DRAM and Flash Cells
Store a bit as charge. DRAM keeps it on a tiny capacitor that leaks. Flash keeps it on a floating gate that does not.
You Will Be Able To
- Describe how a 1T1C DRAM cell is written and why the word line is boosted.
- Compute the bit-line signal of a DRAM read from the two capacitances.
- Explain why a DRAM read is destructive and why cells must be refreshed.
- Describe how a floating-gate flash cell stores a bit as a threshold shift.
- Lesson
- G9.17
- Time
- 30 min
- Runs on
- ngspice
- Check
- 5 questions
A DRAM cell is one transistor and a capacitor of about 20 fF, on a much larger bit line. In this run a stored 1 moved a 100 fF bit line by +35 mV. A stored 0 moved it by −142 mV: that is the whole signal.
Writing a 1T1C Cell
A DRAM cell is an access NMOS and a storage capacitor. The word line opens the transistor; the bit line then charges or empties the capacitor.
| Write | Word line | Stored, word line on | Stored after the word line falls |
|---|---|---|---|
| 1 | 1.8 V | 1.082 V | 1.059 V |
| 0 | 1.8 V | 0 mV | |
| 1 | 2.6 V (boosted) | 1.734 V |
Measured: the lab, 20 fF, access NMOS W 1 µm, word line on for 4 ns.
The 1 stops a threshold below the word line, as in lesson G9.12. DRAMs drive the word line above VDD, to about VDD + Vt, so that the 1 is stored nearly in full. With 2.6 V it stored 1.734 V here. This run overdrives a 1.8 V logic transistor to show the effect. Real DRAM access devices have a thicker gate oxide, made for that voltage.
Reading by Charge Sharing
For a read the bit line is precharged to VDD/2 and left floating. When the word line opens, the cell and the bit line settle at one voltage. Charge is conserved, so the bit line moves by (V_cell − V_DD/2) × C_s / (C_s + C_bl).
| Bit line | Stored | Measured ΔV | Formula | Cell after the read |
|---|---|---|---|---|
| 100 fF | 1.06 V | +35 mV | +26 mV | 0.94 V |
| 100 fF | 0.00 V | −142 mV | −150 mV | 0.76 V |
| 250 fF | 1.06 V | +16 mV | +12 mV | 0.92 V |
| 250 fF | 0.00 V | −64 mV | −67 mV | 0.84 V |
| 150 fF | 1.06 V | +25 mV | +19 mV | 0.93 V |
| 150 fF | 0.00 V | −101 mV | −106 mV | 0.80 V |
| 150 fF | 1.80 V | +113 mV | +106 mV | 1.01 V |
Measured: 20 fF cell, bit lines precharged to 0.9 V, read 19.5 ns after the word line rises. The 1.8 V row is the level a boosted write leaves.
The formula fits every row within a few millivolts, and each measured row sits a few millivolts above it. The rising word line couples a little charge in through the access transistor. The share takes time. For the boosted 1, the bit line had moved only +78 mV 3.5 ns after the word line rose. The access NMOS conducts slowly: its source is the bit line near 1 V, and its gate is only at 1.8 V. DRAMs boost the word line for reads too.
The read leaves the cell at the shared voltage, and the data is gone. The sense amplifier must drive the bit line back to the full level while the word line is still open. That write-back is part of every DRAM read.
Leakage and Refresh
Charge leaks away through the off access transistor and the capacitor itself. The cell must be read and written back, refreshed, before its level drifts out of the sense amplifier's reach. Commodity DDR DRAMs refresh every row every 64 ms; the example cell in KL needs it every 2 to 4 ms.
The worst case for a stored 1 is a bit line held at 0 V while another row of the column is written. The off access NMOS then has the full cell voltage across it. A short run measures that current, and t = C_s × ΔV / I_leak turns it into a retention time.
| Temperature | Leakage of the off access NMOS | Time to lose 100 mV from 20 fF |
|---|---|---|
| 27 °C | 0.11 pA | 17.7 ms |
| 85 °C | 3.5 pA | 569 µs |
Measured: a 1 µm by 0.15 µm SKY130 logic NMOS, gate at 0 V, cell held at 1.059 V, bit line at 0 V. The models have no junction or gate leakage here, so real cells leak more.
Subthreshold leakage grows fast with temperature: 31 times from 27 °C to 85 °C here. A logic transistor at 85 °C would need a refresh well under a millisecond. That is why DRAM processes build their access transistors for low leakage.
Flash: Charge on a Floating Gate
- Control / FSM
- Storage
- Device
- I/O
- Structure
- Control
- Erased
The floating gate holds little charge. The transistor turns on at a low control-gate voltage.
- Programmed
Electrons are forced onto the floating gate, by hot-carrier injection or by tunnelling through the thin oxide. Their charge raises the threshold seen from the control gate.
- Reading
A read voltage between the two thresholds is put on the control gate. An erased cell conducts and a programmed one does not.
Read the diagram as text
- Control gate: the word line
- Floating gate: surrounded by oxide
- Channel: between source and drain
- Source and drain: the read current flows between them
- Control gate → Floating gate: couples through oxide
- Floating gate → Channel: thin tunnel oxide
- Channel → Source and drain: conducts or not
The oxide around the floating gate holds the charge for years with no power, so flash needs no refresh.
A floating gate can be modelled as a capacitive divider. The control gate couples to the floating gate through C_CG, and the floating gate is the gate of an ordinary NMOS. Stored electrons are a charge Q_FG on that isolated node. Seen from the control gate, they move the threshold by ΔV_T = −Q_FG / C_CG.
| Cell | Floating gate with the control gate at 0 V | Threshold seen from the control gate |
|---|---|---|
| erased | 0 V | 0.91 V |
| programmed | −1 V | 2.26 V |
Measured. C_CG is 3 fF onto the gate of an NMOS of 1 µm by 0.15 µm, with the drain at 0.1 V. Threshold: a drain current of 1 µA.
The control gate moves the floating gate by only 0.77 V per volt here. The rest of the swing is lost across C_CG. So a 1 V shift on the floating gate looks like 1 / 0.77 = 1.30 V from outside. The run gives 1.35 V. A read voltage between the two thresholds, about 1.6 V here, makes an erased cell conduct and a programmed one stay off.
NOR flash connects cells in parallel on a bit line, for fast random reads. NAND flash connects them in series strings, for density. Lessons G8.4 and G8.5 cover the arrays.
Asked in interviews
Why is a DRAM read destructive?
The cell shares its charge with the much larger bit line and ends at a level near the precharge. The sense amplifier must write the full level back while the word line is still open.
How big is the DRAM read signal?
ΔV = (V_cell − V_DD/2) × C_s / (C_s + C_bl). With a 20 fF cell and a few hundred femtofarads of bit line it is tens to about a hundred millivolts.
How does a flash cell store a bit?
As charge on a floating gate, which shifts the threshold seen from the control gate. A read voltage between the erased and programmed thresholds tells them apart.
Try It
v1 is the stored 1 with the word line still on: 1.082 V. v1end is the same after the word line falls: 1.059 V.
vbend is the stored 1 after a write with the boosted word line: 1.734 V.
Change the boost from 2.6 V to 2.2 V in VWB and run. The stored 1 falls a little.
Run the netlist to draw its plots here.
v1 = 1.081854e+00 v0 = -4.162489e-11 v1end = 1.058907e+00 vbend = 1.734123e+00Check Yourself
Common Mistakes
Wrong Expect a written 1 to reach the full supply.
The access NMOS passes a weak 1. With the word line at 1.8 V the cell reached 1.082 V, and 1.059 V after the word line fell.
Fix Boost the word line a threshold above the supply. At 2.6 V the cell stored 1.734 V.
Wrong Expect the bit line to swing like a logic signal on a read.
The 20 fF cell shares its charge with a much larger bit line. With 100 fF the bit line moved only +35 mV for a 1.
Fix Precharge the bit line to VDD/2 and sense the small difference with a sense amplifier.
Wrong Treat a read as harmless to the cell.
After the read the cell holds the shared voltage, not its old one. It was left at 0.935 V for a stored 1 and 0.758 V for a stored 0.
Fix Write the sensed value back on every read. The sense amplifier does it by driving the bit line to the full level.
Wrong Size the signal against a precharge of 0 V or VDD.
The precharge sits halfway, so a 1 moves the line up and a 0 moves it down. The signal is set by the stored level minus VDD/2, not minus 0.
Fix Use ΔV = (V_cell − V_DD/2) × C_s / (C_s + C_bl).
Remember
- A 1T1C cell: an access NMOS and a capacitor. A boosted word line stores a full 1.
- Read by charge sharing: ΔV = (V_cell − V_DD/2) × C_s / (C_s + C_bl), tens of millivolts.
- A read is destructive; the sense amplifier writes the value back.
- Charge leaks, so DRAM is refreshed periodically.
- Flash stores charge on a floating gate, which shifts the threshold; it keeps data without power.