 |
Description  |
|
|
FIELD OF THE INVENTION
This present invention relates generally to digital memory devices and
systems, and more particularly to power-saving modes for such devices and
systems.
BACKGROUND OF THE INVENTION
Microprocessors, digital signal processors, digital imaging devices, and
many other types of digital data processing devices rely on an attached
high-speed memory system to hold data and/or processor instructions needed
by the processing device. FIG. 1 depicts a typical memory system
configuration 20. A host processor 22 issues data store and retrieve
requests to a memory controller 24 over a front-side bus FSB. Memory
controller 24 acts as an intermediary for the exchange of data between
processor 22 and memory devices 26A and 26B. The memory controller
performs memory transfers by relating the processor's memory requests
(into its view of a contiguous memory space) to the individual memory
devices' activation, addressing, timing, and bus signaling requirements.
The memory controller's techniques for partitioning and accessing the
memory devices may be highly optimized in order to avoid wasted bus
cycles; the controller may also perform memory management functions such
as device initialization, refresh for memories that require refresh, etc.
Various circumstances can exist to cause any particular memory device's
duty cycle (i.e., the percent of the time that the device is actually
performing memory operations) to be relatively low, either for brief
periods (e.g., a microsecond) or for extended periods. For instance, the
processor may be sleeping, lightly loaded, or successfully hitting an
intermediate cache (and therefore not needing to access the main memory).
Normally, memory devices are maintained in a ready state that allows them
to respond to a new request from a memory controller within a few clock
cycles. In order to reduce power consumption, e.g., for battery-powered or
energy-conserving devices, many existing memory devices offer a
"power-down" mode that deactivates input and output buffers and other
clocked circuitry, without loss of data. During a time when a memory
device is idle, the memory controller can instruct that device to enter
the power-down mode, thereby saving a significant fraction of the device's
ready-state power.
BRIEF DESCRIPTION OF THE DRAWINGS
The invention may be best understood by reading the disclosure with
reference to the drawing, wherein:
FIG. 1 illustrates a typical memory system configuration;
FIG. 2 shows a block diagram for a prior art SDRAM with power-down
capability;
FIG. 3 contains a timing diagram for a read, power-down, and power-up
sequence for the SDRAM of FIG. 2;
FIG. 4 contains a block diagram for an SDRAM with early power-down
capability according to an embodiment of the invention;
FIG. 5 contains a block diagram for the power-down logic used in the device
of FIG. 4;
FIG. 6 contains a timing diagram for a read, early power-down, and power-up
sequence for the SDRAM of FIG. 4;
FIG. 7 contains a second timing diagram for a read-with-auto-precharge,
early power-down, and power-up sequence for the SDRAM of FIG. 4;
FIG. 8 illustrates a block diagram for an SDRAM with early power-down
capability according to a second embodiment of the invention;
FIG. 9 contains a block diagram for the power-down logic used in the device
of FIG. 8; and
FIG. 10 contains a timing diagram for a read, early power-down, and
power-up sequence for the SDRAM of FIG. 8.
DETAILED DESCRIPTION OF THE EMBODIMENTS
Although existing power-down modes are useful, what apparently has been
heretofore unrecognized is that further increases in power savings are
possible, with little added memory device complexity. Embodiments
described herein allow a memory controller to issue a power-down command
to a memory device much earlier, e.g., while a read or write command is
still in process. For instance, a controller may request an
eight-clock-cycle burst read from a device. A prior art controller has to
wait until the memory device sets up the read, transfers the data to the
controller over eight bus cycles, and possibly executes a precharge
command to release the active word line--all before the controller can
issue a power-down command. But in many of the embodiments described
below, no waiting is required. Instead, the controller can issue the
power-down command immediately after (or concurrently with) the last read,
write, or other active-mode command. The memory device then shuts off some
unneeded components, such as address/command registers, immediately. As
other components finish their tasks, they can be shut down as well. The
net result can be additional power savings, and possibly reduced
complexity in the controller, which no longer has to worry about when it
should issue a power-down command in a wide variety of circumstances. In
some embodiments, it becomes feasible to power-down for even very short
time intervals, as power savings occur during an active command, and
power-up can be accomplished quickly.
Several terms have been assigned particular meanings within the context of
this disclosure. A memory device is a solid-state electrical component
containing addressable memory space that can be used to store and later
retrieve digital data. In this respect, a memory device can be an
integrated circuit, a packaged circuit, a rank of such circuits, a memory
module containing one or more ranks, etc. The term multi-component memory
device recognizes that such devices can usually be divided functionally
into circuitry to accomplish various memory device tasks, e.g., control,
external bus communication, refresh, row selection, column selection, data
storage, etc. The types, speeds, sizes, or numbers of memory devices
selected for use with the present invention are not critical, although
these parameters may impact how an early power-down strategy is employed.
Some possible device types for which the present invention is useful
include synchronous dynamic random access memory (SDRAM) devices
(including double-data-rate (DDR) SDRAM devices and quad-data-rate (QDR)
SDRAM devices), static RAM, and flash memory devices.
An active-mode command is a command that normally leaves a memory device at
full readiness (although in some embodiments of the invention, an
active-mode command can include a power-down-upon-completion option). Some
examples of active mode commands are a command to load a register,
activate a word line, read from memory, write to memory, precharge bit
lines, terminate a memory burst, auto-refresh, etc. A power-down command
is an instruction to a memory device to revert to a state of incomplete
readiness that consumes less power than the device consumes at full
readiness. Some examples of power-down commands include a standby mode
command, sleep mode command, self-refresh command, etc. An early
power-down command is issued while an active-mode command is pending or
executing, i.e., at a time when an immediate power-down would leave an
active-mode command unfinished.
As an introduction to the embodiments, and to provide a point of contrast
with the present invention, a prior-art SDRAM device 26A with prior-art
power-down capability is presented in FIGS. 2 and 3.
FIG. 2 contains a block diagram for a prior-art SDRAM device. Device 26A
maintains several external connections to a controlling device. Command
register 50 receives L-bit-wide commands over a command bus CMD. Address
register 52 receives N-bit-wide row addresses, M/W-bit-wide column
addresses, or N-bit-wide mode register data, depending on the contents of
command register 50, over an address bus ADD. Two registers connect to
W-bit-wide data bus DQ. Data output register 54 is capable of driving read
data onto bus DQ when output enable OUT EN is asserted; otherwise,
register 54 remains in a high-impedance state. Data input register 56 is
capable of receiving write data from bus DQ. Delay-locked-loop (DLL)
circuit 66 receives an external clock signal CLK and generates a stable
internal clock CK based on CLK. Clock enable (CKE) register 58 receives an
external clock enable signal CKE. The output of register 58 is combined
with the output of DLL 66 to provide the gated internal clock signal CK
(the operation of CKE and CK will be described further in connection with
FIG. 3). Note that additional external connections (e.g., power, ground,
bank select, data strobes, data masks) have not been illustrated, although
such connections exist in many/all devices.
Control logic 60 coordinates the overall operation of the memory device.
Command decoder 62 interprets the contents of command register 50 at each
clock cycle; if the contents require a memory operation, control logic 60
executes that operation, in conjunction with configuration data stored in
mode register 64. Execution of an operation generally involves proper
sequencing and timing of operations in other device components controlled
by control logic 60.
Memory array 70 contains 2.sup.N.times.2.sup.M addressable cells, each cell
lying at the intersection of one of 2.sup.N rows and one of 2.sup.M
columns (additional redundant cells may also exist for the purpose of
replacing defective cells, not shown). Cells along each row share a common
word line; cells along each column share a common bit line. Each cell
stores one bit of information as a charge on a capacitor. Cell information
can be "read" by charge-sense amplifiers 72 that connect to the bit lines
(shown here as one sense amp/bit line, although some embodiments use less
sense amps, each multiplexed to multiple bit lines). Block 72 also
contains write driver circuitry to write data to an active row of array
70.
Row access circuitry comprises row latch and decoder 80, row address
multiplexer 82, and refresh counter 84. When control logic 60 signals
latch/decoder 80, latch/decoder 80 reads in an N-bit row address from
multiplexer 82 and activates the word line corresponding to that row
address. Multiplexer 82 selects the row address from one of two sources:
an address latched into address register 52 during an activate word line
command; or an address supplied by refresh counter 84, for auto-refresh or
self-refresh operation.
Control logic 60, in "self-refresh" mode, or an external controller, via
"auto-refresh" commands, executes periodic refresh cycles, using refresh
counter addressing, to maintain data integrity. Like all DRAM devices,
SDRAM 26A requires regular refresh (read/writeback) of the cells on each
word line to prevent data loss due to capacitor discharge. Refresh counter
74 cycles through all word line addresses in a repeating cycle, one
address each time control logic 60 strobes the counter.
Data access circuitry comprises column address counter/latch 90, column
decoder 92, and page register 94. Page register 94 holds 2.sup.M bits of
data read from a row of memory array 70, or up to 2.sup.M bits of data to
be written to a row of memory array 70. Page register 94 connects to the
data input and output registers by a W-bit-wide bus 96. Column decoder 92
connects data bus 96 to an appropriate set of W register bits in page
register 94 for each memory operation. Column decoder 92 determines the
appropriate set of register bits by decoding a column address supplied by
column address counter/latch 90. When the memory device receives a read or
write command, it also receives a starting column address in address
register 52. This address is latched into column address counter/latch 90
by control logic 60; if the read or write is a "burst" operation, i.e.,
multiple W-bit data bus transactions will occur as part of the read or
write, counter/latch 90 increments the starting column address for each
subsequent bus transaction.
FIG. 3 illustrates, for the device of FIG. 2, a four-word burst read
operation, followed by a power-down, a power-up, and the initiation of a
four-word burst write to the same row used for the burst read. New memory
instructions and addresses are clocked into registers 50 and 52 on the
rising edges of CK, i.e., at T1, T2, . . . , T11. Thus at T1, a "READ4"
command and address "COL a" are registered to the memory device. At
T2-T10, "NOP" (No Operation) commands are received (a NOP could include a
command to another memory device, which would be received but ignored by
this memory device). At T11, a "WRITE4" command and address "COL b" are
registered to the memory device.
The address "COL a" latched into address register 52 at T1 is latched into
column address counter/latch 90 after the command decoder 62 determines
that this is a READ command. The column decoder then selects the data word
at "COL a" of the page register for output, and that data word is
subsequently clocked into data output register 54. At about T3, OUT EN is
asserted and the word is driven onto DQ as "Dout a". Meanwhile, column
address counter/latch 80 has been incremented to address column a+1, and
the data word from that column of the page register is selected to bus 96.
Prior to T4, OUT EN is deasserted, the new word is clocked into data
output register 54, and OUT EN is reasserted at about T4--driving "Dout
a+1" onto DQ. This process repeats for the data words at columns a+2 and
a+3, with the fourth and last word of the burst driven onto DQ at T6.
Prior to T7, CKE is deasserted. At T7, the deasserted CKE is clocked into
CKE register 58. This forces the CK output of AND gate 68 low, such that
CK no longer follows CLK (i.e., CK is "suspended"). The suspension of CK
freezes the device's synchronous logic in its current state. As that logic
is typically CMOS-based, the logic requires very little current while the
clock is frozen, thus placing the device in a low-power standby mode. In
this mode, changes in CMD, ADD, and DQ are not clocked into their
respective input registers and no refresh operations are performed.
Prior to T10, CKE is reasserted to "wake up" the circuit. CK once again
goes high at T10, and the device is ready to receive a next command, in
this case a "WRITE4" command, once clock cycle later at T11.
From FIGS. 2 and 3, it can be appreciated that CK cannot be suspended prior
to the completion of the READ4 command, as it is needed through clock
cycle T6 to output data. Were CKE deasserted sooner, data loss would
occur, and the data output register may be frozen in a state that drives
the data bus continually.
It is recognized herein that during the majority of the read operation of
FIG. 3, the only elements of the memory device involved in that operation
are the data access circuitry (blocks 90, 92, and 94), the data output
register, and the portion of the control logic that controls those blocks.
And yet the device continues to operate all of its synchronous logic, and
continues to read commands and addresses every clock cycle during the
burst read--looking for a next command--and not knowing that the memory
controller intends to sleep the device as soon as the current read
operation ends. Meanwhile, the memory controller must wait until the read
operation completes and then issue a separate command to put the device to
sleep.
The present disclosure describes embodiments that can employ different
power-down behavior than the prior art. It is recognized herein that the
use of a large multipurpose clock tree within a memory device can waste
power unnecessarily. Within the control logic of a memory device, multiple
clock signals can be formed to selectively power-down components of the
device at different times, based on an understanding of when during an
active-mode command each component is used. Using such capability, the
prior-art constraint that a power-down command must be issued only to an
idle memory device can be removed. The result is that a memory device can
respond to an early power-down command by shutting down the part of the
memory device that is not necessary to complete a pending active-mode
command.
FIG. 4 illustrates one embodiment of the invention, SDRAM 100, in block
diagram form. Control logic 110 contains a power-down logic block 120 that
receives the external clock enable signal CKE and the clock reference CK.
Power-down logic 120 generates three separately controllable clock signals
CK1, CK2, and CK3, based on CK and the internal state of control logic
110.
FIG. 5 shows one possible implementation for power-down logic 120,
consisting of two OR gates 121, 122 and three AND gates 123, 124, and 125.
OR gate 121 receives inputs CKE and RAP (Row Access Pending), and asserts
its output whenever either of these signals is high. RAP is asserted by
control logic 110 whenever command decoder 62 receives a row access
command, e.g., an activate row command, precharge command, or refresh
command. RAP is deasserted when that command completes.
OR gate 122 receives inputs CKE and CAP (Column Access Pending), and
asserts its output whenever either of these signals is high. CAP is
asserted by control logic 110 whenever command decoder 62 receives a
column access command, e.g., a read, write, or burst-terminate command.
CAP is deasserted when that command completes.
AND gate 123 forms clock signal CK1 by ANDing the output of OR gate 121
with CK--thus CK1 follows CK whenever CKE and/or RAP is asserted, and is
low otherwise. AND gate 124 forms clock signal CK2 by ANDing the output of
OR gate 122 with CK--thus CK2 follows CK whenever CKE and/or CAP is
asserted, and is low otherwise. AND gate 125 forms clock signal CK3 by
ANDing CKE with CK--thus CK3 follows CK whenever CKE is asserted, and is
low otherwise.
Clock distribution for the clock signals in device 100 is partitioned
according to memory device component function. Clock signal CK serves
power-down logic 120. Clock signal CK1 serves row access control logic
114, i.e., the portion of control logic 110 that controls the row access
circuitry (blocks 72, 80, 82, and 84). Clock signal CK2 serves cell access
control logic 116, i.e., the portion of control logic 110 that controls
the cell access circuitry (blocks 90, 92, and 94); clock signal CK2 also
serves the two registers that share the DQ bus. Clock signal CK3 serves
command register 50 and address register 52.
FIGS. 6 and 7 contain illustrative timing diagrams for the embodiment shown
in FIGS. 4 and 5. FIG. 6 illustrates the same read and write commands and
read/write command timing as shown in FIG. 3. But in FIG. 6, the memory
controller takes CKE low prior to T2, signaling the memory device that the
controller will consider the device to be on standby after it completes
the pending read operation.
At T1, the device decodes the READ4 command and enters a state sequence to
perform the requested burst read operation. When the device enters this
state sequence, control logic 110 asserts CAP to power-down logic 120.
Since the READ4 command does not require row-access states, RAP remains
low.
Prior to T2, memory controller takes CKE low, thus blocking CK1 and CK3
from transitioning at T2. As CAP is high prior to the deassertion of CKE,
CK2 is unaffected by the CKE transition. Thus those synchronous components
of the device that attach to the C2 clock tree continue to operate, while
those components of the device that attach to the CK1 and CK3 clock trees
freeze at T2.
At T6, cell access control logic 116 asserts OUT EN to data output register
54, which drives Dout a+3 onto DQ. After the preset bus drive time is
passed, cell access control logic 116 deasserts OUT EN, and the control
logic has completed the last state in the requested burst read.
Consequently, CAP goes low, such that at T7, CK2 is prevented from
transitioning. At this point, only DLL 66 and power-down logic 120
continue clocked operation, while the remainder of the circuit remains
frozen.
Prior to T11, the memory controller transitions CKE high again and places a
new command/address on the CMD and ADD buses. CKE high re-enables CK1,
CK2, and CK3, such that all components are ready at T11 when the WRITE4
command is received.
Comparing FIG. 6 to FIG. 3, most of device 100 sleeps for nine clock
cycles, while none of device 26A sleeps for more than three clock cycles.
Also, with device 100 it becomes possible to sleep part of the device even
when the controller does not intend to have the device sleep between
commands--the sleeping components can merely be "woken up" when the next
command is issued.
FIG. 7 shows a two-word burst read followed by an auto-precharge
instruction at T1, and activation of a new row b at T8. Here, RAP remains
active three clock cycles longer than CAP, keeping CK1 enabled to allow
the device to perform the auto-precharge operation. At T8, activation of
the new array row b requires the row access logic, but not the column
access logic--thus CKE can transition low again prior to T9, stopping CK2
and CK3 immediately while CK1 remains active until the activation command
completes.
An embodiment of the present invention need not use the clock enable signal
to signal an early power-down. FIG. 8 illustrates an SDRAM 200 that uses
an address register bit as an early power-down option to an active-mode
command. This embodiment frees the memory controller from having to issue
a separate power-down command following each active-mode command. It also
allows CKE to be used for other purposes, and allows multiple memory
device ranks to be powered down independently, even though those ranks
share a common CKE connection.
On many memory devices, the row address width, N, is greater than the
column address width M/W (where M is the number of columns in the cell
array and W is the width of the data bus). Since address register 52 must
be at least N bits wide, one or more extra address lines may be available
when a command is not accompanied by a row address. Device 200 uses one of
these address lines, An, as a power-down option bit.
FIG. 9 shows a block diagram for power-down logic 220 of device 200. Bit An
from address register 52 forms the input to register 226. An is registered
into 226 only upon a Latch Column Address (LCA) signal, which could be the
same signal used to latch a column address into column address
counter/latch 90. If An is set high when register 226 receives LCA, Clock
Enable (CE)--which has a similar use as CKE in FIG. 5--remains high and
all clocks run. If An is unset at this time, however, CE is deasserted,
CK3 is disabled, and CK1 and CK2 are controlled by RAP and CAP as
previously illustrated. Thus the memory controller can utilize otherwise
unused addressing bits to control a device's power-down behavior.
One problem with using a registered bit to power-down is that power-down
shuts off the registers, making it impossible to signal a power-up in the
same fashion as a power-down. Device 200 avoids this issue by running the
Chip Select (CS) bit of the CMD bus through a different register, CS REG
206, which runs off the external clock CLK instead of CK3. Since every
command addressed to device 200 must assert CS, the CS bit can be used to
"wake up" the device when a valid command is received. FIG. 9 uses
register 227, AND gate 228, and inverter 229 to perform a wake-up
operation.
Register 227 uses CK as a clock. Every clock cycle, register 227 copies and
holds the current value of CE. The output of register 227 passes through
inverter 229 to form one input to AND gate 228. Thus whenever CE
transitions low (due to an issued power-down command), one input to AND
gate 228 will be asserted on the next clock cycle--and every clock cycle
thereafter as long as CE remains low. In this condition, the output of AND
gate 228 is asserted whenever CS is asserted, i.e., when a new command is
received. AND gate 228 is tied to SET on register 226, an input that sets
CE high when asserted. This re-enables the clocks CK1, CK2, and CK3,
allowing the remainder of the command, address, and possibly data to be
clocked in. This behavior is shown in the timing diagram of FIG. 10, where
the READ4 command issued at T1 contains a standby option "STBY" attached
to the column address, and the WRITE2 command issued at T11 does not.
One other feature of device 200, which may or may not be selected for
combination with the features described above, is the capability to have a
different power-down function that responds to CKE. In FIG. 8, CKE passes
through CKE register 204 to form one input to AND gate 202. The other
input to AND gate 202 is CLK, and the AND gate output is the input to DLL
66. Thus CLK is prevented from reaching DLL 66, or any circuitry depending
on CK, when CKE is deasserted. This feature could be used with a more
comprehensive (but not early) power-down mode that allows CLK to go away
for a time period, and then be re-established prior to assertion of CKE,
and finally CS.
Many other variations are possible. More than one bit of the address
register could be used to signal different power-down options along with a
command. For instance, a short-term power down mode could leave more
circuitry active than a long-term power-down mode, in order to respond to
a wake-up signal in less time.
Another variable is the number of different clock signals controlled by the
power-down logic. In a simple embodiment, the number of clock signals
could be just two (e.g., one for the command/address registers and one for
everything else). In other embodiments, the number of separate clock trees
could be more than three. For instance, the data input register and data
output register clock signals could be separately controlled, such that
the clock to the output register can be turned off during write commands,
and the clock to the input register can be turned off during read
commands.
The control of more than one input clock signal is also possible. Many
faster memory systems use source-synchronous data-bus clocking, i.e., the
source of a data word also drives one or more data strobe signals DQS on
the bus when a data word is driven. Power-down logic can gate incoming DQS
signals such that those signals only affect the data input register when
the device has not been issued a power-down command. Further, when a
power-down command has been issued while a read command is pending, the
power down logic can gate DQS so that the device's outgoing DQS does not
affect its own input registers.
Although the power-down logic is illustrated in the figures in a single
block, this logic could be distributed within the components themselves,
and each component could receive a single power-down flag signal that it
would respond to according to its own state. It is preferable, however, to
have the logic concentrated near the source of the clock tree so that less
clock buffers are driven when in power-down.
One of ordinary skill in the art will recognize that the concepts taught
herein can be tailored to a particular application in many other
advantageous ways. In particular, those skilled in the art will recognize
that the illustrated embodiments are but one of many alternative
implementations that will become apparent upon reading this disclosure.
Whether a particular signal is asserted when high or low is a design
issue, but either can be made to work with the present invention. Various
buffers not shown in the drawing may be used to adjust signal timing,
reduce loading on a gate or register output, etc. Only a limited set of
exemplary memory devices has been presented--many devices could use a
banked array architecture, may not need some disclosed components (e.g.,
refresh circuitry), etc. A memory controller may not be a dedicated unit,
but may be encompassed within, e.g., a host processor. Such minor
modifications are encompassed within the invention, and are intended to
fall within the scope of the claims.
The preceding embodiments are exemplary. Although the specification may
refer to "an", "one", "another", or "some" embodiment(s) in several
locations, this does not necessarily mean that each such reference is to
the same embodiment(s), or that the feature only applies to a single
embodiment.
* * * * *
|
|
 |
|
 |
Description |
|
