How Does DRAM Work? Cells, Refresh, Rows and Banks

A DRAM cell stores a bit as charge on a small capacitor behind one access transistor. Reading it connects the capacitor to a much larger bitline, which moves only a little, so a sense amplifier has to detect and restore it; and because the charge leaks away, every row must be refreshed every few tens of milliseconds. Cells are arranged in rows and banks, and whether the row you want is already open decides most of the latency you see.

What you'll learn

  • The 1T1C cell and why a read is destructive
  • Charge sharing, with numbers
  • Refresh: how often, and what it costs
  • Rows, the row buffer and banks
  • DDR4 timing in nanoseconds: hit, closed row and conflict

Step 1: the 1T1C cell

Each cell is one access transistor and one storage capacitor. The transistor's gate is the wordline, shared by a row; its drain is the bitline, shared by a column. A charged capacitor is a 1 and an empty one a 0. One transistor per bit, against six in an SRAM cell, is why DRAM gives far more bits per chip, and why it is used for main memory.

Step 2: reading, by charge sharing

The bitline is precharged to VDD/2. When the wordline rises, the cell capacitor C_s and the bitline capacitance C_bl share their charge and settle to one voltage. The bitline moves by ΔV = (V_cell − VDD/2) × C_s / (C_s + C_bl). Because C_bl is several times C_s, the change is small, and its sign says whether the cell held a 1 or a 0.

QuantityValueHow
Cell capacitance C_s25 fFExample value
Bitline capacitance C_bl100 fFExample value
VDD, precharge1.2 V, 0.6 VBitline starts at VDD/2
Change for a stored 1+120 mV(1.2 − 0.6) × 25 / 125
Change for a stored 0−120 mV(0 − 0.6) × 25 / 125

After charge sharing the cell no longer holds its full value: the read has destroyed it. The sense amplifier, a cross-coupled latch on each bitline pair, amplifies the ±120 mV to a full rail, and because the wordline is still on, that full level is written back into the cell. Every read is therefore a read and a restore.

Step 3: refresh

The capacitor leaks through the transistor and the junction, so a stored 1 decays. The memory controller must refresh every row within the retention period, by reading it into the sense amplifiers and writing it back. For DDR4 the JEDEC standard sets that period at 64 ms at normal temperatures, covered by 8,192 refresh commands, so one refresh is due every 64 ms / 8,192 = 7.8125 µs (tREFI, quoted as 7.8 µs).

During a refresh the device is busy for tRFC, which grows with density: 350 ns for an 8 Gb DDR4 device. The share of time lost is 350 ns / 7.8125 µs = 4.48%. Above 85 °C leakage is faster, the refresh interval halves and the overhead doubles.

Step 4: rows, the row buffer and banks

A device is split into banks, each its own array of rows. Accessing data takes up to three commands. ACTIVATE opens a row: it reads the whole row (thousands of bits) into that bank's sense amplifiers, which then act as a row buffer. READ or WRITE then moves a burst of data from the open row. PRECHARGE closes the row, restoring the bitlines to VDD/2 before another row in that bank can open. Different banks can have different rows open, so the controller overlaps work across banks.

DRAM read from a closed bankACTIVATE opens the row; after tRCD a READ is issued; after CL the data burst appears; PRECHARGE closes the row afterwards. Not to scale: each delay is 22 clocks in DDR4-3200AA.tRCDCLclkcommandACTRDPREdataD0D1
Opening a row, reading it and closing it. Not to scale: tRCD and CL are 22 clocks each in DDR4-3200AA.

Step 5: what that means for latency

DDR4-3200 transfers 3,200 million times a second on a 1,600 MHz clock, so one clock is 0.625 ns. The JEDEC DDR4-3200AA speed bin has CL = tRCD = tRP = 22 clocks, 13.75 ns each.

CaseCommandsLatency
Row hit (row already open)READCL = 13.75 ns
Row closed (bank precharged)ACTIVATE, READtRCD + CL = 27.5 ns
Row conflict (another row open)PRECHARGE, ACTIVATE, READtRP + tRCD + CL = 41.25 ns

A conflict costs three times a hit, which is why controllers reorder requests to hit open rows and why access patterns that walk along a row are fast. These figures are the device's part only; the controller and interconnect add their own delay. Peak bandwidth on a 64-bit channel is 3,200 MT/s × 8 bytes = 25.6 GB/s, reached only with long runs of row hits.

Common mistakes

  • Thinking a DRAM read leaves the cell unchanged. It is destructive and the sense amplifier restores it.
  • Quoting CL alone as DRAM latency. A row miss adds tRCD and possibly tRP.
  • Mixing clocks and transfers. DDR4-3200 is 3,200 MT/s on a 1,600 MHz clock, so CL22 is 13.75 ns, not 6.9 ns.
  • Forgetting refresh overhead, which grows with density and temperature.
  • Assuming SRAM and DRAM differ only in speed. DRAM needs refresh and is far denser.

Check your understanding

  • Q1. With C_s = 30 fF, C_bl = 120 fF and VDD = 1.2 V, how far does the bitline move for a stored 1? Answer: 0.6 × 30 / 150 = 120 mV.
  • Q2. Why is a DRAM read destructive? Answer: charge sharing leaves the capacitor at nearly VDD/2, so the value has to be written back.
  • Q3. What does a row conflict cost in DDR4-3200AA? Answer: tRP + tRCD + CL = 41.25 ns.

Related

Sources

Last verified

Charge sharing, refresh and timing figures are computed in the site's tests (lib/memoryNumbers.test.js) on 16 November 2026. Timing values are from the JEDEC DDR4 standard; a specific device's datasheet may differ.