electrical charge in a storage cell consisting of a capacitor
and a
transistor. Due to leakage the capacitor discharges
gradually and the memory cell loses the information.
Therefore, to preserve the information, the memory has to be
refreshed periodically. Despite this inconvenience, the DRAM
is a very popular memory technology because of its high
density and consequent low price.
The first commercially available DRAM chip was the
Intel1103, introduced in 1970.
The early DRAM chips up to a 16k x 1 (16384 locations of one
bit each) model needed 3 supply voltages (+5V, -5V and +12V).
included on-chip to create the necessary supply voltages out
of a single +5V supply. This was necessary to fit the device
into a 16-pin
DIL package, which was the preferred package
at the time, and also made them easier to use.
To reduce the pin count, thereby helping miniaturisation,
DRAMs generally had a single data line which meant that a
computer with an N bit wide
data bus needed a "bank" of (at
least) N DRAM chips. In a bank, the address and control
signals of all chips were common and the data line of each
chip was connected to one of the data bus lines.
Beginning with the 256 kilobit DRAM, a tendency toward
data line appeared (e.g. 64k x 4), reducing the number of
chips per bank. This trend has continued and DRAM chips with
up to 36 data lines are available today. Furthermore,
together with surface mount packages, memory manufacturers
began to offer memory modules, where a bank of memory chips
Single Inline Pin Module, SIMM = Single Inline Memory Module,
DIMM = Dual Inline Memory Module). Today, this is the
DRAM bit cells are arranged on a chip in a grid of rows and
columns where the number of rows and columns are usually a
power of two. Often, but not always, the number of rows and
columns is the same. A one megabit device would then have
1024 x 1024 memory cells. A single memory cell can be
selected by a 10-bit row address and a 10-bit column address.
To access a memory cell, one entire row of cells is selected
and its contents are transferred into an on-chip buffer. This
discharges the storage capacitors in the bit cells. The
desired bits are then read or written in the buffer. The
(possibly altered) information is finally written back into
the selected row, thereby refreshing all bits (recharging the
capacitors) in the row.
To prevent data loss, all bit cells in the memory need to be
refreshed periodically. This can be done by reading all rows
in regular intervals. Most DRAMs since 1970 have been
specified such that one of the rows needs to be refreshed at
least every 15.625 microseconds. For a device with 1024 rows,
a complete refresh of all rows would then take up to 16 ms; in
other words, each cell is guaranteed to hold the data for 16
ms without refresh. Devices with more rows have accordingly
longer retention times.
Many varieties of DRAM exist today. They differ in the way
they are interfaced to the system - the structure of the
memory cell itself is essentially the same.
"Traditional" DRAMs have multiplexed address lines and
separate data inputs and outputs. There are three control
signals: RAS (row address strobe), CAS (column address
strobe), and WE (write enable) (the backslash indicates an
active low signal). Memory access procedes as follows:
1. The control signals initially all being inactive (high), a
memory cycle is started with the row address applied to the
address inputs and a falling edge of RAS . This latches the
row address and "opens" the row, transferring the data in the
row to the buffer. The row address can then be removed from
the address inputs since it is latched on-chip. 2. With RAS
still active, the column address is applied to the address
pins and CAS is made active as well. This selects the
desired bit or bits in the row which subsequently appear at
the data output(s). By additionally activating WE the data
applied to the data inputs can be written into the selected
location in the buffer. 3. Deactivating CAS disables the
data input and output again. 4. Deactivating RAS causes the
data in the buffer to be written back into the memory array.
Certain timing rules must be obeyed to guarantee reliable
operation. 1. RAS must remain inactivate for a while before
the next memory cycle is started to provide sufficient time
for the storage capacitors to charge (Precharge Time). 2. It
takes some time from the falling edge of the RAS or CAS
signals until the data appears at the data output. This is
specified as the Row Access Time and the Column Access Time.
Current DRAM's have Row Access Times of 50-100 ns and Column
Access Times of 15-40 ns. Speed grades usually refer to the
former, more important figure.
Note that the Memory Cycle Time, which is the minimum time
from the beginning of one access to the beginning of the next,
is longer than the Row Access Time (because of the Precharge
Time).
Multiplexing the address pins saves pins on the chip, but
usually requires additional logic in the system to properly
generate the address and control signals, not to mention
further logic for refresh. Therefore, DRAM chips are usually
preferred when (because of the required memory size) the
additional cost for the control logic is outweighed by the
lower price.
Based on these principles, chip designers have developed many
varieties to improve performance or ease system integration of
DRAMs:
PSRAMs (Pseudo Static Random Access Memory) are essentially
controller. This saves some system logic and makes the device
look like a normal
SRAM. This has been popular as a lower
complete SRAM substitute because it is sometimes busy when
doing self-refresh, which can be tedious.
line by clocking the CAS line.
accesses to the currently open row can be made while the RAS
signal is kept active.
Static Column DRAM is similar to Page Mode DRAM, but to access
different bits in the open row, only the column address needs
to be changed while the CAS signal stays active. The row
buffer essentially behaves like SRAM.
data from one address while setting up a new address, for use
DRAM used for Video RAM (
VRAM) has an additional long
shift register that can be loaded from the row buffer. The
shift register can be regarded as a second interface to the
memory that can be operated in parallel to the normal
CRT displays. These frame buffers generate a serial data
stream that is sent to the CRT to modulate the electron beam.
By using the shift register in the VRAM to generate this
stream, the memory is available to the computer through the
normal interface most of the time for updating the display
data, thereby speeding up display data manipulations.
SDRAM (Synchronous DRAM) adds a separate clock signal to the
the chip and high speed "burst" accesses that clock a series
of successive bits out (similar to the nibble mode).
CDRAM (Cached DRAM) adds a separate static RAM array used for
caching. It essentially combines main memory and
cachememory in a single chip. The cache memory controller needs to
be added externally.
RDRAM (Rambus DRAM) changes the system interface of DRAM
completely. A byte-wide bus is used for address, data and
command transfers. The bus operates at very high speed: 500
million transfers per second. The chip operates synchronously
with a 250MHz clock. Data is transferred at both rising and
falling edges of the clock. A system with signals at such
frequencies must be very carefully designed, and the signals
on the Rambus Channel use nonstandard signal levels, making it
incompatible with standard system logic. These disadvantages
are compensated by a very fast data transfer, especially for
burst accesses to a block of successive locations.
A number of different refresh modes can be included in some of
the above device varieties:
RAS only refresh: a row is refreshed by an ordinary read
access without asserting CAS. The data output remains
disabled.
CAS before RAS refresh: the device has a built-in counter
for the refresh row address. By activating CAS before
activating RAS, this counter is selected to supply the row
address instead of the address inputs.
Self-Refresh: The device is able to generate refresh cycles
internally. No external control signal transitions other than
those for bringing the device into self-refresh mode are
needed to maintain data integrity.
(1996-07-11)