U.S. patent application number 11/474076 was filed with the patent office on 2007-01-18 for method and circuit for configuring memory core integrated circuit dies with memory interface integrated circuit dies.
Invention is credited to Suresh N. Rajan.
Application Number | 20070014168 11/474076 |
Document ID | / |
Family ID | 37595841 |
Filed Date | 2007-01-18 |
United States Patent
Application |
20070014168 |
Kind Code |
A1 |
Rajan; Suresh N. |
January 18, 2007 |
Method and circuit for configuring memory core integrated circuit
dies with memory interface integrated circuit dies
Abstract
A memory device comprises a first and second integrated circuit
dies. The first integrated circuit die comprises a memory core as
well as a first interface circuit. The first interface circuit
permits full access to the memory cells (e.g., reading, writing,
activating, pre-charging and refreshing operations to the memory
cells). The second integrated circuit die comprises a second
interface that interfaces the memory core, via the first interface
circuit, an external bus, such as a synchronous interface to an
external bus. A technique combines memory core integrated circuit
dies with interface integrated circuit dies to configure a memory
device. A speed test on the memory core integrated circuit dies is
conducted, and the interface integrated circuit die is electrically
coupled to the memory core integrated circuit die based on the
speed of the memory core integrated circuit die.
Inventors: |
Rajan; Suresh N.; (San Jose,
CA) |
Correspondence
Address: |
STATTLER JOHANSEN & ADELI LLP
540 UNIVERSITY AVENUE
SUITE 350
PALO ALTO
CA
94301
US
|
Family ID: |
37595841 |
Appl. No.: |
11/474076 |
Filed: |
June 23, 2006 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
60693631 |
Jun 24, 2005 |
|
|
|
Current U.S.
Class: |
365/201 |
Current CPC
Class: |
G11C 11/401 20130101;
G11C 29/24 20130101; G11C 11/4076 20130101; G11C 29/022 20130101;
G11C 7/1012 20130101; G11C 8/12 20130101; G11C 7/1045 20130101;
G11C 29/50012 20130101; G11C 7/10 20130101; G11C 11/4093 20130101;
G11C 29/02 20130101; G11C 29/50 20130101; G11C 29/028 20130101;
G11C 2207/2254 20130101 |
Class at
Publication: |
365/201 |
International
Class: |
G11C 29/00 20060101
G11C029/00 |
Claims
1. A method for configuring a memory device, said method
comprising: conducting at least one test on a plurality of memory
core integrated circuit dies; identifying at least one
characteristic of said memory core integrated circuit dies from
said test; and electrically coupling an interface integrated
circuit die to said memory core integrated circuit die based on
said memory core integrated circuit characteristic.
2. A method as set forth in claim 1, wherein identifying at least
one characteristic of said memory core integrated circuit die from
said test comprises identifying a speed of operation for said
memory core integrated circuit die.
3. A method as set forth in claim 2, wherein: identifying a speed
of operation for said memory core integrated circuit comprises
identifying at least one memory core integrated circuit die that
operates at a slow speed relative to operation of other memory core
integrated circuit dies; and electrically coupling an interface
integrated circuit die to said memory core integrated circuit die
comprises coupling an interface integrated circuit die, designed
for low power markets, to said memory core integrated circuit die
that operates at a relatively slow speed.
4. A method as set forth in claim 2, wherein: identifying a speed
of operation for said memory core integrated circuit comprises
identifying at least one memory core integrated circuit die that
operates at a high speed relative to operation of other memory core
integrated circuit dies; and electrically coupling an interface
integrated circuit die to said memory core integrated circuit die
comprises coupling an interface integrated circuit die, designed
for high performance markets, to said memory core integrated
circuit die that operates at a high speed.
5. A method as set forth in claim 2, wherein: identifying a speed
of operation for said memory core integrated circuit comprises
identifying at least one memory core integrated circuit die that
operates at an average speed relative to operation of other memory
core integrated circuit dies; and electrically coupling an
interface integrated circuit die to said memory core integrated
circuit die comprises coupling an interface integrated circuit die,
designed for power conservation and performance markets, to said
memory core integrated circuit die that operates at an average
speed.
6. The method as set forth in claim 1, wherein conducting at least
one test on a plurality of memory core integrated circuit dies
comprises conducting a speed test to measure the amount of time
required for a read, write, activate, pre-charge, or refresh
operations.
7. The method as set forth in claim 1, wherein: conducting at least
one test on a plurality of memory core integrated circuit dies
comprises conducting a test to determine if a memory core
integrated circuit die comprises defective rows or columns;
identifying at least one characteristic of said memory core
integrated circuit dies from said test comprises identifying at
least one non-defective memory core integrated circuit die that
does not contain any defective rows or columns; and electrically
coupling an interface integrated circuit die to said memory core
integrated circuit die comprises electrically coupling an interface
integrated circuit die, designed for high performance markets, to
said non-defective memory core integrated circuit die.
8. A method for configuring a memory device, said method
comprising: conducting at least one test on a plurality of memory
core integrated circuit dies to characterize said memory core
integrated circuit dies based on speed of operation; configuring an
effective data rate for an interface on said memory core integrated
circuit die based on said speed of operation of said memory core
integrated circuit die; and electrically coupling an interface
integrated circuit die to said memory core integrated circuit die
to operate with an external bus and to operate in conjunction with
said effective data rate of said interface on said memory core
integrated circuit die.
9. A method as set forth in claim 8, wherein: conducting at least
one test on a plurality of memory core integrated circuit dies to
characterize said memory core integrated circuit dies comprises
characterizing at least one memory core integrated circuit die as a
slow memory core integrated circuit die; configuring an effective
data rate for an interface on said memory core integrated circuit
die comprises configuring a relatively high data rate for an
interface on said memory core integrated circuit die; and
electrically coupling an interface integrated circuit die to said
memory core integrated circuit die comprises electrically coupling
an interface integrated circuit die that operates at said
relatively high data rate.
10. A method as set forth in claim 8, wherein: conducting at least
one test on a plurality of memory core integrated circuit dies to
characterize said memory core integrated circuit dies comprises
characterizing at least one memory core integrated circuit die as a
slow memory core integrated circuit die; configuring an effective
data rate for an interface on said memory core integrated circuit
die comprises configuring a relatively medium speed data rate for
an interface on said memory core integrated circuit die; and
electrically coupling an interface integrated circuit die to said
memory core integrated circuit die comprises electrically coupling
an interface integrated circuit die to said external bus that
operates at said medium speed data rate.
11. A memory device comprising: first integrated circuit die
comprising: memory core comprising a plurality of memory cells,
said memory cell comprising an access time for executing operations
on said memory cells; first interface circuit coupled to said
memory cells, comprising an internal data rate, corresponding to
said access time, for transferring data between said memory cells
and said first interface circuit; and second integrated circuit
die, electrically coupled to said first integrated circuit die,
comprising a second interface comprising an external data rate for
accessing data, at a data rate from said first interface circuit
compatible with said internal data rate of said first interface
circuit, and for interfacing said data to an external circuit.
12. A memory device as set forth in claim 11, wherein said access
time of said memory cells comprises a relatively slow access time,
said internal data rate comprises a relatively high data rate, and
said external data rate comprises a relatively high data rate.
13. A memory device as set forth in claim 11, wherein said access
time of said memory cells comprises a relatively slow access time,
said internal data rate comprises a an average data rate, and said
external data rate comprises an average data rate.
Description
RELATED APPLICATIONS
[0001] This patent application claims the benefit to U.S.
Provisional Patent Application entitled "Methods and Apparatus for
Integrating Multi-Chip Memory Devices," Ser. No. 60/693,631, filed
on Jun. 24, 2005.
BACKGROUND OF THE INVENTION
[0002] 1. Field of the Invention
[0003] The present invention is directed toward the field of
building custom memory systems cost-effectively for a wide range of
markets.
[0004] 2. Art Background
[0005] Dynamic Random Access Memory (DRAM) is the most popular type
of volatile memory and is widely used in a number of different
markets. The popularity of DRAMs is mostly due to their
cost-effectiveness (Mb/$). The PC main memory market has
traditionally been the largest consumer of DRAMs. However, in
recent times, other important markets have adopted DRAMs. A report
published by De Dios and Associates showed that in 2004, the PC
main memory market consumed only 50% of the total DRAM bits.
[0006] Several of the non-PC markets use specialty or legacy
memories. Specialty memory is typically memory that is not used by
the PC main memory but is memory that is designed for one or more
niche markets. For example, the PC graphics market uses GDDR
(Graphics Dual Data Rate) DRAM. Similarly, some segments of the
network infrastructure market use FCRAM (Fast Cycle RAM) or RLDRAM
(Reduced Latency DRAM). Legacy memory is typically memory that was
used in the past but is not used in that particular market segment
now. For example, SDRAM (Synchronous DRAM) was used for PC main
memory from .about.1997 to .about.2001 but is no longer used today
for PC main memory. Instead, most cellular phones and handheld (or
mobile) devices use SDRAM today.
[0007] Bringing a new DRAM architecture into the market requires
significant investment of time and money. For example, it typically
takes 4 years for JEDEC to approve a new DRAM architecture. DRAM
makers must then spend hundreds of millions of dollars to
productize the new architecture. Unless the investment is amortized
over an extremely large number of devices, the cost of the new
devices will be high. In addition, the DRAM makers have optimized
their manufacturing flow for high volumes. Any deviation from the
norm disrupts the flow. This is the reason why specialty and legacy
memory typically carry a price premium over memory used by the PC
main memory market (which is usually referred to as commodity
memory).
[0008] Given the time and money required to bring a new DRAM
architecture into the market, it is obvious that the industry does
not have the luxury of being able to define a DRAM architecture
that exclusively meets the needs of the smaller markets for DRAMs.
For example, it is difficult for the DRAM makers to
cost-effectively produce a DRAM that perfectly meets the needs of
the cell phone market. Therefore, it is even more difficult for
each cell phone maker (e.g. Nokia or Motorola) to design DRAMs
tailor made for its phones. So, cell phone designers are forced to
choose the DRAM architecture that is least objectionable from their
perspective. Hence the selection of SDRAM for cell phones.
[0009] This situation will become even worse in the future. Most
analyst projections show that not only will DRAM usage expand
rapidly into newer markets but also that the DRAM bit consumption
growth rate will be higher for non-PC markets. The needs of these
markets are quite different from the needs of the PC main memory
market. Clearly, there is a need in the market for a way to quickly
and cost-effectively build custom memory that is tailor made for a
customer's exact needs.
SUMMARY OF THE INVENTION
[0010] A method combines memory core integrated circuit dies with
interface integrated circuit dies to configure a memory device. At
least one test on the memory core integrated circuit dies is
conducted. In one embodiment, the test measures the amount of time
required for a read, write, activate, pre-charge, or refresh
operations. From the test, at least one characteristic of the
memory core integrated circuit dies is identified. In one
embodiment, the characteristic comprises the speed of operation for
the memory core integrated circuit die. The interface integrated
circuit die is electrically coupled to the memory core integrated
circuit die based on the memory core integrated circuit
characteristic.
[0011] In one embodiment, the memory core integrated circuit dies
are speed binned based on "slow", "typical" and "high" speed. The
slow memory core integrated circuit dies are coupled to an
interface integrated circuit die designed for low power markets.
The typical speed memory core integrated circuit dies are coupled
to an interface integrated circuit die designed for power
conservation and performance markets. Also, the high-speed memory
core integrated circuit dies are coupled to an interface integrated
circuit die designed for high performance markets.
[0012] In other embodiments, an effective data rate for an
interface on the memory core integrated circuit die is configured
based on the speed of operation of the memory core integrated
circuit die. An interface integrated circuit die is electrically
coupled to the memory core integrated circuit die to operate with
an external bus and to operate in conjunction with the effective
data rate of the interface on the memory core integrated circuit
die.
BRIEF DESCRIPTION OF THE DRAWINGS
[0013] FIG. 1 is a block diagram illustrating a conventional DRAM
chip.
[0014] FIG. 2 illustrates a typical organization of a 4-bank modern
SDRAM.
[0015] FIG. 3 is a block diagram illustrating one embodiment of
banks arranged in sub-arrays.
[0016] FIG. 4 illustrates a block diagram of an interface and
multiple banks in a DRAM.
[0017] FIG. 5 is a block diagram illustrating a DRAM chip with an
interface removed.
[0018] FIG. 6 illustrates one embodiment for a center bonded DRAM
core chip.
[0019] FIG. 7 illustrates one embodiment for an edge bonded DRAM
core chip.
[0020] FIG. 8 illustrates one embodiment for a concentrated-bank
architecture.
[0021] FIG. 9 illustrates one embodiment for a distributed-bank
architecture.
[0022] FIG. 10 illustrates one embodiment for a quadrant in a
distributed-bank architecture DRAM core chip.
[0023] FIG. 11 is block diagram illustrating one embodiment of a
distributed-bank architecture universal DRAM core chip.
[0024] FIG. 12 is a block diagram illustrating a distributed-bank
architecture DRAM core chip configured to support DDR2 speeds.
[0025] FIG. 13 is a block diagram illustrating a distributed-bank
architecture DRAM core chip configured to support DDR speeds and
external bus widths from 17 to 32 bits.
[0026] FIG. 14 is a block diagram illustrating a distributed-bank
architecture DRAM core chip configured to support DDR speeds and
external bus widths from 9 to 16 bits.
[0027] FIG. 15 is a block diagram illustrating a distributed-bank
architecture DRAM core chip configured to support DDR speeds and
external bus widths from 1 to 8 bits.
[0028] FIG. 16 illustrates one embodiment for a portion of a DRAM
core chip that includes a decoder for selecting a mode of
operation.
[0029] FIG. 17a is a block diagram illustrating the relationship
between the internal data bus rate and the external data bus rate
for 4n pre-fetching.
[0030] FIG. 17b is a block diagram illustrating the relationship
between the internal data bus rate and the external data bus rate
for burst mode with a length of 4n.
[0031] FIG. 18 is a block diagram illustrating an example
multi-chip memory implementation.
[0032] FIG. 19 illustrates techniques for stacking two DRAM core
chips behind a single interface chip.
DETAILED DESCRIPTION
[0033] The disclosure of U.S. Provisional Patent Application Ser.
No. 60/693,631, entitled "Methods and Apparatus for Integrating
Multi-Chip Memory Devices", filed on Jun. 24, 2005, is hereby
expressly incorporated herein by reference.
[0034] By examining several different DRAM architectures, it is
clear that the internal organizations of the DRAMs are quite
similar. For example, a .times.16 (.times.16 denotes the external
data width) 256 Mb SDRAM, .times.16 256 Mb DDR SDRAM, and .times.16
256 Mb DDR2 SDRAM consist of:
[0035] Control logic block;
[0036] Address input register and decoder;
[0037] Memory arrays that store the data;
[0038] Data selection circuit (I/O gating);
[0039] Data read circuit; and
[0040] Data write circuit.
[0041] It is obvious that most of the blocks are common across all
the three architectures. The main differences are in the control
logic block (which implements the protocol, among other functions),
in the width of the data that is accessed per column address, and
in the data I/O section. These are usually considered part of the
interface section of the DRAM while the rest of the circuits
(address decoder, memory arrays, and data selection) are considered
part of the memory core. The core timing parameters are typically
specified in absolute units of time (seconds) rather than in terms
of clock periods. For example, the Micron 256 Mb DDR2 SDRAM data
sheet lists the following core timing parameters: TABLE-US-00001
TABLE 1 Speed Grade t.sub.RCD (ns) t.sub.RP (ns) t.sub.RC (ns) -5E
15 15 55 -37E 15 15 55 -3 15 15 55 -3E 12 12 54
[0042] Similarly, the Micron 256 Mb DDR SDRAM data sheet identifies
the following timing specifications: TABLE-US-00002 TABLE 2 Speed
Grade t.sub.RCD (ns) t.sub.RP (ns) t.sub.RC (ns) -75Z 20 20 65 -75
20 20 65
[0043] The Micron 256 Mb SDRAM data sheet discloses the following
specifications: TABLE-US-00003 TABLE 3 Speed Grade t.sub.RCD (ns)
t.sub.RP (ns) t.sub.RC (ns) -7E 15 15 60 -75 20 20 66
[0044] So, even though the protocol and speed of SDRAM, DDR SDRAM,
and DDR2 SDRAM are quite different, it is clear that the internal
core or array of all these types of DRAMs has similar
characteristics. In fact, we can go even further and observe that
all synchronous DRAMs are composed of an asynchronous core and an
interface that defines the protocol, synchronous operation, speed,
and signaling. The memory core typically comprises .about.90%-95%
of the total die area.
[0045] Current practice is to integrate the memory core and the
interface onto a common die. The drawback with this approach is
that a change in the protocol, speed, or signaling for example
requires a re-design of the entire chip. This is usually very
expensive and time consuming, and hence the inability to bring
specialty or custom DRAMs to the market quickly and
cost-effectively. One embodiment of the invention comprises a
multi-chip implementation, wherein one or more DRAM core chips are
attached to an interface chip. The interface chip sits between the
host electronic system and the DRAM core chips. In other words, the
interface chip can be thought of as a "wrapper" that surrounds the
DRAM core chips. The partitioning of the conventional DRAM into
DRAM core chip and interface chip should preferably be done is such
a way that that the functions and circuits that are relatively
constant across many different architectures are retained in the
DRAM core chip while the functions and circuits that vary between
the different architectures are moved to the interface chip.
[0046] The DRAM core chip can be designed to be suitable for a
large number of markets (i.e. a "universal core"). The interface
chip can now be designed to meet the exact needs of a market, and
even the exact needs of individual customers in that market. To
illustrate, the proposed solution enables the design of an
interface chip to meet the exact needs of Nokia for the cell phone
market and another interface chip to meet the exact needs of
Motorola for the cell phone market.
[0047] In order to accommodate the needs of the different markets,
the DRAM core chip must be capable of operating across a wide range
of frequencies, be capable of supporting high data rates, and must
be low cost. In one embodiment, the DRAM core chip is asynchronous,
wide, and operates at its natural speed. For the case of modern
DRAM cores, the natural speed is between 5 ns to 10 ns per column
access, which is equivalent to 100 MHz to 200 MHz synchronous
operation. That is, a modern DRAM core can keep up with an external
memory bus or interface that runs at a speed from 100 MHz to 200
MHz. So, for the case of a synchronous DRAM that operates at 100
MHz to 200 MHz and is n-bits wide (1.ltoreq.n.ltoreq.32 typically),
n bits can be fetched from the DRAM core once every clock cycle. In
fact, this is how SDRAMs operate.
[0048] Newer synchronous DRAMs run at higher clock speeds. JEDEC
defines the DDR SDRAM specification with external data rates of 200
MHz, 266 MHz, 333 MHz, and 400 MHz. An even newer specification
called DDR2 SDRAM has been defined with external data rates of 400
MHz, 533 MHz, 667 MHz, and 800 MHz. Effort is currently underway in
JEDEC to define a DDR3 SDRAM specification that spans data rates
from 800 MHz to 1600 MHz. GDDR, GDDR2, and GDDR3 SDRAMs typically
run faster than the DDR, DDR2, and DDR3 SDRAMs. However, even
though the external data rate has been increasing quite rapidly,
the speed of the DRAM core has not kept pace. In order to bridge
the gap between the external data rate and the internal core speed,
the DRAM industry has adopted a technique called
"pre-fetching."
[0049] Pre-fetching involves accessing more bits than the external
data bus width on every column access. To illustrate, an n-bit wide
DDR SDRAM accesses 2n bits every column access. This allows the
external data bus to run at 200 MHz to 400 MHz while the internal
memory core runs at 100 MHz to 200 MHz respectively. FIG. 1 is a
block diagram illustrating a conventional DRAM chip. A DRAM chip
100 comprises a DRAM core 110, Internal Data Bus 120, DRAM
interface 130 and External data bus 140. TABLE 4 shows the concept
of pre-fetching for a DRAM chip. TABLE-US-00004 TABLE 4 External
Internal Data Data External Internal Bus 140 Bus 120 Data Data
Protocol Width Width Rate (MHz) Rate (MHz) SDRAM N n 66-133 66-133
DDR SDRAM N 2n 200-400 100-200 DDR2 SDRAM n 4n 400-800 100-200 DDR3
SDRAM n 8n 800-1600 100-200 (proposed)
[0050] This implies that the universal DRAM core chip must be
sufficiently wide enough to support the data rates required by many
different markets. Obviously there is a limit to how wide the
universal DRAM core chip can be before it starts to negatively
impact the cost of the chip. In general, if the width of the DRAM
core chip is so large so as to make either the core chip or the
interface chip pad limited (especially the core chip), the cost of
this solution would be very high.
[0051] Modern DRAMs also feature multiple banks. A bank is a
section of the DRAM core that can be accessed independently. The
DRAM core is broken up into banks that can be active
simultaneously. Within each bank, only one row can be open at any
given time. Most DRAMs up to 512 Mb densities are organized into 4
banks. 1 Gb (and possibly, up to 4 Gb) DRAMs are organized into 8
banks but only 4 banks can be activated within a specific time
window. This is dictated by power and thermal considerations. So,
the universal DRAM core chip must be capable of supporting multiple
banks.
[0052] Let us consider the internal organization of a .times.16 256
Mb SDRAM. A .times.16 256 Mb SDRAM may have 4 banks, each of which
is 64 Mb. Each bank can be conceptualized as consisting of 16
sub-arrays, each sub-array being a 8192.times.512 matrix of memory
cells. That is, each sub-array has 8192 or 8k rows and 512 columns.
So, when a bank is accessed, a particular row is accessed
(activated) in each of the 16 sub-arrays in the bank. The row is
determined by the row address. After the 16 rows are activated, a
particular bit in each row is selected. The bit is specified by the
column address. So, on each access to a bank, 16 bits are
accessed.
[0053] FIG. 2 illustrates a typical organization of a 4-bank modern
SDRAM. The memory cells are arranged into four banks: bank 0 (220),
bank 1 (210), bank 2 (230) and bank 3 (240). Each bank contains
P.times.Q.times.16 cells (e.g., P=8192 and Q=512 for a 256 Mb
SDRAM). Each bank has associated word line drivers (275, 280, 285
and 290) and sense amplifiers (255, 260, 265 and 270). The banks
are selected through use of MUX 250.
[0054] In one embodiment, the banks are organized in sub-arrays.
FIG. 3 is a block diagram illustrating one embodiment of banks
arranged in sub-arrays. For this embodiment, each bank has 16
sub-arrays (each sub-array being 8K.times.512) because the DRAM is
organized as a .times.16 memory.
[0055] Consider the internal organization of a .times.16 256 Mb DDR
SDRAM. The .times.16 256 Mb DDR SDRAM is organized similar to the
.times.16 256 Mb SDRAM with some changes to the memory core. The
more important changes to the core organization are:
[0056] Each bank has 32 sub-arrays; and
[0057] Each sub-array is now 8192.times.256 matrix (i.e. P=8192,
Q=256). The reason for having 32 sub-arrays is that DDR SDRAM
memory uses a pre-fetching of 2n. Since this is a .times.16 DDR
memory, 32 bits must be accessed from each bank for a read or write
operation.
[0058] Note that pre-fetching can be done in a number of ways.
Consider a memory array that is organized as a P.times.Q matrix
that needs to support 2n pre-fetching. One approach is to divide
the P.times.Q array into two arrays (i.e. two P.times.Q/2 arrays)
and access both arrays in parallel, so that we get 2 bits per
column address. Another approach is to not split the array but
modify the column decoder so that 2 bits are selected for each
column address (in other words, the least significant bit of the
column address is not used). Some embodiments of the invention are
described that use the first approach. However, the teachings of
the present invention are applicable to different pre-fetching
implementations.
[0059] Looking at the organization of a .times.16 256 Mb DDR2
SDRAM, the .times.16 256 Mb DDR2 SDRAM is organized similar to the
.times.16 256 Mb SDRAM (and the .times.16 256 Mb DDR SDRAM). The
following identify some of the changes to the memory core:
[0060] Each bank has 64 sub-arrays. Each sub-array is now a
8192.times.128 matrix (i.e. P=8192, Q=128). The reason for the 64
sub-arrays per bank is that DDR2 SDRAM uses a 4n pre-fetching.
Since this is a .times.16 DDR2 memory (n=16), 64 bits must be
accessed from each bank for a read or write operation.
[0061] In all cases (SDRAM, DDR, DDR2), data bits to/from each bank
are brought to a multiplexer/de-multiplexer (hereafter referred to
as a MUX), which in turn is connected to the external DQ pins. This
MUX is typically in the middle of the DRAM chip. FIG. 4 illustrates
a block diagram of an interface and multiple banks in a DRAM. As
shown in FIG. 4, bank 0 (410), bank 1 (420), bank 2 (430) and bank
3 (440) are accessed by interface 450. Note that for SDRAM, m=n,
for DDR SDRAM, m=2n, DDR2 SDRAM, m=4n, and DDR3 SDRAM (proposed),
m=8n, Also note that the data MUX is typically part of the
interface.
[0062] This arrangement works well for a conventional DRAM since
everything is on a single die. However, in one embodiment of the
invention, the interface is on a separate die. If we were to just
move the interface alone (to another die) without disturbing the
memory core, then the number of I/O pads on both the memory core
chip and the interface chip will become quite large, as shown in
TABLE 5 below for a .times.16 4-bank implementation. FIG. 5 is a
block diagram illustrating a DRAM chip with an interface removed.
For this embodiment, bank 0 (510), bank 1 (520), bank 2 (530) and
bank 3 (540) are coupled to I/O pads 550, 560, 570 and 580,
respectively. If we look at only the data pins and ignored the
address, command, power and ground pins, we can see that the number
of data signals that have to go off-chip between the DRAM core chip
and the interface chip is 4m. TABLE-US-00005 TABLE 5 Total Number
of External Data Bus Off-Chip Data Protocol Width (n) m Pins (4m)
SDRAM 16 16 64 DDR SDRAM 16 32 128 DDR2 16 64 256 SDRAM DDR3 16 128
512 SDRAM
[0063] So, it is quite obvious that removing the interface without
disturbing the rest of the memory core quickly leads to a very
large number of off-chip connections, especially for wider external
data bus widths and higher data speeds (because, the amount of
pre-fetching will increase with higher data speeds). Under these
conditions, either the DRAM core chip or the interface chip or both
will become pad limited, which will increase the cost of the total
solution.
[0064] In one embodiment, in order to reduce the number of off-chip
connections between the DRAM core chip and the interface chip, some
part or all of the multiplexing of the data I/O from the banks is
done in the core chip itself.
[0065] One option is to route all the data bits from each bank to a
central MUX, and then connect the other side of the MUX to off-chip
drivers. This is quite similar to the current practice for center
bonded DRAMs. FIG. 6 illustrates one embodiment for a center bonded
DRAM core chip. For this example, integrated circuit 600 includes
bank 0 (610), bank 1 (620), bank 2 (630) and bank 3 (640) coupled
to MUX 650. MUX 650 is connected to substrate bonding pads 670 via
bond wires 680 through I/O pads 660.
[0066] The drawback with this approach is that bond wires 680 that
connect I/O pads 660 on the DRAM core die to the substrate bonding
pads 670 become quite long. Long bond wires have significant
inductance and limit the speed at which the memory chip can
operate.
[0067] In another embodiment, edge bonding for the core chip is
used. FIG. 7 illustrates one embodiment for an edge bonded DRAM
core chip. For this embodiment, integrated circuit 700 includes
bank 0 (710), bank 1 (720), bank 2 (730) and bank 3 (740). I/O pads
750, located on the silicon die 705, are connected to the substrate
bonding pads 760 via bond wires 780.
[0068] If the DRAM core was organized with one bank per quadrant,
then the number of data I/O pads will be equal to 4m as illustrated
previously. The other option is to route the data bits from each
bank to a centrally located MUX (as shown in FIG. 6) and then route
the signals from the other side of the MUX to the periphery of the
die. However, this means that the data signals will have to
traverse the die twice--once from the bank to the central MUX and
once from the central MUX to the periphery. This increases routing
complexity, may possibly require an extra metal layer on the die
(higher cost), and adds to the latency of the memory core.
[0069] In another embodiment, an inventive "Distributed-Bank"
architecture is used. In this architecture, a bank is distributed
(or spread) across all 4 quadrants instead of concentrating a bank
in only one quadrant. Using this architecture, data MUXs, located
in all 4 quadrants, select the appropriate bank, and the data
signals corresponding to the selected bank can be easily routed to
the periphery of the chip.
[0070] FIG. 8 illustrates one embodiment for a concentrated-bank
architecture. For the purpose of this illustration, a .times.16,
4-bank, 256 Mb DDR2 SDRAM core is used. However, any type of DRAM,
with different external data widths, different number of banks,
different density, and different amount of pre-fetching may be used
without deviating from the spirit or scope of the invention. As
previously shown, each bank (810, 820, 830 and 840) in a .times.16,
4-bank, 256 Mb DDR2 SDRAM consists of 64 sub-arrays, with each
sub-array organized as a 8192.times.128 array of memory cells.
[0071] FIG. 9 illustrates one embodiment for a distributed-bank
architecture. For this embodiment, the DRAM core chip is divided
into four quadrants (910, 920, 930 and 940). Each quadrant includes
a portion of a bank.
[0072] FIG. 10 illustrates one embodiment for a quadrant in a
distributed-bank architecture DRAM core chip. As discussed
previously, there are 64 sub-arrays per quadrant with each
sub-array being a 8192.times.128 matrix. Instead of assigning all
64 sub-arrays in a single quadrant to a single bank in the
concentrated-bank architecture, there are 16 sub-arrays to each of
the 4 banks within a single quadrant in the distributed-bank
architecture. In the distributed-bank architecture embodiment,
local data MUXs are located in each quadrant to select one of the
four banks.
[0073] FIG. 11 is block diagram illustrating one embodiment of a
distributed-bank architecture universal DRAM core chip. For this
embodiment, the banks of DRAM cells are distributed among quadrant
1110, 1120, 1130 and 1140. An address decoder 1150, located in the
center of the chip, controls word line drivers 1155, 1164, 1170 and
1176 in quadrants 1110, 1120, 1130 and 1140 respectively. The data
from the bank sub-arrays are output to sense amplifiers (1157,
1166, 1172 and 1178) and input to the respective bank select MUXs
(1160, 1168, 1174 and 1180). The data is then routed to data I/O
pads 1162 located in the proximity for each of the quadrants.
[0074] Since 64 data bits are accessed from the core chip for every
read or write operation, the interface chip may be designed or
configured to act similar to a .times.16 DDR2 SDRAM, .times.8 DDR2
SDRAM, .times.4 DDR2 SDRAM, .times.2 DDR2 SDRAM, or .times.1 DDR2
SDRAM. In fact, the interface chip may be designed to support any
data width between .times.1 and .times.16 when operating in a 4n
pre-fetch mode.
[0075] The distributed-bank architecture is flexible enough to
support protocols like SDRAM, DDR SDRAM, DDR2 SDRAM, and DDR3
SDRAM. For example, the DRAM core chip shown in FIG. 11 may be
configured or used as shown in FIG. 12 to support DDR2 speeds. FIG.
12 is a block diagram illustrating a distributed-bank architecture
DRAM core chip configured to support DDR2 speeds. MUX 1210 selects
64 bits of data for one of the banks (1220, 1230, 1240 and 1250).
MUX 1210 represents the data MUXs located in all four quadrants on
the DRAM core chip.
[0076] The DRAM core chip shown in FIG. 11 may also be configured
or used as shown in FIG. 13 to support DDR SDRAM speeds when it is
operated in a 2n pre-fetch mode. FIG. 13 is a block diagram
illustrating a distributed-bank architecture DRAM core chip
configured to support DDR speeds and external bus widths from 17 to
32 bits. MUX 1310 selects 64 bits of data for one of the banks
(1320, 1330, 1340 and 1350). The mode of operation shown in FIG. 13
may be used with the appropriate interface chip to support external
data widths between (and inclusive) of 17 and 32 in a 2n pre-fetch
mode. The same DRAM core chip may be used with the appropriate
interface chip to support external data widths between (and
inclusive) 9 and 16 in a 2n pre-fetch mode when operated as shown
in FIG. 14. FIG. 14 is a block diagram illustrating a
distributed-bank architecture DRAM core chip configured to support
DDR speeds and external bus widths from 9 to 16 bits. For this
embodiment, MUX 1410 selects 32 bits of data for one of the banks
(1420, 1430, 1440 and 1450).
[0077] Also, the same DRAM core chip can be used with the
appropriate interface chip to support external data widths between
(and inclusive of) 1 and 8 in a 2n pre-fetch mode. FIG. 15 is a
block diagram illustrating a distributed-bank architecture DRAM
core chip configured to support DDR speeds and external bus widths
from 1 to 8 bits. For this embodiment, MUX 1510 selects 16 bits of
data for one of the banks (1520, 1530, 1540 and 1550).
[0078] From these architectures, the internal data bus width (the
width of the bus between the DRAM core chip and the interface chip)
may be configured to match the amount of pre-fetching required
(which is determined by the external data rate) and the width of
the external data bus. The external data bus is the bus from the
interface chip to the ASIC or memory controller. The DRAM core chip
as shown in FIG. 11 may be configured to support the following
modes and requirements shown in TABLE 6. TABLE-US-00006 TABLE 6
Internal Data Pre- Minimum External Maximum External Bus Width
Fetching Used Data Bus Width Data Bus Width 64 8n 1 8 4n 1 16 2n 1
32 1n 1 64 32 8n 1 4 4n 1 8 2n 1 16 1n 1 32 16 8n 1 2 4n 1 4 2n 1 8
1n 1 16
[0079] Note that:
[0080] The proposed DDR3 SDRAM is an example of an 8n pre-fetch
protocol;
[0081] DDR2 SDRAM is an example of a 4n pre-fetch protocol;
[0082] DDR SDRAM is an example of a 2n pre-fetch protocol; and
[0083] SDRAM is an example of a in pre-fetch protocol (i.e. no
pre-fetching needed).
[0084] Again, for the DRAM core chip shown in FIG. 11, 3 modes of
operation may be defined and a 2-bit binary code may be assigned to
represent them as shown in TABLE 7. TABLE-US-00007 TABLE 7 Mode of
Operation (Mode[1:0]) Internal Data Bus Width 00 64 01 32 10 16 11
Undefined/Reserved
[0085] These two bits (Mode[1:0]) may be inputs to the DRAM core
chip so that the internal data bus width is selected through
external means. For example, the Mode[1:0] inputs to the core chip
may be selected by means of fuses on the core chip or on the
interface chip, by means of pull-up or pull-down resistors in the
package of either chip (or in the common package) or on the printed
circuit board, or may be driven by a register on the interface
chip, or may be part of the address input to the core chip.
[0086] Let us assume that the Mode[1:0] inputs to the DRAM core
chip are controlled by a register in the interface chip. FIG. 16
illustrates one embodiment for a portion of a DRAM core chip that
includes a decoder for selecting a mode of operation. The decoder
1610 in the core chip is aware of the Mode[1:0] inputs as shown in
FIG. 16.
[0087] Note that the embodiments disclosed below are based on the
DRAM core chip shown in FIG. 11 only for the purpose of explaining
the concept, and that the following embodiments are applicable to
DRAM core chips of different densities, number of banks, internal
organization, and number of sub-arrays. For the purpose of
simplicity, only bank 0 is shown being accessed in the different
modes of operation.
[0088] The mode decoder truth table is shown below in TABLE 8. In
TABLE 8 below, TABLE-US-00008 TABLE 8 Mode[1:0] RA[14] RA[13]
Bank0a_En Bank0b_En Bank0c_En Bank0d_En 00 X X H H H H 01 X L H H L
L X H L L H H 10 L L H L L L L H L H L L H L L L H L H H L L L H RA
= Row Address X = Don't Care H = Asserted L = Not Asserted
[0089] Based on the techniques of the present invention, a
universal DRAM core chip, such as the embodiment shown in FIG. 11,
may be configured to support a wide variety of data speeds and
widths. For example, the core chip shown in FIG. 11 may support
data rates up to 8*f MB/s, where f is the maximum clock rate (in
MHz) at which the DRAM core can run in sync with the external data
bus without using pre-fetching (i.e. using a pre-fetching of in).
For modern DRAM processes and designs, f is typically between 100
MHz and 200 MHz. So, the DRAM core chip shown in FIG. 11 supports
maximum data rates between 800 MB/s and 1600 MB/s (1.6 GB/s).
[0090] In order to build custom memory cost-effectively, it is
imperative that the DRAM core chip be used in a variety of markets.
This will reduce the cost of the core chip because of economies of
scale. Since the memory core is typically 90% to 95% of the silicon
area, the overall cost may be lowered. Here, we can make two
observations:
[0091] Some markets for the universal DRAM core chip value low
power at the expense of performance (e.g. cell phones and other
handheld devices) whereas other markets will sacrifice power to
achieve higher speed (e.g. PC graphics and game console
markets).
[0092] Semiconductor fabrication process is inherently statistical
in nature. That is, if we fabricate a statistically significant
number of identical chips, some of the chips will only be capable
of operating below the target speed, some of the chip will be
capable of operating at the target speed, and some of the chips
will be capable of operating above the target speed. These are
known in the industry as slow, typical, and fast parts
respectively. Usually, the fast parts are sold at a price premium
over the other parts while the slow parts are sold at lower prices
compared to the typical parts.
[0093] However, slow parts typically consume less power than the
typical parts, which in turn typically consume less power than the
fast parts. So, if we can sort the DRAM core chips according to
their maximum speeds of operation (usually called "speed binning")
before they are attached to the interface chips, we can:
[0094] Attach interface chips designed for low power markets (e.g.
an SDRAM-like interface for the cell phone market) to the slow core
parts;
[0095] Attach interface chips designed for the high
performance/speed markets (e.g. a GDDR3-like interface for the game
console market) to the fast core parts; and
[0096] Attach interface chips designed for markets sensitive to
both power and performance (e.g. a DDR-like interface for the
server market) to the typical core parts.
[0097] This allows us to maximize the ASP (average selling price or
average sales price) of all the solutions since all the core chips
have natural homes.
[0098] Speed binning of memory chips is typically done after it is
packaged. Note that it is possible to do a simple speed sort of the
memory chips at the wafer level itself. In order to do a speed sort
or speed bin, we need to use ATE (automatic test equipment), also
known as tester.
[0099] When we speed bin the DRAM core chips, we need to measure
the time required for basic operations like Read, Write, Activate
(open one or more pages), Precharge (close one or more pages), and
Refresh. To satisfy this requirement, the DRAM core chips, defined
by the present invention, are fully functional asynchronous DRAM
chips capable of stand-alone operation. In other words, the DRAM
core chips contain all the necessary circuits and capabilities
needed to access the internal array used to store the data.
[0100] Memory makers, especially DRAM manufacturers, build
redundancy into the memory core. For example, if the memory array
is to be organized as P.times.Q (P rows and Q columns), the actual
array is designed as (P+i).times.(Q+j), where i and j are small
compared to P and Q respectively. This allows the memory makers to
replace up to i defective rows in the main array with the redundant
rows, and up to j defective columns in the main array with the
redundant columns. With the help of the redundant rows and columns,
memory makers can increase the yield (i.e. the percentage of fully
functional chips) to .gtoreq.90%. In a typical DRAM manufacturing
flow, the individual dies on a wafer are tested at low speed and
the partially functional dies (i.e. those with some defective rows
and/or columns) are marked. The defective rows and/or columns on
these marked dies are replaced with the redundant rows and/or
columns respectively.
[0101] However, a die that uses the redundant rows and/or columns
(because it had some defective rows and/or columns) will be slower
than a die that does not use redundant rows and/or columns. This is
due to the nature of how redundancy is built into the memory and
how it is enabled. Therefore:
[0102] Interface chips that are designed for high performance
markets may be attached to the DRAM core dies that do not have any
defective rows and/or columns in the main array.
[0103] In another embodiment, the defective rows and/or columns of
memory core dies are not replaced with the redundant rows and/or
columns, but are configured to operate the memory core dies as
(P/y).times.(Q/z), where y and z are preferably powers of 2
(including 2.sup.0=1). These DRAM core chips may then be attached
to interface chips that are designed for high performance
markets.
[0104] As we have seen previously, the DRAM makers use pre-fetching
to support higher external data rates. For example, DDR2 SDRAM uses
4n pre-fetching. This means that for an n-bit wide external data
bus, 4n data bits are accessed from the memory core for every read
or write. In a conventional DRAM (where the memory core and the
interface are on the same die), increasing the amount of
pre-fetching increases the amount of metal interconnects on the
die, which has a modest impact on the cost. In the invention
described herein, increasing the amount of pre-fetching may make
either the memory core chip or the interface chip or both pad
limited. Being pad limited can increase the cost substantially.
[0105] Burst mode is another technique that can be used to increase
the data rate of memory chips. In burst mode, the memory chip reads
or writes multiple data bits per column address. For example, an
n-bit wide (external data bus width) memory chip that is configured
for a burst mode of 4n will access 4n bits from the memory core for
a given column address. So this is quite similar to a 4n pre-fetch
except that in burst mode, the same data wires are used. In other
words, in a memory chip that supports 4n burst mode (but not 4n
pre-fetching), the internal data bus between the memory core and
the interface is only n-bits wide. Each line in the internal bus
carries 4 data bits that are separated in time.
[0106] The difference between pre-fetching and burst mode is shown
in FIGS. 17a and 17b. FIG. 17a is a block diagram illustrating the
relationship between the internal data bus rate and the external
data bus rate for 4n pre-fetching. Memory core 1710 is coupled to
memory interface 1720 via internal data bus 1715 at 4n @ f.sub.1,
Hz data rate. The memory interface 1720 is coupled to external data
bus 1725, and under these conditions, the external data bus
operates at a n @ 4*f.sub.1, Hz data rate. FIG. 17b is a block
diagram illustrating the relationship between the internal data bus
rate and the external data bus rate for burst mode with a length of
4n. Memory core 1730 is coupled to memory interface 1750 via
internal data bus 1740 at n @ 4*f.sub.2, Hz data rate. The memory
interface 1750 is coupled to external data bus 1760, and for the
burst mode of operation, the external data bus operates at a n @
4*f.sub.2, Hz data rate.
[0107] Typically, pre-fetching will provide higher external data
rates than burst mode. However, burst mode does not increase the
amount of off-chip connections between the core chip and the
interface chip. So, in some embodiments, it is preferable to design
the DRAM core chip of this invention with burst mode
capability.
[0108] As mentioned previously, one of the aspects of this
invention is the ability to test and speed bin the memory core
chips and then attach the appropriate interface chips. Testing and
speed binning of the DRAM core chip is usually done on a tester.
This requires the core chip to have sufficiently strong output
drivers to drive the inputs of the tester, which are usually some
distance (several inches) from the outputs of the core chip.
However, in the normal mode of operation, the inputs of the
interface chip will be much closer (<1'') to the outputs of the
core chip. So, it is not necessary to have strong output drivers in
the core chip in the normal mode of operation. In order to satisfy
both requirements, in some embodiments the DRAM core chip
preferably has output drivers whose strength or drive capability is
adjustable. For example, the core chip may have, by default, normal
strength output drivers that are capable of driving signals across
some distance to the inputs of the tester. However, when an
interface chip is attached to the core chip, a signal from the
interface chip decreases the drive strength of the core chip's
output drivers.
[0109] In some embodiments, the output drivers of interface chip
that interface with the DRAM core chip have similar adjustable
drive strength capability. This allows testing of the interface
chips separately prior to attaching them to the core chips. Note
that the adjustable drive strength drivers are not necessarily
required on the interface chip on the pins that interface to the
electronic host system. However, it is preferable to have the
adjustable strength capability on these drivers as well so that the
drive strength can be tailored to the requirements of the system or
external world. As with the DRAM core chips, the strength of the
output drivers on the interface chip that communicate with the core
chip are preferably controlled by a signal from the core chip.
[0110] FIG. 18 is a block diagram illustrating an example
multi-chip memory implementation. The solution includes a DRAM core
chip 1810 and an interface chip 1820. In some embodiments, the main
characteristics of the DRAM core chip of this invention are:
[0111] Asynchronous or synchronous DRAM that is capable of
stand-alone operation;
[0112] A plurality of DRAM core chips may be coupled together with
one or more interface chips;
[0113] A plurality of interface chips may be coupled together with
one or more DRAM core chips;
[0114] The interface on the DRAM core chip may include a custom
and/or industry standard interface;
[0115] Has address inputs (bank address, row address, column
address--row and column address can be on separate inputs or
multiplexed on same pins);
[0116] Has command inputs like address strobes, read/write, output
enable, and data masks);
[0117] Has control inputs that determine mode of
operation--examples are inputs that determine the width of the
internal data bus (bus between the memory core chip and interface
chip) and inputs that determine the strength of the output
drivers;
[0118] Has control outputs that determine some aspect of the
functions performed by the interface chip;
[0119] Internal data bus width .gtoreq.external data bus width (bus
from interface chip to memory controller or ASIC);
[0120] Optional burst mode capability;
[0121] Adjustable drive strength on output drivers;
[0122] Capable of well-defined standard operations like Read,
Write, Activate, Precharge, and Refresh that can be clearly
characterized in terms of speed; and
[0123] May be tested, burnt in, and speed binned independently
(i.e. in stand-alone mode).
[0124] In some embodiments, the main characteristics of the
interface chip of this invention are:
[0125] Implements the protocol used by the memory controller;
[0126] Interface chip implements an industry standard protocol like
DDR SDRAM, DDR2 SDRAM, GDDR2 SDRAM, etc.;
[0127] Interface chip implements an industry standard protocol with
custom extensions (e.g. GDDR2 SDRAM with extensions as specified by
mutual agreement with one or more customers);
[0128] Interface chip implements a fully custom protocol as
specified by one or more customers or a fully custom protocol
developed in-house;
[0129] Interface chip operates as a transformer to convert
protocols from the external interface to the interface of the DRAM
core chip (e.g., synchronous to asynchronous and asynchronous to
synchronous);
[0130] Interface chip determines the signaling used by the external
interface;
[0131] For example, single ended, pseudo-differential, fully
differential;
[0132] For example, push-pull outputs, open drain/collector
outputs;
[0133] For example, asynchronous, synchronous, source synchronous,
SerDes-like where clock is encoded/embedded in the data stream;
[0134] Determines the width of the external data bus;
[0135] Determines the speed of operation of the memory chip (by
memory chip, we mean the combination of the DRAM core chip and the
interface chip);
[0136] Determines the pin out of the memory chip as seen by the
external world;
[0137] Allows the pin out of the memory chip to better match the
pin out of the ASIC/memory controller to reduce the board routing
complexity;
[0138] Implements special or custom functions and modes of
operation;
[0139] Example, special power management functions and operating
modes; and
[0140] Example, special error detection and error correction
capabilities, as well as other types of redundancy capabilities and
functions.
[0141] The DRAM core chip and the interface chip of this invention
may be attached together in a number of different ways:
[0142] One or more DRAM core chip dies and one or more interface
chip dies may be electrically connected to each other and the whole
combination be put into a single package (e.g., a single DRAM core
chip die with a single interface chip die, multiple DRAM core chip
dies with a single interface chip die, or a single DRAM core chip
die with multiple interface chip dies).
[0143] The core chip die may be put in a separate package, and then
the interface chip die may then be electrically attached to the
package containing the core chip die;
[0144] The interface chip die may be put in a separate package, and
then the core chip die may then be electrically attached to the
package containing the interface chip die;
[0145] The core chip die can be put in a separate package; the
interface chip die may be put in a separate package; and the two
packages can be electrically attached to each other;
[0146] Multiple DRAM core chip dies may be put in a separate
package; the interface chip die may be put in a separate package;
and the two packages can be electrically attached to each
other;
[0147] A DRAM core chip die may be put in a separate package;
multiple interface chip dies may be put in a separate package; and
the two packages can be electrically attached to each other;
[0148] The DRAM core chip die and the interface chip die may be
electrically attached in any way without deviating from the spirit
or scope of the invention.
[0149] One aspect of this invention is that a multi-chip DRAM that
is built according to this invention, might have higher cost than a
traditional DRAM, especially if the DRAM core chip die and the
interface chip die were packaged separately and then attached to
each other. This is due to the cost associated with the extra
package. One way to ameliorate this is to put multiple DRAM core
chip dies in a single package. For the purpose of this discussion,
we shall consider putting two DRAM core chip dies in a single
package, each die being a 256 Mb density device. A typical DRAM
manufacturing process might have the following sequences after the
wafer has been fully processed:
[0150] DRAM dies on a wafer are tested at low speed and dies with
defective rows and/or columns are marked;
[0151] The defective rows and/or columns are replaced with
redundant rows and/or columns;
[0152] Wafer is diced into individual dies, which are then
packaged;
[0153] Packaged parts are tested for functionality--parts damaged
by the packaging process are eliminated;
[0154] Tested packaged parts undergo long term burn in to eliminate
infant mortality parts; and
[0155] Burnt in parts are optionally tested again for functionality
and shipped.
[0156] So, if we place two 256 Mb DRAM core dies in a single
package, the following 3 bins may be generated after the parts have
been packaged and burnt-in:
[0157] Bin A--both the DRAM core chip dies are functional, so the
total capacity is 512 Mb
[0158] Bin B--only one of the DRAM core chip dies is functional, so
the total capacity is 256 Mb
[0159] Bin C--neither of the DRAM core chip dies is functional, so
the total capacity is 0 Mb
[0160] The bin C parts should be discarded. We can now use the bin
B parts for those markets and/or customers who require only 256 Mb
devices. For example, a handheld device manufacturer might require
only a 256 Mb DRAM. So, the bin B parts can be attached to the
interface chips designed for this manufacturer. Other markets
and/or manufacturers might require 512 Mb devices. For example, a
network router manufacturer might need 512 Mb DRAMs. So, we can use
bin A parts for this manufacturer by attaching the appropriate
interface chips to the bin A parts. This concept can be extended to
cover more than two DRAM core chip dies in a single package as well
as DRAM core chip dies of all densities.
[0161] As disclosed above, some embodiments cover the idea of
attaching multiple DRAM core chips to a single interface chip. The
core chips may be attached to the interface chip in a number of
different ways. Some of these ways are described below and in FIG.
19. Again, for the purpose of illustration, we shall assume that
two 4-bank 256 Mb DRAM core chips (FIG. 11) are attached to the
interface chip. Each of the core chips has a 64-bit wide data bus
designed to connect to the interface chip. Note that the idea
explained below can be applied to DRAM core chips with different
number of banks, density, data bus width, etc.
[0162] In one method (1910), the row addresses are used to select
the DRAM core chip. In a second method (1920), bank addresses are
used to select DRAM core chips. In a third method (1930), the
column addresses are used to select the DRAM core chip.
[0163] The two DRAM core chips are attached to the interface chip
such that only one of the core chips is accessible at any given
time. That is, the two core chips look to the electronic host
system as a single 512 Mb DRAM with 4 banks. This implies that the
interface chip will use the row address to select one or the other
core chip.
[0164] The two DRAM core chips are attached to the interface such
that both of the core chips are accessible at any given time, and
that the two chips look to the external world electronic host
system as a single 512 Mb DRAM with 8 banks. This implies that the
interface chip will use the bank address to select one or the other
core chip.
[0165] The two DRAM core chips are attached to the interface chip
such that both of the core chips are accessible at any given time,
and that the two chips look to the electronic host system as a
single 512 Mb DRAM with 4 banks. The interface chip uses the column
address to select one or the other core chip. Note that in other
embodiments an interface chip always accesses both the core chips
in parallel, so that the data bus between the core chips and the
interface chip becomes 128-bits wide. For this embodiment, the
interface chip doubles the external data rate. In other words, the
amount of pre-fetching has been doubled.
[0166] In other embodiments the multi-chip solution is configured
such that the attachment of the DRAM core chips to the interface
chip is programmable. For example, the customer may choose between
one of the three methods listed above to attach two 4-bank, 64-bit
wide, 256 Mb DRAM core chips to an interface chip by programmable
means. These means include using fuses on the interface chip or the
core chips, pull-up or pull-down resistors on the package
substrates or printed circuit board, or by means of a register on
the interface chip or core chips. However, any way of attaching the
DRAM core chips to the interface chip may be accomplished without
deviating from the spirit or scope of the invention.
[0167] Other embodiments of the invention include building
redundant memory systems by attaching multiple DRAM core chips to
an interface chip. For example, when more than one core chip is
attached to an interface chip, redundancy is added by several means
including:
[0168] Storing identical copies of the data in corresponding
locations of each core chip when data is written to the memory by
the memory controller; when the data is read back by the memory
controller, the interface chip can read the multiple copies of the
data from the different core chips and select the correct copy and
transmit it to the memory controller. The correct copy can be
determined by means like majority voting, and/or by the use of
parity or ECC bits;
[0169] Using (n+m) bits to store n data bits.
[0170] Another aspect of this invention is placing memory core
chips of different types behind a common interface chip. For
example, we can place any combination of DRAM core, SRAM (Static
Random Access Memory), and Flash chips behind a common interface
chip. Multi-Chip Package (MCP) memory solutions are fairly common
in the cell phone and handheld markets today. The issue with
current MCP solutions is that each of these memories (DRAM, SRAM,
Flash) has different interfaces, which complicates the design of
the memory controller, the packaging, and the board routing.
Placing any possible combination of DRAM core chip, SRAM, and Flash
behind a common interface chip simplifies the memory controller
design since the idiosyncrasies of each of these memory types is
hidden from the memory controller. In addition, the board routing
is simplified.
[0171] Another aspect of this invention is placing a large and slow
memory as well as a smaller and faster memory behind a common
interface chip and using the faster memory as a cache for the
slower memory. For example, an SRAM chip might be used as the cache
for a DRAM core chip or a DRAM core chip can be used as the cache
for a Flash chip. The cache management logic may be built into the
interface chip so that the cache is transparent to the memory
controller. Alternately, the cache may be made visible to the
memory controller and managed by the memory controller. Let us
consider the case of an interface chip that has been designed to
interface to one or more DRAM core chips and an SRAM chip. The SRAM
chip can be used to cache the rows in the DRAM core chips that were
recently opened. The SRAM may be used to cache the entire contents
of the recently opened rows in the DRAM or cache part of the
contents of the recently opened rows in the DRAM. The properties of
the cache (associatively of the cache lines, mapping between DRAM
rows and SRAM cache lines, etc.) may be determined by programming
certain registers in the interface chip. By storing data that has a
high likelihood of being accessed in the near future, system
performance is improved.
[0172] Using a cache also allows the interface chip to do
speculative pre-fetching of data from the DRAM core chip (and
storing it in the SRAM cache chip), which again improves system
performance. In addition, the interface chip may operate with SRAM
chips (that are used as caches) of different densities. This allows
the same interface chip to be used across several different
segments within a market. For example, a DRAM/SRAM combination
memory device may includes a common interface where the DRAM
capacity is 512 Mb and the SRAM capacity ranges from 0 to 32 Mb.
This allows a DRAM supplier to ship the 512 Mb DRAM+32 Mb SRAM
combination in the high performance segment of the market, ship a
512 Mb DRAM+8 Mb SRAM combination in the mainstream segment of the
market, and ship a 512 Mb DRAM (no SRAM cache) device in the value
segment of the market.
[0173] The techniques of the present invention are not just
applicable to DRAM. As such, the DRAM core chip need not
necessarily be a trench-capacitor or stacked-capacitor device. The
present invention is applicable to a variety of memory technologies
like MRAM (Magnetic RAM), FRAM (Ferro-electric RAM), Ovonics
memory, molecular memory (e.g. memory technology developed by
ZettaCore), carbon nanotube memory (e.g. memory technology
developed by Nantero Inc.), etc.
[0174] Another aspect of this invention is that it can be used with
DRAM core chips that have different architectures like FCRAM (Fast
Cycle RAM), RLDRAM (Reduced Latency DRAM), ESDRAM (Enhanced
SDRAM).
[0175] Some embodiments of invention allow the use of a common
memory core across a wide range of markets while varying the
interface of the memory core according to the market and customer
needs. It also allows the interface to be kept constant while
changing the memory core behind the interface to address the needs
of the different segments within a market.
[0176] Although the present invention has been described in terms
of specific exemplary embodiments, it will be appreciated that
various modifications and alterations might be made by those
skilled in the art without departing from the spirit and scope of
the invention.
* * * * *