Memory module with memory stack and interface with enhanced capabilities

[0001] This application is a divisional application of, and claims priority to, U.S. patent application Ser. No. 11/702,981, entitled “Memory Module with Memory Stack and Interface with Enhanced Capabilities,” which was filed on Feb. 5, 2007. U.S. patent application Ser. No. 11/702,981 claims the benefit to U.S. Provisional Patent Application No. 60/772,414, entitled “Multi-Rank Memory Buffer and Memory Stack”, filed on Feb. 9, 2006. U.S. patent application Ser. No. 11/702,981 also claims the benefit to U.S. Provisional Patent Application No. 60/865,624, entitled “Memory Subsystem and Method,” filed on Nov. 13, 2006; and this application is a continuation-in-part and further claims the benefit to U.S. patent application Ser. No. 11/461,437, entitled “Memory Refresh System and Method”, filed on Jul. 31, 2006. The disclosures of the above-identified patent applications are expressly incorporated herein by reference.

[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] The memory capacity requirements of computers in general, and servers in particular, are increasing at a very rapid pace due to several key trends in the computing industry. The first trend is 64-bit computing, which enables processors to address more than 4 GB of physical memory. The second trend is multi-core CPUs, where each core runs an independent software thread. The third trend is server virtualization or consolidation, which allows multiple operating systems and software applications to run simultaneously on a common hardware platform. The fourth trend is web services, hosted applications, and on-demand software, where complex software applications are centrally run on servers instead of individual copies running on desktop and mobile computers. The intersection of all these trends has created a step function in the memory capacity requirements of servers.

[0006] However, the trends in the DRAM industry are not aligned with this step function. As the DRAM interface speeds increase, the number of loads (or ranks) on the traditional multi-drop memory bus decreases in order to facilitate high speed operation of the bus. In addition, the DRAM industry has historically had an exponential relationship between price and DRAM density, such that the highest density ICs or integrated circuits have a higher $/Mb ratio than the mainstream density integrated circuits. These two factors usually place an upper limit on the amount of memory (i.e. the memory capacity) that can be economically put into a server.

[0007] One solution to this memory capacity gap is to use a fully buffered DIMM (FB-DIMM), and this is currently being standardized by JEDEC. FIG. 1A illustrates a fully buffered DIMM. As shown in FIG. 1A, memory controller 100 communicates with FB-DIMMs (130 and 140) via advanced memory buffers (AMB) 110 and 120 to operate a plurality of DRAMs. As shown in FIG. 1B, the FB-DIMM approach uses a point-to-point, serial protocol link between the memory controller 100 and FB-DIMMs 150, 151, and 152. In order to read the DRAM devices on, say, the third FB-DIMM 152, the command has to travel through the AMBs on the first FB-DIMM 150 and second FB-DIMM 151 over the serial link segments 141, 142, and 143, and the data from the DRAM devices on the third FB-DIMM 152 must travel back to the memory controller 100 through the AMBs on the first and second FB-DIMMs over serial link segments 144, 145, and 146.

[0008] The FB-DIMM approach creates a direct correlation between maximum memory capacity and the printed circuit board (PCB) area. In other words, a larger PCB area is required to provide larger memory capacity. Since most of the growth in the server industry is in the smaller form factor servers like 1 U/2 U rack servers and blade servers, the FB-DIMM solution does not solve the memory capacity gap for small form factor servers. So, clearly there exists a need for dense memory technology that fits into the mechanical and thermal envelopes of current memory systems.

[0009] A memory module comprises at least one memory stack. The memory stack includes a plurality of DRAM integrated circuits. A buffer circuit, which couples the memory module to a host system, interfaces the memory stack to the host system for transforming one or more physical parameters between the DRAM integrated circuits and the host system. In other embodiments, the buffer circuit interfaces the memory stack to the host system for configuring one or more of the DRAM integrated circuits in the memory stack.

[0028] In one embodiment of this invention, multiple buffer integrated circuits are used to buffer the DRAM integrated circuits or devices on a DIMM as opposed to the FB-DIMM approach, where a single buffer integrated circuit is used to buffer all the DRAM integrated circuits on a DIMM. That is, a bit slice approach is used to buffer the DRAM integrated circuits. As an option, multiple DRAMs may be connected to each buffer integrated circuit. In other words, the DRAMs in a slice of multiple DIMMs may be collapsed or coalesced or stacked behind each buffer integrated circuit, such that the buffer integrated circuit is between the stack of DRAMs and the electronic host system. FIGS. 2A-2C illustrate one embodiment of a DIMM with multiple DRAM stacks, where each DRAM stack comprises a bit slice across multiple DIMMs. As an example, FIG. 2A shows four DIMMs (e.g., DIMM A, DIMM B, DIMM C and DIMM D). Also, in this example, there are 9 bit slices labeled DAO, . . . , DA6, . . . DA8 across the four DIMMs. Bit slice “6” is shown encapsulated in block 210. FIG. 2B illustrates a buffered DRAM stack. The buffered DRAM stack 230 comprises a buffer integrated circuit (220) and DRAM devices DA6, DB6, DC6 and DD6. Thus, bit slice 6 is generated from devices DA6, DB6, DC6 and DD6. FIG. 2C is a top view of a high density DIMM with a plurality of buffered DRAM stacks. A high density DIMM (240) comprises buffered DRAM stacks (250) in place of individual DRAMs.

[0029] Some exemplary embodiments include:

• • [0030] (a) a configuration with increased DIMM density, that allows the total memory capacity of the system to increase without requiring a larger PCB area. Thus, higher density DIMMs fit within the mechanical and space constraints of current DIMMs.
  • [0031] (b) a configuration with distributed power dissipation, which allows the higher density DIMM to fit within the thermal envelope of existing DIMMs. In an embodiment with multiple buffers on a single DIMM, the power dissipation of the buffering function is spread out across the DIMM.
  • [0032] (c) a configuration with non-cumulative latency to improve system performance. In a configuration with non-cumulative latency, the latency through the buffer integrated circuits on a DIMM is incurred only when that particular DIMM is being accessed.

[0033] In a buffered DRAM stack embodiment, the plurality of DRAM devices in a stack are electrically behind the buffer integrated circuit. In other words, the buffer integrated circuit sits electrically between the plurality of DRAM devices in the stack and the host electronic system and buffers some or all of the signals that pass between the stacked DRAM devices and the host system. Since the DRAM devices are standard, off-the-shelf, high speed devices (like DDR SDRAMs or DDR2 SDRAMs), the buffer integrated circuit may have to re-generate some of the signals (e.g. the clocks) while other signals (e.g. data signals) may have to be re-synchronized to the clocks or data strobes to minimize the jitter of these signals. Other signals (e.g. address signals) may be manipulated by logic circuits such as decoders. Some embodiments of the buffer integrated circuit may not re-generate or re-synchronize or logically manipulate some or all of the signals between the DRAM devices and host electronic system.

[0034] The buffer integrated circuit and the DRAM devices may be physically arranged in many different ways. In one embodiment, the buffer integrated circuit and the DRAM devices may all be in the same stack. In another embodiment, the buffer integrated circuit may be separate from the stack of DRAM integrated circuits (i.e. buffer integrated circuit may be outside the stack). In yet another embodiment, the DRAM integrated circuits that are electrically behind a buffer integrated circuit may be in multiple stacks (i.e. a buffer integrated circuit may interface with a plurality of stacks of DRAM integrated circuits).

[0035] In one embodiment, the buffer integrated circuit can be designed such that the DRAM devices that are electrically behind the buffer integrated circuit appear as a single DRAM integrated circuit to the host system, whose capacity is equal to the combined capacities of all the DRAM devices in the stack. So, for example, if the stack contains eight 512 Mb DRAM integrated circuits, the buffer integrated circuit of this embodiment is designed to make the stack appear as a single 4 Gb DRAM integrated circuit to the host system. An un-buffered DIMM, registered DIMM, S0-DIMM, or FB-DIMM can now be built using buffered stacks of DRAMs instead of individual DRAM devices. For example, a double rank registered DIMM that uses buffered DRAM stacks may have eighteen stacks, nine of which may be on one side of the DIMM PCB and controlled by a first integrated circuit select signal from the host electronic system, and nine may be on the other side of the DIMM PCB and controlled by a second integrated circuit select signal from the host electronic system. Each of these stacks may contain a plurality of DRAM devices and a buffer integrated circuit.

[0036] FIG. 3A illustrates a DIMM PCB with buffered DRAM stacks. As shown in FIG. 3A, both the top and bottom sides of the DIMM PCB comprise a plurality of buffered DRAM stacks (e.g., 310 and 320). Note that the register and clock PLL integrated circuits of a registered DIMM are not shown in this figure for simplicity's sake. FIG. 3B illustrates a buffered DRAM stack that emulates a 4 Gb DRAM.

[0037] In one embodiment, a buffered stack of DRAM devices may appear as or emulate a single DRAM device to the host system. In such a case, the number of memory banks that are exposed to the host system may be less than the number of banks that are available in the stack. To illustrate, if the stack contained eight 512 Mb DRAM integrated circuits, the buffer integrated circuit of this embodiment will make the stack look like a single 4 Gb DRAM integrated circuit to the host system. So, even though there are thirty two banks (four banks per 512 Mb integrated circuit*eight integrated circuits) in the stack, the buffer integrated circuit of this embodiment might only expose eight banks to the host system because a 4 Gb DRAM will nominally have only eight banks. The eight 512 Mb DRAM integrated circuits in this example may be referred to as physical DRAM devices while the single 4 Gb DRAM integrated circuit may be referred to as a virtual DRAM device. Similarly, the banks of a physical DRAM device may be referred to as a physical bank whereas the bank of a virtual DRAM device may be referred to as a virtual bank.

[0038] In another embodiment of this invention, the buffer integrated circuit is designed such that a stack of n DRAM devices appears to the host system as m ranks of DRAM devices (where n≧m, and m≧2). To illustrate, if the stack contained eight 512 Mb DRAM integrated circuits, the buffer integrated circuit of this embodiment may make the stack appear as two ranks of 2 Gb DRAM devices (for the case of m=2), or appear as four ranks of 1 Gb DRAM devices (for the case of m=4), or appear as eight ranks of 512 Mb DRAM devices (for the case of m=8). Consequently, the stack of eight 512 Mb DRAM devices may feature sixteen virtual banks (m=2; eight banks per 2 Gb virtual DRAM*two ranks), or thirty two virtual banks (m=4; eight banks per 1 Gb DRAM*four ranks), or thirty two banks (m=8; four banks per 512 Mb DRAM*eight ranks).

[0039] In one embodiment, the number of ranks may be determined by the number of integrated circuit select signals from the host system that are connected to the buffer integrated circuit. For example, the most widely used JEDEC approved pin out of a DIMM connector has two integrated circuit select signals. So, in this embodiment, each stack may be made to appear as two DRAM devices (where each integrated circuit belongs to a different rank) by routing the two integrated circuit select signals from the DIMM connector to each buffer integrated circuit on the DIMM. For the purpose of illustration, let us assume that each stack of DRAM devices has a dedicated buffer integrated circuit, and that the two integrated circuit select signals that are connected on the motherboard to a DIMM connector are labeled CS0# and CS1#. Let us also assume that each stack is 8-bits wide (i.e. has eight data pins), and that the stack contains a buffer integrated circuit and eight 8-bit wide 512 Mb DRAM integrated circuits. In this example, both CS0# and CS1# are connected to all the stacks on the DIMM. So, a single-sided registered DIMM with nine stacks (with CS0# and CS1# connected to all nine stacks) effectively features two 2 GB ranks, where each rank has eight banks.

[0040] In another embodiment, a double-sided registered DIMM may be built using eighteen stacks (nine on each side of the PCB), where each stack is 4-bits wide and contains a buffer integrated circuit and eight 4-bit wide 512 Mb DRAM devices. As above, if the two integrated circuit select signals CS0# and CS1# are connected to all the stacks, then this DIMM will effectively feature two 4 GB ranks, where each rank has eight banks. However, half of a rank's capacity is on one side of the DIMM PCB and the other half is on the other side. For example, let us number the stacks on the DIMM as S0 through S17, such that stacks S0 through S8 are on one side of the DIMM PCB while stacks S9 through S17 are on the other side of the PCB. Stack S0 may be connected to the host system's data lines DQ[3:0], stack S9 connected to the host system's data lines DQ[7:4], stack 51 to data lines DQ[11:8], stack S10 to data lines DQ[15:12], and so on. The eight 512 Mb DRAM devices in stack S0 may be labeled as S0_M0 through S0_M7 and the eight 512 Mb DRAM devices in stack S9 may be labeled as S9_M0 through S9_M7. In one example, integrated circuits S0_M0 through S0_M3 may be used by the buffer integrated circuit associated with stack S0 to emulate a 2 Gb DRAM integrated circuit that belongs to the first rank (i.e. controlled by integrated circuit select CS0#). Similarly, integrated circuits S0_M4 through S0_M7 may be used by the buffer integrated circuit associated with stack S0 to emulate a 2 Gb DRAM integrated circuit that belongs to the second rank (i.e. controlled by integrated circuit select CS1#). So, in general, integrated circuits Sn_M0 through Sn_M3 may be used to emulate a 2 Gb DRAM integrated circuit that belongs to the first rank while integrated circuits Sn_M4 through Sn_M7 may be used to emulate a 2 Gb DRAM integrated circuit that belongs to the second rank, where n represents the stack number (i.e. 0≦n≦17). It should be noted that the configuration described above is just for illustration. Other configurations may be used to achieve the same result without deviating from the spirit or scope of the claims. For example, integrated circuits S0_M0, S0_M2, S0_M4, and S0_M6 may be grouped together by the associated buffer integrated circuit to emulate a 2 Gb DRAM integrated circuit in the first rank while integrated circuits S0_M1, S0_M3, S0_M5, and S0_M7 may be grouped together by the associated buffer integrated circuit to emulate a 2 Gb DRAM integrated circuit in the second rank of the DIMM.

[0041] FIG. 4A illustrates an example of a registered DIMM that uses buffer integrated circuits and DRAM stacks. For simplicity sake, note that the register and clock PLL integrated circuits of a registered DIMM are not shown. The DIMM PCB 400 includes buffered DRAM stacks on the top side of DIMM PCB 400 (e.g., S5) as well as the bottom side of DIMM PCB 400 (e.g., S 15). Each buffered stack emulates two DRAMs. FIG. 4B illustrates a physical stack of DRAM devices in this embodiment. For example, stack 420 comprises eight 4-bit wide, 512 Mb DRAM devices and a buffer integrated circuit 430. As shown in FIG. 4B, a first group of devices, consisting of Sn_M0, Sn_M1, Sn_M2 and Sn_M3, is controlled by CS0#. A second group of devices, which consists of Sn_M4, Sn_M5, Sn_M6 and Sn_M7, is controlled by CS1#. It should be noted that the eight DRAM devices and the buffer integrated circuit are shown as belonging to one stack in FIG. 4B strictly as an example. Other implementations are possible. For example, the buffer integrated circuit 430 may be outside the stack of DRAM devices. Also, the eight DRAM devices may be arranged in multiple stacks.

[0042] In an optional variation of the multi-rank embodiment, a single buffer integrated circuit may be associated with a plurality of stacks of DRAM integrated circuits. In the embodiment exemplified in FIGS. 5A and 5B, a buffer integrated circuit is dedicated to two stacks of DRAM integrated circuits. FIG. 5B shows two stacks, one on each side of the DIMM PCB, and one buffer integrated circuit B0 situated on one side of the DIMM PCB. However, this is strictly for the purpose of illustration. The stacks that are associated with a buffer integrated circuit may be on the same side of the DIMM PCB or may be on both sides of the PCB.

[0043] In the embodiment exemplified in FIGS. 5A and 5B, each stack of DRAM devices contains eight 512 Mb integrated circuits, the stacks are numbered S0 through S17, and within each stack, the integrated circuits are labeled Sn_M0 through Sn_M7 (where n is 0 through 17). Also, for this example, the buffer integrated circuit is 8-bits wide, and the buffer integrated circuits are numbered B0 through B8. The two integrated circuit select signals, CS0# and CS1#, are connected to buffer B0 as are the data lines DQ[7:0]. As shown, stacks S0 through S8 are the primary stacks and stacks S9 through S17 are optional stacks. The stack S9 is placed on the other side of the DIMM PCB, directly opposite stack S0 (and buffer B0). The integrated circuits in stack S9 are connected to buffer B0. In other words, the DRAM devices in stacks S0 and S9 are connected to buffer B0, which in turn, is connected to the host system. In the case where the DIMM contains only the primary stacks S0 through S8, the eight DRAM devices in stack S0 are emulated by the buffer integrated circuit B0 to appear to the host system as two 2 Gb devices, one of which is controlled by CS0# and the other is controlled by CS1#. In the case where the DIMM contains both the primary stacks S0 through S8 and the optional stacks S9 through S17, the sixteen 512 Mb DRAM devices in stacks S0 and S9 are together emulated by buffer integrated circuit B0 to appear to the host system as two 4 Gb DRAM devices, one of which is controlled by CS0# and the other is controlled by CS1#.

[0044] It should be clear from the above description that this architecture decouples the electrical loading on the memory bus from the number of ranks. So, a lower density DIMM can be built with nine stacks (S0 through S8) and nine buffer integrated circuits (B0 through B8), and a higher density DIMM can be built with eighteen stacks (S0 through S17) and nine buffer integrated circuits (B0 through B8). It should be noted that it is not necessary to connect both integrated circuit select signals CS0# and CS1# to each buffer integrated circuit on the DIMM. A single rank lower density DIMM may be built with nine stacks (S0 through S8) and nine buffer integrated circuits (B0 through B8), wherein CS0# is connected to each buffer integrated circuit on the DIMM. Similarly, a single rank higher density DIMM may be built with seventeen stacks (S0 through S17) and nine buffer integrated circuits, wherein CS0# is connected to each buffer integrated circuit on the DIMM.

[0045] A DIMM implementing a multi-rank embodiment using a multi-rank buffer is an optional feature for small form factor systems that have a limited number of DIMM slots. For example, consider a processor that has eight integrated circuit select signals, and thus supports up to eight ranks. Such a processor may be capable of supporting four dual-rank DIMMs or eight single-rank DIMMs or any other combination that provides eight ranks Assuming that each rank has y banks and that all the ranks are identical, this processor may keep up to 8*y memory pages open at any given time. In some cases, a small form factor server like a blade or 1 U server may have physical space for only two DIMM slots per processor. This means that the processor in such a small form factor server may have open a maximum of 4*y memory pages even though the processor is capable of maintaining 8*y pages open. For such systems, a DIMM that contains stacks of DRAM devices and multi-rank buffer integrated circuits may be designed such that the processor maintains 8*y memory pages open even though the number of DIMM slots in the system are fewer than the maximum number of slots that the processor may support. One way to accomplish this, is to apportion all the integrated circuit select signals of the host system across all the DIMM slots on the motherboard. For example, if the processor has only two dedicated DIMM slots, then four integrated circuit select signals may be connected to each DIMM connector. However, if the processor has four dedicated DIMM slots, then two integrated circuit select signals may be connected to each DIMM connector.

[0046] To illustrate the buffer and DIMM design, say that a buffer integrated circuit is designed to have up to eight integrated circuit select inputs that are accessible to the host system. Each of these integrated circuit select inputs may have a weak pull-up to a voltage between the logic high and logic low voltage levels of the integrated circuit select signals of the host system. For example, the pull-up resistors may be connected to a voltage (VTT) midway between VDDQ and GND (Ground). These pull-up resistors may be on the DIMM PCB. Depending on the design of the motherboard, two or more integrated circuit select signals from the host system may be connected to the DIMM connector, and hence to the integrated circuit select inputs of the buffer integrated circuit. On power up, the buffer integrated circuit may detect a valid low or high logic level on some of its integrated circuit select inputs and may detect VTT on some other integrated circuit select inputs. The buffer integrated circuit may now configure the DRAMs in the stacks such that the number of ranks in the stacks matches the number of valid integrated circuit select inputs.

[0047] FIG. 6A illustrates a memory controller that connects to two DIMMS. Memory controller (600) from the host system drives 8 integrated circuit select (CS) lines: CS0# through CS7#. The first four lines (CS0#-CS3#) are used to select memory ranks on a first DIMM (610), and the second four lines (CS4#-CS7#) are used to select memory ranks on a second DIMM (620). FIG. 6B illustrates a buffer and pull-up circuitry on a DIMM used to configure the number of ranks on a DIMM. For this example, buffer 630 includes eight (8) integrated circuits select inputs (CS0#-CS7#). A pull-up circuit on DIMM 610 pulls the voltage on the connected integrated circuit select lines to a midway voltage value (i.e., midway between VDDQ and GND, VTT). CS0#-CS3# are coupled to buffer 630 via the pull-up circuit. CS4#-CS7# are not connected to DIMM 610. Thus, for this example, DIMM 610 configures ranks based on the CS0#-CS3# lines.

[0048] Traditional motherboard designs hard wire a subset of the integrated circuit select signals to each DIMM connector. For example, if there are four DIMM connectors per processor, two integrated circuit select signals may be hard wired to each DIMM connector. However, for the case where only two of the four DIMM connectors are populated, only 4*y memory banks are available even though the processor supports 8*y banks because only two of the four DIMM connectors are populated with DIMMs. One method to provide dynamic memory bank availability is to configure a motherboard where all the integrated circuit select signals from the host system are connected to all the DIMM connectors on the motherboard. On power up, the host system queries the number of populated DIMM connectors in the system, and then apportions the integrated circuit selects across the populated connectors.

[0049] In one embodiment, the buffer integrated circuits may be programmed on each DIMM to respond only to certain integrated circuit select signals. Again, using the example above of a processor with four dedicated DIMM connectors, consider the case where only two of the four DIMM connectors are populated. The processor may be programmed to allocate the first four integrated circuit selects (e.g., CS0# through CS3#) to the first DIMM connector and allocate the remaining four integrated circuit selects (say, CS4# through CS7#) to the second DIMM connector. Then, the processor may instruct the buffer integrated circuits on the first DIMM to respond only to signals CS0# through CS3# and to ignore signals CS4# through CS7#. The processor may also instruct the buffer integrated circuits on the second DIMM to respond only to signals CS4# through CS7# and to ignore signals CS0# through CS3#. At a later time, if the remaining two DIMM connectors are populated, the processor may then re-program the buffer integrated circuits on the first DIMM to respond only to signals CS0# and CS1#, re-program the buffer integrated circuits on the second DIMM to respond only to signals CS2# and CS3#, program the buffer integrated circuits on the third DIMM to respond to signals CS4# and CS5#, and program the buffer integrated circuits on the fourth DIMM to respond to signals CS6# and CS7#. This approach ensures that the processor of this example is capable of maintaining 8*y pages open irrespective of the number of DIMM connectors that are populated (assuming that each DIMM has the ability to support up to 8 memory ranks). In essence, this approach de-couples the number of open memory pages from the number of DIMMs in the system.

[0050] FIGS. 7A and 7B illustrate a memory system that configures the number of ranks in a DIMM based on commands from a host system. FIG. 7A illustrates a configuration between a memory controller and DIMMs. For this embodiment, all the integrated circuit select lines (e.g., CS0#-CS7#) are coupled between memory controller 730 and DIMMs 710 and 720. FIG. 7B illustrates the coupling of integrated circuit select lines to a buffer on a DIMM for configuring the number of ranks based on commands from the host system. For this embodiment, all integrated circuit select lines (CS0#-CS7#) are coupled to buffer 740 on DIMM 710.

[0051] Virtualization and multi-core processors are enabling multiple operating systems and software threads to run concurrently on a common hardware platform. This means that multiple operating systems and threads must share the memory in the server, and the resultant context switches could result in increased transfers between the hard disk and memory.

[0052] In an embodiment enabling multiple operating systems and software threads to run concurrently on a common hardware platform, the buffer integrated circuit may allocate a set of one or more memory devices in a stack to a particular operating system or software thread, while another set of memory devices may be allocated to other operating systems or threads. In the example of FIG. 6C, the host system (not shown) may operate such that a first operating system is partitioned to a first logical address range 660, corresponding to physical partition 680, and all other operating systems are partitioned to a second logical address range 670, corresponding to a physical partition 690. On a context switch toward the first operating system or thread from another operating system or thread, the host system may notify the buffers on a DIMM or on multiple DIMMs of the nature of the context switch. This may be accomplished, for example, by the host system sending a command or control signal to the buffer integrated circuits either on the signal lines of the memory bus (i.e. in-band signaling) or on separate lines (i.e. side band signaling). An example of side band signaling would be to send a command to the buffer integrated circuits over an SMBus. The buffer integrated circuits may then place the memory integrated circuits allocated to the first operating system or thread 680 in an active state while placing all the other memory integrated circuits allocated to other operating systems or threads 690 (that are not currently being executed) in a low power or power down mode. This optional approach not only reduces the power dissipation in the memory stacks but also reduces accesses to the disk. For example, when the host system temporarily stops execution of an operating system or thread, the memory associated with the operating system or thread is placed in a low power mode but the contents are preserved. When the host system switches back to the operating system or thread at a later time, the buffer integrated circuits bring the associated memory out of the low power mode and into the active state and the operating system or thread may resume the execution from where it left off without having to access the disk for the relevant data. That is, each operating system or thread has a private main memory that is not accessible by other operating systems or threads. Note that this embodiment is applicable for both the single rank and the multi-rank buffer integrated circuits.

[0053] When users desire to increase the memory capacity of the host system, the normal method is to populate unused DIMM connectors with memory modules. However, when there are no more unpopulated connectors, users have traditionally removed the smaller capacity memory modules and replaced them with new, larger capacity memory modules. The smaller modules that were removed might be used on other host systems but typical practice is to discard them. It could be advantageous and cost-effective if users could increase the memory capacity of a system that has no unpopulated DIMM connectors without having to discard the modules being currently used.

[0054] In one embodiment employing a buffer integrated circuit, a connector or some other interposer is placed on the DIMM, either on the same side of the DIMM PCB as the buffer integrated circuits or on the opposite side of the DIMM PCB from the buffer integrated circuits. When a larger memory capacity is desired, the user may mechanically and electrically couple a PCB containing additional memory stacks to the DIMM PCB by means of the connector or interposer. To illustrate, an example multi-rank registered DIMM may have nine 8-bit wide stacks, where each stack contains a plurality of DRAM devices and a multi-rank buffer. For this example, the nine stacks may reside on one side of the DIMM PCB, and one or more connectors or interposers may reside on the other side of the DIMM PCB. The capacity of the DIMM may now be increased by mechanically and electrically coupling an additional PCB containing stacks of DRAM devices to the DIMM PCB using the connector(s) or interposer(s) on the DIMM PCB. For this embodiment, the multi-rank buffer integrated circuits on the DIMM PCB may detect the presence of the additional stacks and configure themselves to use the additional stacks in one or more configurations employing the additional stacks. It should be noted that it is not necessary for the stacks on the additional PCB to have the same memory capacity as the stacks on the DIMM PCB. In addition, if the stacks on the DIMM PCB may be connected to one integrated circuit select signal while the stacks on the additional PCB may be connected to another integrated circuit select signal. Alternately, the stacks on the DIMM PCB and the stacks on the additional PCB may be connected to the same set of integrated circuit select signals.

[0055] FIG. 8 illustrates one embodiment for a DIMM PCB with a connector or interposer with upgrade capability. A DIMM PCB 800 comprises a plurality of buffered stacks, such as buffered stack 830. As shown, buffered stack 830 includes buffer integrated circuit 840 and DRAM devices 850. An upgrade module PCB 810, which connects to DIMM PCB 800 via connector or interposer 880 and 870, includes stacks of DRAMs, such as DRAM stack 820. In this example and as shown in FIG. 8, the upgrade module PCB 810 contains nine 8-bit wide stacks, wherein each stack contains only DRAM integrated circuits 860. Each multi-rank buffer integrated circuit 840 on DIMM PCB 800, upon detection of the additional stack, re-configures itself such that it sits electrically between the host system and the two stacks of DRAM integrated circuits. That is, the buffer integrated circuit is now electrically between the host system and the stack on the DIMM PCB 800 as well as the corresponding stack on the upgrade module PCB 810. However, it should be noted that other embodiments of the buffer integrated circuit (840), the DRAM stacks (820), the DIMM PCB 800, and the upgrade module PCB 810 may be configured in various manners to achieve the same result, without deviating from the spirit or scope of the claims. For example, the stack 820 on the additional PCB may also contain a buffer integrated circuit. So, in this example, the upgrade module 810 may contain one or more buffer integrated circuits.

[0056] The buffer integrated circuits may map the addresses from the host system to the DRAM devices in the stacks in several ways. In one embodiment, the addresses may be mapped in a linear fashion, such that a bank of the virtual (or emulated) DRAM is mapped to a set of physical banks, and wherein each physical bank in the set is part of a different physical DRAM device. To illustrate, let us consider a stack containing eight 512 Mb DRAM integrated circuits (i.e. physical DRAM devices), each of which has four memory banks. Let us also assume that the buffer integrated circuit is the multi-rank embodiment such that the host system sees two 2 Gb DRAM devices (i.e. virtual DRAM devices), each of which has eight banks. If we label the physical DRAM devices M0 through M7, then a linear address map may be implemented as shown below.

[0057]

	Host System Address	DRAM Device
	(Virtual Bank)	(Physical Bank)
	Rank 0, Bank [0]	{(M4, Bank [0]), (M0, Bank [0])}
	Rank 0, Bank [1]	{(M4, Bank [1]), (M0, Bank [1]))
	Rank 0, Bank [2]	{(M4, Bank [2]), (M0, Bank [2])}
	Rank 0, Bank [3]	{(M4, Bank [3]), (M0, Bank [3]))
	Rank 0, Bank [4]	{(M6, Bank [0]), (M2, Bank [0]))
	Rank 0, Bank [5]	{(M6, Bank [1]), (M2, Bank [1]))
	Rank 0, Bank [6]	{(M6, Bank [2]), (M2, Bank [2]))
	Rank 0, Bank [7]	{(M6, Bank [3]), (M2, Bank [3]))
	Rank 1, Bank [0]	{(M5, Bank [0]), (M1, Bank [0]))
	Rank 1, Bank [1]	{(M5, Bank [1]), (M1, Bank [1]))
	Rank 1, Bank [2]	{(M5, Bank [2]), (M1, Bank [2]))
	Rank 1, Bank [3]	{(M5, Bank [3]), (M1, Bank [3]))
	Rank 1, Bank [4]	{(M7, Bank [0]), (M3, Bank [0]))
	Rank 1, Bank [5]	{(M7, Bank [1]), (M3, Bank [1]))
	Rank 1, Bank [6]	{(M7, Bank [2]), (M3, Bank [2]))
	Rank 1, Bank [7]	{(M7, Bank [3]), (M3, Bank [3]))

[0058] FIG. 9 illustrates an example of linear address mapping for use with a multi-rank buffer integrated circuit.

[0059] An example of a linear address mapping with a single-rank buffer integrated circuit is shown below.

[0060]

Host System Address	DRAM Device
(Virtual Bank)	(Physical Banks)
Rank 0, Bank [0]	{(M6, Bank [0]), (M4, Bank[0]), (M2, Bank [0]),
	(M0, Bank [0])}
Rank 0, Bank [1]	{(M6, Bank [1]), (M4, Bank[1]), (M2, Bank [1]),
	(M0, Bank [1])}
Rank 0, Bank [2]	{(M6, Bank [2]), (M4, Bank[2]), (M2, Bank [2]),
	(M0, Bank [2])}
Rank 0, Bank [3]	{(M6, Bank [3]), (M4, Bank[3]), (M2, Bank [3]),
	(M0, Bank [3])}
Rank 0, Bank [4]	{(M7, Bank [0]), (M5, Bank[0]), (M3, Bank [0]),
	(M1, Bank [0])}
Rank 0, Bank [5]	{(M7, Bank [1]), (M5, Bank[1]), (M3, Bank [1]),
	(M1, Bank [1])}
Rank 0, Bank [6]	{(M7, Bank [2]), (M5, Bank[2]), (M3, Bank [2]),
	(M1, Bank [2])}
Rank 0, Bank [7]	{(M7, Bank [3]), (M5, Bank[3]), (M3, Bank [3]),
	(M1, Bank [3])}

[0061] FIG. 10 illustrates an example of linear address mapping with a single rank buffer integrated circuit. Using this configuration, the stack of DRAM devices appears as a single 4 Gb integrated circuit with eight memory banks.

[0062] In another embodiment, the addresses from the host system may be mapped by the buffer integrated circuit such that one or more banks of the host system address (i.e. virtual banks) are mapped to a single physical DRAM integrated circuit in the stack (“bank slice” mapping). FIG. 11 illustrates an example of bank slice address mapping with a multi-rank buffer integrated circuit. Also, an example of a bank slice address mapping is shown below.

[0063]

	Host System Address	DRAM Device
	(Virtual Bank)	(Physical Bank)
	Rank 0, Bank [0]	M0, Bank [1:0]
	Rank 0, Bank [1]	M0, Bank [3:2]
	Rank 0, Bank [2]	M2, Bank [1:0]
	Rank 0, Bank [3]	M2, Bank [3:2]
	Rank 0, Bank [4]	M4, Bank [1:0]
	Rank 0, Bank [5]	M4, Bank [3:2]
	Rank 0, Bank [6]	M6, Bank [1:0]
	Rank 0, Bank [7]	M6, Bank [3:2]
	Rank 1, Bank [0]	M1, Bank [1:0]
	Rank 1, Bank [1]	M1, Bank [3:2]
	Rank 1, Bank [2]	M3, Bank [1:0]
	Rank 1, Bank [3]	M3, Bank [3:2]
	Rank 1, Bank [4]	M5, Bank [1:0]
	Rank 1, Bank [5]	M5, Bank [3:2]
	Rank 1, Bank [6]	M7, Bank [1:0]
	Rank 1, Bank [7]	M7, Bank [3:2]

[0064] The stack of this example contains eight 512 Mb DRAM integrated circuits, each with four memory banks In this example, a multi-rank buffer integrated circuit is assumed, which means that the host system sees the stack as two 2 Gb DRAM devices, each having eight banks.

[0065] FIG. 12 illustrates an example of bank slice address mapping with a single rank buffer integrated circuit. The bank slice mapping with a single-rank buffer integrated circuit is shown below.

[0066]

	Host System Address	DRAM Device
	(Virtual Bank)	(Physical Device)
	Rank 0, Bank [0]	M0
	Rank 0, Bank [1]	M1
	Rank 0, Bank [2]	M2
	Rank 0, Bank [3]	M3
	Rank 0, Bank [4]	M4
	Rank 0, Bank [5]	M5
	Rank 0, Bank [6]	M6
	Rank 0, Bank [7]	M7

[0067] The stack of this example contains eight 512 Mb DRAM devices so that the host system sees the stack as a single 4 Gb device with eight banks. The address mappings shown above are for illustrative purposes only. Other mappings may be implemented without deviating from the spirit and scope of the claims.

[0068] Bank slice address mapping enables the virtual DRAM to reduce or eliminate some timing constraints that are inherent in the underlying physical DRAM devices. For instance, the physical DRAM devices may have a tFAW (4 bank activate window) constraint that limits how frequently an activate operation may be targeted to a physical DRAM device. However, a virtual DRAM circuit that uses bank slice address mapping may not have this constraint. As an example, the address mapping in FIG. 11 maps two banks of the virtual DRAM device to a single physical DRAM device. So, the tFAW constraint is eliminated because the t_RC timing parameter prevents the host system from issuing more than two consecutive activate commands to any given physical DRAM device within a t_RC window (and t_RC>t_FAW). Similarly, a virtual DRAM device that uses the address mapping in FIG. 12 eliminates the t_RRD constraint of the underlying physical DRAM devices.

[0069] In addition, a bank slice address mapping scheme enables the buffer integrated circuit or the host system to power manage the DRAM devices on a DIMM on a more granular level. To illustrate this, consider a virtual DRAM device that uses the address mapping shown in FIG. 12, where each bank of the virtual DRAM device corresponds to a single physical DRAM device. So, when bank 0 of the virtual DRAM device (i.e. virtual bank 0) is accessed, the corresponding physical DRAM device M0 may be in the active mode. However, when there is no outstanding access to virtual bank 0, the buffer integrated circuit or the host system (or any other entity in the system) may place DRAM device M0 in a low power (e.g. power down) mode. While it is possible to place a physical DRAM device in a low power mode, it is not possible to place a bank (or portion) of a physical DRAM device in a low power mode while the remaining banks (or portions) of the DRAM device are in the active mode. However, a bank or set of banks of a virtual DRAM circuit may be placed in a low power mode while other banks of the virtual DRAM circuit are in the active mode since a plurality of physical DRAM devices are used to emulate a virtual DRAM device. It can be seen from FIG. 12 and FIG. 10, for example, that fewer virtual banks are mapped to a physical DRAM device with bank slice mapping (FIG. 12) than with linear mapping (FIG. 10). Thus, the likelihood that all the (physical) banks in a physical DRAM device are in the precharge state at any given time is higher with bank slice mapping than with linear mapping. Therefore, the buffer integrated circuit or the host system (or some other entity in the system) has more opportunities to place various physical DRAM devices in a low power mode when bank slide mapping is used.

[0070] In several market segments, it may be desirable to preserve the contents of main memory (usually, DRAM) either periodically or when certain events occur. For example, in the supercomputer market, it is common for the host system to periodically write the contents of main memory to the hard drive. That is, the host system creates periodic checkpoints. This method of checkpointing enables the system to re-start program execution from the last checkpoint instead of from the beginning in the event of a system crash. In other markets, it may be desirable for the contents of one or more address ranges to be periodically stored in non-volatile memory to protect against power failures or system crashes. All these features may be optionally implemented in a buffer integrated circuit disclosed herein by integrating one or more non-volatile memory integrated circuits (e.g. flash memory) into the stack. In some embodiments, the buffer integrated circuit is designed to interface with one or more stacks containing DRAM devices and non-volatile memory integrated circuits. Note that each of these stacks may contain only DRAM devices or contain only non-volatile memory integrated circuits or contain a mixture of DRAM and non-volatile memory integrated circuits.

[0071] FIGS. 13A and 13B illustrate examples of buffered stacks that contain both DRAM and non-volatile memory integrated circuits. A DIMM PCB 1300 includes a buffered stack (buffer 1310 and DRAMs 1320) and flash 1330. In another embodiment shown in FIG. 13B, DIMM PCB 1340 includes a buffered stack (buffer 1350, DRAMs 1360 and flash 1370). An optional non-buffered stack includes at least one non-volatile memory device (e.g., flash 1390) or DRAM device 1380. All the stacks that connect to a buffer integrated circuit may be on the same PCB as the buffer integrated circuit or some of the stacks may be on the same PCB while other stacks may be on another PCB that is electrically and mechanically coupled by means of a connector or an interposer to the PCB containing the buffer integrated circuit.

[0072] In some embodiments, the buffer integrated circuit copies some or all of the contents of the DRAM devices in the stacks that it interfaces with to the non-volatile memory integrated circuits in the stacks that it interfaces with. This event may be triggered, for example, by a command or signal from the host system to the buffer integrated circuit, by an external signal to the buffer integrated circuit, or upon the detection (by the buffer integrated circuit) of an event or a catastrophic condition like a power failure. As an example, let us assume that a buffer integrated circuit interfaces with a plurality of stacks that contain 4 Gb of DRAM memory and 4 Gb of non-volatile memory. The host system may periodically issue a command to the buffer integrated circuit to copy the contents of the DRAM memory to the non-volatile memory. That is, the host system periodically checkpoints the contents of the DRAM memory. In the event of a system crash, the contents of the DRAM may be restored upon re-boot by copying the contents of the non-volatile memory back to the DRAM memory. This provides the host system with the ability to periodically check point the memory.

[0073] In another embodiment, the buffer integrated circuit may monitor the power supply rails (i.e. voltage rails or voltage planes) and detect a catastrophic event, for example, a power supply failure. Upon detection of this event, the buffer integrated circuit may copy some or all the contents of the DRAM memory to the non-volatile memory. The host system may also provide a non-interruptible source of power to the buffer integrated circuit and the memory stacks for at least some period of time after the power supply failure to allow the buffer integrated circuit to copy some or all the contents of the DRAM memory to the non-volatile memory. In other embodiments, the memory module may have a built-in backup source of power for the buffer integrated circuits and the memory stacks in the event of a host system power supply failure. For example, the memory module may have a battery or a large capacitor and an isolation switch on the module itself to provide backup power to the buffer integrated circuits and the memory stacks in the event of a host system power supply failure.

[0074] A memory module, as described above, with a plurality of buffers, each of which interfaces to one or more stacks containing DRAM and non-volatile memory integrated circuits, may also be configured to provide instant-on capability. This may be accomplished by storing the operating system, other key software, and frequently used data in the non-volatile memory.

[0075] In the event of a system crash, the memory controller of the host system may not be able to supply all the necessary signals needed to maintain the contents of main memory. For example, the memory controller may not send periodic refresh commands to the main memory, thus causing the loss of data in the memory. The buffer integrated circuit may be designed to prevent such loss of data in the event of a system crash. In one embodiment, the buffer integrated circuit may monitor the state of the signals from the memory controller of the host system to detect a system crash. As an example, the buffer integrated circuit may be designed to detect a system crash if there has been no activity on the memory bus for a pre-determined or programmable amount of time or if the buffer integrated circuit receives an illegal or invalid command from the memory controller. Alternately, the buffer integrated circuit may monitor one or more signals that are asserted when a system error or system halt or system crash has occurred. For example, the buffer integrated circuit may monitor the HT_SyncFlood signal in an Opteron processor based system to detect a system error. When the buffer integrated circuit detects this event, it may de-couple the memory bus of the host system from the memory integrated circuits in the stack and internally generate the signals needed to preserve the contents of the memory integrated circuits until such time as the host system is operational. So, for example, upon detection of a system crash, the buffer integrated circuit may ignore the signals from the memory controller of the host system and instead generate legal combinations of signals like CKE, CS#, RAS#, CAS#, and WE# to maintain the data stored in the DRAM devices in the stack, and also generate periodic refresh signals for the DRAM integrated circuits. Note that there are many ways for the buffer integrated circuit to detect a system crash, and all these variations fall within the scope of the claims.

[0076] Placing a buffer integrated circuit between one or more stacks of memory integrated circuits and the host system allows the buffer integrated circuit to compensate for any skews or timing variations in the signals from the host system to the memory integrated circuits and from the memory integrated circuits to the host system. For example, at higher speeds of operation of the memory bus, the trace lengths of signals between the memory controller of the host system and the memory integrated circuits are often matched. Trace length matching is challenging especially in small form factor systems. Also, DRAM processes do not readily lend themselves to the design of high speed I/O circuits. Consequently, it is often difficult to align the I/O signals of the DRAM integrated circuits with each other and with the associated data strobe and clock signals.

[0077] In one embodiment of a buffer integrated circuit, circuitry that adjusts the timing of the I/O signals may be incorporated. In other words, the buffer integrated circuit may have the ability to do per-pin timing calibration to compensate for skews or timing variations in the I/0 signals. For example, say that, the DQ[0] data signal between the buffer integrated circuit and the memory controller has a shorter trace length or has a smaller capacitive load than the other data signals, DQ[7:1]. This results in a skew in the data signals since not all the signals arrive at the buffer integrated circuit (during a memory write) or at the memory controller (during a memory read) at the same time. When left uncompensated, such skews tend to limit the maximum frequency of operation of the memory sub-system of the host system. By incorporating per-pin timing calibration and compensation circuits into the I/0 circuits of the buffer integrated circuit, the DQ[0] signal may be driven later than the other data signals by the buffer integrated circuit (during a memory read) to compensate for the shorter trace length of the DQ[O] signal. Similarly, the per-pin timing calibration and compensation circuits allow the buffer integrated circuit to delay the DQ[O] data signal such that all the data signals, DQ[7:0], are aligned for sampling during a memory write operation. The per-pin timing calibration and compensation circuits also allow the buffer integrated circuit to compensate for timing variations in the I/O pins of the DRAM devices. A specific pattern or sequence may be used by the buffer integrated circuit to perform the per-pin timing calibration of the signals that connect to the memory controller of the host system and the per-pin timing calibration of the signals that connect to the memory devices in the stack.

[0078] Incorporating per-pin timing calibration and compensation circuits into the buffer integrated circuit also enables the buffer integrated circuit to gang a plurality of slower DRAM devices to emulate a higher speed DRAM integrated circuit to the host system. That is, incorporating per-pin timing calibration and compensation circuits into the buffer integrated circuit also enables the buffer integrated circuit to gang a plurality of DRAM devices operating at a first clock speed and emulate to the host system one or more DRAM integrated circuits operating at a second clock speed, wherein the first clock speed is slower than the second clock speed.

[0079] For example, the buffer integrated circuit may operate two 8-bit wide DDR2 SDRAM devices-in parallel at a 533 MHz data rate such that the host system sees a single 8-bit wide DDR2 SDRAM integrated circuit that operates at a 1066 MHz data rate. Since, in this example, the two DRAM devices are DDR2 devices, they are designed to transmit or receive four data bits on each data pin for a memory read or write respectively (for a burst length of 4). So, the two DRAM devices operating in parallel may transmit or receive sixty four bits per data pin per memory read or write respectively in this example. Since the host system sees a single DDR2 integrated circuit behind the buffer, it will only receive or transmit thirty-two data bits per pin per memory read or write respectively. In order to accommodate for the different data widths, the buffer integrated circuit may make use of the DM signal (Data Mask). Say that the host system sends DA[7:0], DB[7:0], DC[7:0], and DD[7:0] to the buffer integrated circuit at a 1066 MHz data rate. The buffer integrated circuit may send DA[7:0], DC[7:0], XX, and XX to the first DDR2 SDRAM integrated circuit and send DB[7:0], DD[7:0], XX, and XX to the second DDR2 SDRAM integrated circuit, where XX denotes data that is masked by the assertion (by the buffer integrated circuit) of the DM inputs to the DDR2 SDRAM integrated circuits.

[0080] In another embodiment, the buffer integrated circuit operates two slower DRAM devices as a single, higher-speed, wider DRAM. To illustrate, the buffer integrated circuit may operate two 8-bit wide DDR2 SDRAM devices running at 533 MHz data rate such that the host system sees a single 16-bit wide DDR2 SDRAM integrated circuit operating at a 1066 MHz data rate. In this embodiment, the buffer integrated circuit may not use the DM signals. In another embodiment, the buffer integrated circuit may be designed to operate two DDR2 SDRAM devices (in this example, 8-bit wide, 533 MHz data rate integrated circuits) in parallel, such that the host system sees a single DDR3 SDRAM integrated circuit (in this example, an 8-bit wide, 1066 MHz data rate, DDR3 device). In another embodiment, the buffer integrated circuit may provide an interface to the host system that is narrower and faster than the interface to the DRAM integrated circuit. For example, the buffer integrated circuit may have a 16-bit wide, 533 MHz data rate interface to one or more DRAM devices but have an 8-bit wide, 1066 MHz data rate interface to the host system.

[0081] In addition to per-pin timing calibration and compensation capability, circuitry to control the slew rate (i.e. the rise and fall times), pull-up capability or strength, and pull-down capability or strength may be added to each I/O pin of the buffer integrated circuit or optionally, in common to a group of I/O pins of the buffer integrated circuit. The output drivers and the input receivers of the buffer integrated circuit may have the ability to do pre-emphasis in order to compensate for non-uniformities in the traces connecting the buffer integrated circuit to the host system and to the memory integrated circuits in the stack, as well as to compensate for the characteristics of the I/O pins of the host system and the memory integrated circuits in the stack.

[0082] Stacking a plurality of memory integrated circuits (both volatile and non-volatile) has associated thermal and power delivery characteristics. Since it is quite possible that all the memory integrated circuits in a stack may be in the active mode for extended periods of time, the power dissipated by all these integrated circuits may cause an increase in the ambient, case, and junction temperatures of the memory integrated circuits. Higher junction temperatures typically have negative impact on the operation of ICs in general and DRAMs in particular. Also, when a plurality of DRAM devices are stacked on top of each other such that they share voltage and ground rails (i.e. power and ground traces or planes), any simultaneous operation of the integrated circuits may cause large spikes in the voltage and ground rails. For example, a large current may be drawn from the voltage rail when all the DRAM devices in a stack are refreshed simultaneously, thus causing a significant disturbance (or spike) in the voltage and ground rails. Noisy voltage and ground rails affect the operation of the DRAM devices especially at high speeds. In order to address both these phenomena, several inventive techniques are disclosed below.

[0083] One embodiment uses a stacking technique wherein one or more layers of the stack have decoupling capacitors rather than memory integrated circuits. For example, every fifth layer in the stack may be a power supply decoupling layer (with the other four layers containing memory integrated circuits). The layers that contain memory integrated circuits are designed with more power and ground balls or pins than are present in the pin out of the memory integrated circuits. These extra power and ground balls are preferably disposed along all the edges of the layers of the stack.

[0084] FIGS. 14A, 14B and 14C illustrate one embodiment of a buffered stack with power decoupling layers. As shown in FIG. 14A, DIMM PCB 1400 includes a buffered stack of DRAMs including decoupling layers. Specifically, for this embodiment, the buffered stack includes buffer 1410, a first set of DRAM devices 1420, a first decoupling layer 1430, a second set of DRAM devices 1440, and an optional second decoupling layer 1450. The stack also has an optional heat sink or spreader 1455.

[0085] FIG. 14B illustrates top and side views of one embodiment for a DRAM die. A DRAM die 1460 includes a package (stack layer) 1466 with signal/power/GND balls 1462 and one or more extra power/GND balls 1464. The extra power/GND balls 1464 increase thermal conductivity.

[0086] FIG. 14C illustrates top and side views of one embodiment of a decoupling layer. A decoupling layer 1475 includes one or more decoupling capacitors 1470, signal/power/GND balls 1485, and one or more extra power/GND balls 1480. The extra power/GND balls 1480 increases thermal conductivity.

[0087] The extra power and ground balls, shown in FIGS. 14B and 14C, form thermal conductive paths between the memory integrated circuits and the PCB containing the stacks, and between the memory integrated circuits and optional heat sinks or heat spreaders. The decoupling capacitors in the power supply decoupling layer connect to the relevant power and ground pins in order to provide quiet voltage and ground rails to the memory devices in the stack. The stacking technique described above is one method of providing quiet power and ground rails to the memory integrated circuits of the stack and also to conduct heat away from the memory integrated circuits.

[0088] In another embodiment, the noise on the power and ground rails may be reduced by preventing the DRAM integrated circuits in the stack from performing an operation simultaneously. As mentioned previously, a large amount of current will be drawn from the power rails if all the DRAM integrated circuits in a stack perform a refresh operation simultaneously. The buffer integrated circuit may be designed to stagger or spread out the refresh commands to the DRAM integrated circuits in the stack such that the peak current drawn from the power rails is reduced. For example, consider a stack with four 1 Gb DDR2 SDRAM integrated circuits that are emulated by the buffer integrated circuit to appear as a single 4 Gb DDR2 SDRAM integrated circuit to the host system. The JEDEC specification provides for a refresh cycle time (i.e. t_RFC) of 400 ns for a 4 Gb DRAM integrated circuit while a 1 Gb DRAM integrated circuit has a t_RFC specification of 110 ns. So, when the host system issues a refresh command to the emulated 4 Gb DRAM integrated circuit, it expects the refresh to be done in 400 ns. However, since the stack contains four 1 Gb DRAM integrated circuits, the buffer integrated circuit may issue separate refresh commands to each of the 1 Gb DRAM integrated circuit in the stack at staggered intervals. As an example, upon receipt of the refresh command from the host system, the buffer integrated circuit may issue a refresh command to two of the four 1 Gb DRAM integrated circuits, and 200 ns later, issue a separate refresh command to the remaining two 1 Gb DRAM integrated circuits. Since the 1 Gb DRAM integrated circuits require 110 ns to perform the refresh operation, all four 1 Gb DRAM integrated circuits in the stack will have performed the refresh operation before the 400 ns refresh cycle time (of the 4 Gb DRAM integrated circuit) expires. This staggered refresh operation limits the maximum current that may be drawn from the power rails. It should be noted that other implementations that provide the same benefits are also possible, and are covered by the scope of the claims.

[0089] In one embodiment, a device for measuring the ambient, case, or junction temperature of the memory integrated circuits (e.g. a thermal diode) can be embedded into the stack. Optionally, the buffer integrated circuit associated with a given stack may monitor the temperature of the memory integrated circuits. When the temperature exceeds a limit, the buffer integrated circuit may take suitable action to prevent the overheating of and possible damage to the memory integrated circuits. The measured temperature may optionally be made available to the host system.

[0090] Other features may be added to the buffer integrated circuit so as to provide optional features. For example, the buffer integrated circuit may be designed to check for memory errors or faults either on power up or when the host system instructs it do so. During the memory check, the buffer integrated circuit may write one or more patterns to the memory integrated circuits in the stack, read the contents back, and compare the data read back with the written data to check for stuck-at faults or other memory faults.

[0091] 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 claims.