SiO2 films on n+ polycrystalline silicon: Electrical conduction and breakdown. IEEE Transactions on Electron Devices. ED-33, 1785.
[20] Mori, S., Kaneko, Y., Arai, N., Ohshima, Y., Araki, H., Narita, K., Sakagami, E., and Yoshikawa, K. (1990) Reliability study of thin inter-poly dielectrics for nonvolatile memory application. Proceedings 1985 IEEE IRPS. 132.
[21] Aminzadeh, M. (1988) Conduction and charge trapping in polysilicon nitride-oxide-silicon structures under positive gate bias. IEEE Transactions on Electron Devices. ED-35, 205.
[22] Shiner, R. E. (1980) Data retention in EPROMs. Proceedings IRPS. 238.
[23] Mielke, N. (1983) New EPROM data-loss mechanisms. Proceedings IRPS. 106.
[24] Verma, G. and Mielke, N. (1988) Reliability aspects of ETOX based flash memories. IRPS Proceedings, 158
Anatomy of Linux flash file systems
Options and architectures

Document options

Print this page

PDF - Fits A4 and Letter
69KB (10 pages)

Get Adobe® Reader®

E-mail this page
Rate this page

Help us improve this content
Level: Intermediate
M. Tim Jones (mtj@mtjones.com), Consultant Engineer, Emulex Corp.
20 May 2008
You've probably heard of Journaling Flash File System (JFFS) and Yet Another Flash File System (YAFFS), but do you know what it means to have a file system that assumes an underlying flash device? This article introduces you to flash file systems for Linux®, explores how they care for their underlying consumable devices (flash parts) through wear leveling, and identifies the various flash file systems available along with their fundamental designs.
Solid-state drives are all the rage these days, but embedded systems have used solid-state devices for storage for quite some time. You'll find flash file systems used in personal digital assistants (PDAs), cellphones, MP3 players, digital cameras, USB flash drives (UFDs), and even laptop computers. In many cases, the file systems for commercial devices can be custom and proprietary, but they face the same challenges discussed below.

Anatomy of Security-Enhanced Linux (SELinux)Anatomy of real-time Linux architecturesAnatomy of the Linux SCSI subsystemAnatomy of the Linux file systemAnatomy of the Linux networking stackAnatomy of the Linux kernelAnatomy of the Linux slab allocatorAnatomy of Linux synchronization methods
All of Tim's Anatomy of... articlesAll of Tim's articles on developerWorks
Flash-based file systems come in a variety of forms. This article explores a couple of the read-only file systems and also reviews the various read/write file systems available today and how they work. But first, let's explore the flash devices and the challenges that they introduce.
Flash memory, which can come in several different technologies, is non-volatile memory, which means that its contents persist after its source of power is removed. For a great history of flash memory devices, seeResources.
Two of the most common types of flash devices are defined by their respective technologies: NOR and NAND. NOR-based flash is the older technology that supported high read performance at the expense of smaller capacities. NAND flash offers higher capacities with significantly faster write and erase performance. NAND also requires a much more complicated input/output (I/O) interface.
Flash parts are commonly divided into partitions, which allows multiple operations to occur simultaneously (erasing one partition while reading from another). Partitions are further divided into blocks (commonly 64KB or 128KB in size). Firmware that uses the partitions can further apply unique segmenting to the blocks—for example, 512-byte segments within a block, not including metadata.
Flash devices exhibit a common constraint that requires device management when compared to other storage devices such as RAM disks. The only Write operation permitted on a flash memory device is to change a bit from a one to a zero. If the reverse operation is needed, then the block must be erased (to reset all bits to the one state). This means that other valid data within the block must be moved for it to persist. NOR flash memory can typically be programmed a byte at a time, whereas NAND flash memory must be programmed in multi-byte bursts (typically, 512 bytes).
The process of erasing a block differs between the two memory types. Each requires a special Erase operation that covers an entire block of the flash memory. NOR technology requires a precursor step to clear all values to zero before the Erase operation can begin. An Erase is a special operation with the flash device and can be time-consuming. Erasing is an electrical operation that drains the electrons from each cell in an entire block.
NOR flash devices typically require seconds for the Erase operation, whereas a NAND device can erase in milliseconds. A key characteristic of flash devices is the number of Erase operations that can be performed. In a NOR device, each block in the flash memory can be erased up to 100,000 times. NAND flash memories can be erased up to one million times.


Back to top
In addition to and as a result of the constraints explored in the previous section, managing flash devices presents several challenges. The three most important are garbage collection, managing bad blocks, and wear leveling.
Garbage collection is the process of reclaiming invalid blocks (those that contain some amount of invalid data). Reclamation involves moving the valid data to a new block, and then erasing the invalid block to make it available. This process is commonly done in the background or as needed, if the file system is low on available space.
Over time, flash devices can develop bad blocks through use and can even ship from the manufacturer with blocks that are bad and cannot be used. You can detect the presence of back blocks from a failed flash operation (such as an Erase) or an invalid Write operation (discovered through an invalid Error Correction Code, or ECC).
After bad blocks have been identified, they are marked within the flash itself in a bad block table. How this is done is device-dependent but can be implemented with a separate set of reserved blocks managed separately from normal data blocks. The process of handling bad blocks—whether they ship with the device or appear over time—is called bad block management. In some cases, this functionality is implemented in hardware by an internal microcontroller and is therefore transparent to the upper-level file system.
Recall that flash devices are consumable parts: You can perform a finite number of Erase cycles on each block before the block becomes bad (and must therefore be tagged by bad block management). To maximize the life of the flash, wear-leveling algorithms are provided. Wear leveling comes in two varieties: dynamic wear leveling and static wear leveling.
Dynamic wear leveling addresses the problem of a limited number of Erase cycles for a given block. Rather than randomly using blocks as they are available, dynamic wear-leveling algorithms attempt to evenly distribute the use of blocks so that each gets uniform use. Static wear-leveling algorithms address an even more interesting problem. In addition to a maximum number of Erase cycles, certain flash devices suffer from a maximum number of Read cycles between Erase cycles. This means that if data sits for too long in a block and is read too many times, the data can dissipate and result in data loss. Static wear-leveling algorithms address this by periodically moving stale data to new blocks.


Back to top
So far, I've explored flash devices and their fundamental challenges. Now, look at how these pieces come together as part of a layered architecture (see Figure 1). At the top is the virtual file system (VFS), which presents a common interface to higher-level applications. The VFS is followed by the flash file system, which will be covered in the next section. Next is the Flash Translation Layer (FTL), which provides for overall management of the flash device, including allocation of blocks from the underlying flash device as well as address translation, dynamic wear leveling, and garbage collection. In some flash devices, a portion of the FTL can be implemented in hardware.

The Linux kernel uses the Memory Technology Device (MTD) interface, which is a generic interface for flash devices. The MTD can automatically detect the width of the flash device bus and the number of devices necessary for implementing the bus width.


Back to top
Several flash file systems are available for Linux. The next sections explain the design and advantages of each.
One of the earliest flash file systems for Linux is called the Journaling Flash File System. JFFS is a log-structured file system that was designed for NOR flash devices. It was unique and addressed a variety of problems with flash devices, but it created another.
JFFS viewed the flash device as a circular log of blocks. Data written to the flash is written to the tail, and blocks at the head are reclaimed. The space between the tail and head is free space; when this space becomes low, the garbage collector is executed. The garbage collector moves valid blocks to the tail of the log, skips invalid or obsolete blocks, and erases them (see Figure 2). The result is a file system that is automatically wear leveled both statically and dynamically. The fundamental problem with this architecture is that the flash device is erased too often (instead of an optimal erase strategy), which wears the device out too quickly.

When a JFFS is mounted, the structural details are read into memory, which can be slow at mount-time and consume more memory than desired.
Although JFFS was very useful in its time, its wear-leveling algorithm tended to shorten the life of NOR flash devices. The result was a redesign of the underlying algorithm to remove the circular log. The JFFS2 algorithm was designed for NAND flash devices and also includes improved performance with compression.
In JFFS2, each block in the flash is treated independently. JFFS2 maintains block lists to sufficiently wear-level the device. The clean list represents blocks on the device that are full of valid nodes. The dirty list contains blocks with at least one obsoleted node. Finally, the free list represents the blocks that have been erased and are available for use.
The garbage collection algorithm can then intelligently decide what to reclaim in a reasonable way. Currently, the algorithm probabilistically selects from the clean or dirty list. The dirty list is selected 99 percent of the time to reclaim blocks (moving the valid contents to another block), and the clean list is selected 1 percent of the time (simply moving the contents to a new block). In both cases, the selected block is erased and placed on the free list (see Figure 3). This allows the garbage collector to re-use blocks that are obsoleted (or partially so) but still move data around the flash to support static wear leveling.

YAFFS is another flash file system developed for NAND flash. The initial version (YAFFS) supported flash devices with 512-byte pages, but the newer version (YAFFS2) supports newer devices with larger page sizes and greater Write constraints.
In most flash file systems, obsolete blocks are marked as such, but YAFFS2 additionally marks blocks with monotonically increasing sequence numbers. When the file system is scanned at mount time, the valid inodes can be quickly identified. YAFFS also maintains trees in RAM to represent the block structure of the flash device, including fast mounting through checkpointing —the process of saving the RAM tree structure to the flash device on a normal unmount so that it can be quickly read and restored to RAM at mount time (see Figure 4). Mount-time performance is a great advantage of YAFFS2 over other flash file systems.



Back to top
In some embedded systems, there's no need to provide a mutable file system: An immutable one will suffice. Linux supports a variety of read-only file systems, two of the most useful are cramfs and SquashFS.
The cramfs file system is a compressed read-only Linux file system that can exist within flash devices. The primary characteristics of cramfs are that it is both simple and space-efficient. This file system is used in small-footprint embedded designs.
While cramfs metadata is not compressed, cramfs uses zlib compression on a per-page basis to allow random page access (pages are decompressed upon access).
You can play with cramfs using the mkcramfs utility and the loopback device.
SquashFS is another compressed read-only Linux file system that is useful within flash devices. You'll also find SquashFS in numerous Live CD Linux distributions. In addition to supporting zlib for compression, SquashFS uses Lembel-Ziv-Markov chain Algorithm (LZMA) for improved compression and speed.
Like cramfs, you can use SquashFS on a standard Linux system with mksquashfs and the loopback device.


Back to top
Like most of open source, software continues to evolve, and new flash file systems are under development. An interesting alternative still in development is LogFS, which includes some very novel ideas. For example, LogFS maintains a tree structure on the flash device itself so that the mount times are similar to traditional file systems, such as ext2. It also uses a wandering tree for garbage collection (a form of B+tree). What makes LogFS particularly interesting, however, is that it is very scalable and can support large flash parts.
With the growing popularity of flash file systems, you'll see a considerable amount of research being applied toward them. LogFS is one example, but other options, such as UbiFS, are also growing. Flash file systems are interesting architecturally and will continue to be a source of innovation in the future.
Learn
See Wikipedia'soverview of flash memory technologies andlist of file systems, including disk-based file systems, distributed file systems, and special-purpose file systems (including solid-state media file systems).
Read more aboutNAND vs. NOR flash technology.
Tim's "Anatomy of the Linux file system" (developerWorks, Oct 2007) introduces the VFS, its major structures, and its architecture. It also provides an introduction to file systems and illustrates how Linux can support so many file systems concurrently.
Read howJFFS and its successor, JFFS2 (PDF file) take different approaches to flash device management.
CompareYAFFS to other popular flash file systems.
LogFS andUbiFS are two new flash file systems that solve many problems associated with current flash file systems.
TheMTD is a generic subsystem for memory devices, such as flash parts. It provides a generic interface between lower-level hardware drivers and the upper layers of the file system.
Cramfs andSquashFS are read-only compressed file systems for Linux. Both were designed for memory-constrained embedded systems, but you'll also find SquashFS used in Live CD distributions. Sometimes, you'll find these used withUnionFS, which implements a union mount for a file system (overlaying one file system on another). SquashFS has the advantage of better compression and performance usingLZMA and aset of patches.
Readall of Tim's Anatomy of... articles on developerWorks.
Readall of Tim's Linux articles on developerWorks.
In thedeveloperWorks Linux zone, find more resources for Linux developers, and scan ourmost popular articles and tutorials.
See allLinux tips andLinux tutorials on developerWorks.
Stay current withdeveloperWorks technical events and Webcasts.
Get products and technologies
Order the SEK for Linux, a two-DVD set containing the latest IBM trial software for Linux from DB2®, Lotus®, Rational®, Tivoli®, and WebSphere®.
WithIBM trial software, available for download directly from developerWorks, build your next development project on Linux.
Discuss
Get involved in thedeveloperWorks community through blogs, forums, podcasts, and community topics in ournew developerWorks spaces.



M. Tim Jones is an embedded software architect and the author of GNU/Linux Application Programming, AI Application Programming, and BSD Sockets Programming from a Multilanguage Perspective. His engineering background ranges from the development of kernels for geosynchronous spacecraft to embedded systems architecture and networking protocols development. Tim is a Consultant Engineer for Emulex Corp. in Longmont, Colorado.