TECHNICAL FIELD

[001] Embodiments of the invention relate to memory devices, and, more particularly, in one or more embodiments to a memory device that can be operated in either a direct mode, in which conventional memory control signals are coupled to the memory devices, or an indirect mode, in which command packets are coupled to the memory devices.

BACKGROUND OF THE INVENTION

[002] As memory devices of all types have evolved, continuous strides have been made in improving their performance in a variety of respects. For example, the storage capacity of memory devices has continued to increase at geometric proportions. This increased capacity, coupled with the geometrically higher operating speeds of electronic systems containing memory devices, has made high memory device bandwidth ever more critical. One application in which memory devices, such as dynamic random access memory ("DRAM") devices, require a higher bandwidth is their use as system memory in computer systems. As the operating speed of processors has increased, processors are able to read and write data at correspondingly higher speeds. Yet conventional DRAM devices often do not have the bandwidth to read and write data at these higher speeds, thereby slowing the performance of conventional computer systems. This problem is exacerbated by the trend toward multi-core processors and multiple processor computer systems. It is currently estimated that computer systems operating as high-end servers are idle as many as 3 out of every 4 clock cycles because of the limited data bandwidth of system memory devices. In fact, the limited bandwidth of DRAM devices operating as system memory can reduce the performance of computer systems to as low as 10% of the performance of which they would otherwise be capable. [003] Various attempts have been made to increase the data bandwidth of memory devices. For example, wider internal data buses have been used to transfer data to and from arrays with a higher bandwidth. However, doing so usually requires that write data be serialized and read data deserialized at the memory device interface. Another approach has been to simply scale up the size of memory devices or conversely shrink their feature sizes, but, for a variety of reasons, scaling has been incapable of keeping up with the geometric increase in the demand for higher data bandwidths.

[004] More recently, proposals have also been made to stack several integrated circuit memory devices in the same package, but doing so threatens to create a large number of other problems to be overcome. These problems can be solved to a large extent by connecting the stack of interconnected memory devices to a logic die on which the memory devices are stacked. The logic die can then serve as a high-speed interface to the memory devices. However, taking advantage of the increased capabilities of this arrangement is more easily achieved if memory command and address signals are placed in a packet and coupled to the logic die through a high-speed bus. Yet many computer and other systems are designed to interface with memory devices using conventional memory command signals and conventional row and column address signals. Advanced memory systems formed by stacking memory devices on a logic die would therefore be unusable with such systems. However, memory device manufacturers generally desire to standardize their product offerings to the greatest extent possible to lessen the number of different memory devices that are manufactured, marketed, etc.

[005] Therefore, a need exists for a method and system to allow advanced memory system formed by stacking interconnected memory device dice to be interfaced with systems by either using conventional memory commands and addresses or by using packets containing commands and addresses. BRIEF DESCRIPTION OF THE DRAWINGS

[006] Figure 1 is a block diagram of a computer system that includes a dual mode memory system according to an embodiment of the invention. [007] Figure 2 is a block diagram of a dual mode memory system according to an embodiment of the invention. [008] Figure 3 is a more detailed block diagram of a dual mode memory system according to an embodiment of the invention. [009] Figure 4 is a packet diagram showing the format of a downstream packet that can be coupled to the memory system of Figures 1 , 2 or 3 or a memory system according to some other embodiment of the invention for the indirect operating mode. [010] Figure 5 is a chart showing how the commands in the first field of the downstream packet of Figure 4 are modified for the direct operating mode. [011] Figure 6 is a chart showing the commands in the downstream packet of

Figure 4 for the indirect operating mode. [012] Figure 7 is a packet diagram showing the format of an upstream packet that can be coupled from the memory system of Figures 1, 2 or 3 or a memory system according to some other embodiment of the invention. [013] Figure 8 is a chart showing the commands in the upstream packet of

Figure 7 for the indirect operating mode.

DETAILED DESCRIPTION

[014] A computer system including a high-capacity, high bandwidth memory device 10 according to an embodiment of the invention is shown in Figure 1 connected to a processor 12 through a relatively narrow high-speed bus 14 that may be divided into downstream lanes and separate upstream lanes (not shown in Figure 1). The memory device 10 includes 4 DRAM die 20, 22, 24, 26, which may be identical to each other, stacked on top of each other. Although the memory device 10 includes 4 DRAM die 20, 22, 24, 26, other embodiments of the memory device use a greater or lesser number of DRAM die. The DRAM die 20, 22, 24, 26 are stacked with (e.g., on top of) a logic die 30, which serves as the interface with the processor 12. The logic die 30 can implement a variety of functions in the memory device 10, such as to limit the number of functions that are be implemented in the DRAM die 20, 22, 24, 26. For example, the logic die 30 may perform memory management functions, such as power management and refresh of memory cells in the DRAM die 20, 22, 24, 26. In some embodiments, the logic die 30 may implement test and/or repair capabilities, and it may perform error checking and correcting ("ECC") functions.

[015] The DRAM die 20, 22, 24, 26 are connected to each other and to the logic die 30 by a relatively wide bus 34. The bus 34 may be implemented with through silicon vias ("TSVs"), which comprise a large number of conductors extending through the DRAM die 20, 22, 24, 26 at the same locations on the DRAM die and connect to respective conductors formed on the die 20, 22, 24, 26. In one embodiment, each of the DRAM die 20, 22, 24, 26 are divided into 16 autonomous partitions, each of which may contain 2 or 4 independent memory banks. In such case, the partitions of each die 20, 22, 24, 26 that are stacked on top of each other may be independently accessed for read and write operations. Each set of 16 stacked partitions may be referred to as a "vault." Thus, the memory device 10 may contain 16 vaults.

[016] As shown in Figure 2, in one embodiment, the bus 34 may be divided into 16 36-bit bi-directional sub-buses 38a-p, with each of the 16 36-bit sub-buses coupled to the 4 partitions in a respective vault. Each of these sub-buses couples 32 bits of a data and 4 ECC bits between the logic die 30 and the DRAM die 20, 22, 24, 26. However, the number of stacked DRAM die 20, 22, 24, 26, the number of partitions in each DRAM die, the number of banks in each partition, and the number of bits in each of the sub-buses 38a-p can vary as desired. The relatively narrow high-speed bus 14 connecting the processor 12 to the logic die may be divided into 4 16-bit downstream lanes 40a-d and 4 separate 16-bit upstream lanes 42a-d. The 4 downstream lanes 40a-d may be connected to a single processor 12 as shown in Figure 1, which may be a multi-core processor, to multiple processors (not shown), or to some other memory access device like a memory controller. The 4 downstream lanes 40a-d may operate independently of each other so that packets (in the indirect mode) or memory command, address, and data signals (in the direct mode) are coupled through the lanes 40a-d at different times and to the same or different vaults.

[017] As explained in greater detail below, one of the functions performed by the logic die 30 can be to serialize the read data bits coupled from the DRAM die 20, 22, 24, 26 into a serial stream of 16 serial data bits coupled through 16 parallel bits of each upstream lane 42a-d of the bus 14. Similarly, the logic die 30 may perform the functions of deserializing 16 serial data bits coupled through one of the 16-bit downstream lanes 40a-d of the bus 14 to obtain 256 parallel data bits. The logic die 30 then couples these 256 bits through one of the 32-bit sub-buses 38a-p in a serial stream of 8 bits. However, other embodiments may use different numbers of lanes 40, 42 having different widths or different numbers of sub-buses 38a-p having different widths, and they may couple data bits having different structures. As will be appreciated by one skilled in the art, the stacking of multiple DRAM die results in a memory device having a very large capacity. Further, the use of a very wide bus connecting the DRAM die allows data to be coupled to and from the DRAM die with a very high bandwidth.

[018] A logic die 30 according to an embodiment of the invention is shown in

Figure 3 connected to the processor 12 and the DRAM die 20, 22, 24, 26. As shown in Figure 3, each of the 4 downstream lanes 40a-d may be connected to a respective link interface 50a-d. Each link interface 50a-d includes a deserializer 54 that converts each serial stream of 16 data bits on each of the 16-bit lanes 40a-d to 256 parallel bits. Insofar as there are 4 link interfaces 50a-d, the link interfaces can together output 1024 output parallel bits.

[019] Each of the link interfaces 50a-d applies its 256 parallel bits to a respective downstream target 60a-d, which decodes the command and address portions of the received packet (in the indirect mode) or the commands and addresses (in the direct mode) and buffers write data in the event a memory request is for a write operation. The downstream targets 60a-d output their respective commands, addresses and possibly write data to a switch 62. The switch 62 contains 16 multiplexers 64 each of which direct the command, addresses and any write data from any of the downstream targets 60a-d to its respective vault of the DRAM die 20, 22, 24, 26. Thus, each of the downstream targets 60a-d can access any of the 16 vaults in the DRAM die 20, 22, 24, 26. The multiplexers 64 use the address in the received memory requests to determine if its respective vault is the target of a memory request. Each of the multiplexers 64 apply the memory request to a respective one of 16 vault controllers 70a-p.

[020] Each vault controller 70a-p includes a respective memory controller 80, each of which includes a write buffer 82, a read buffer 84 and a command pipeline 86. The commands and addresses in memory requests received from the switch 62 are loaded into the command pipeline 86, which subsequently outputs the received commands and corresponding addresses. Any write data in the memory requests are stored in the write buffer 82. The read buffer 84 may be used to store read data from the respective vault, as will be explained in greater detail below. The write data from the write buffer 82 are applied to a memory interface 88.

[021] According to an embodiment of the invention, the commands and addresses from the command pipeline 86 are applied to a memory interface 88 through a command processing circuit, such as a command register 90. The command register 90 can be a free running interface register. In the direct mode, the commands and addresses from the command pipeline are applied to the memory interface 88. These commands and addressed may be applied to the memory interface 88 as they are received by the memory device 10. In the indirect mode, the command register 90 creates the commands and addresses and sends it to the memory interface 88. The command register 90 includes a sequencer (not shown) that transmits the commands and addresses to the memory interface in the proper order and at the proper times.

[022] The memory interface 88 couples the received command and address signals from the command register 90 to the DRAM die 20, 22, 24, 26 through a command/address bus 92. The memory interface 88 also couples 32-bits of write data from the write buffer 82. In some embodiments, the memory interface 88 may include an ECC system (not shown), which uses ECC techniques to check and correct the data read from the DRAM die 20, 22, 24, 26. hi such case, in addition to coupling write data to the DRAM die 20, 22, 24, 26, the memory interface 88 couples 4 bits of ECC from the ECC system to the DRAM die 20, 22, 24, 26 through a 36-bit data bus 94.

[023] Although write data are loaded into the write buffer 82 as 256 parallel bits, they are output from the buffer 82 in two sets, each set being 128 parallel bits. These 128 bits may then be further serialized by the ECC system (not shown) to 4 sets of 32-bit data, which are coupled through the data bus 94. In the embodiment shown in Figure 3, write data are coupled to the write buffer 82 in synchronism with a 500 MHz clock so the data are stored in the write buffer at 16 gigabytes ("GB") per second. The write data are coupled from the write buffer 82 to the DRAM die 20, 22, 24, 26 using a 2 GHz clock so the data are output from the write buffer 82 at 8 GB/s. Therefore, as long as more than half of the memory requests are not write operations to the same vault, the write buffers 82 will be able to couple the write data to the DRAM die 20, 22, 24, 26 at least as fast as the data are coupled to the write buffer 82.

[024] In the event a memory request is for a read operation, the command and address for the request are coupled to the DRAM die 20, 22, 24, 26 in the same manner as a write request, as explained above. In response to a read request, 32 bits of read data and 4 ECC bits are output from the DRAM die 20, 22, 24, 26 through the 36-bit data bus 94. The ECC bits are passed to the ECC system (not shown), which uses the ECC bits to check and correct the read data before passing the read data on to the read buffer 84. The ECC system also deserializes the 32 bits of read data into two sets of 128-bit read data. However, in some embodiments, the memory system does not include the ECC system.

[025] After 2 sets of 128-bit read data have been stored in the read buffer 84, the read buffer transmits 256 bits to the switch 62. The switch includes 4 output multiplexers 104 coupled to respective upstream masters 110a-d. Each multiplexer 104 can couple 256 bits of parallel data from any one of the vault controllers 70a-p to its respective upstream master 110a-d. The upstream masters 110a-d format the 256 bits of read data into packet data (in the indirect mode) and couple the packet to respective upstream link interfaces 114a-d. In the direct mode, the read data are simply coupled to respective upstream link interfaces 114a-d. Each of the link interfaces 114a-d include a respective serializer 120 that converts the incoming 256 bits to a serial stream of 16 bits on each bit of a respective one of the 16-bit upstream links 42a-d.

[026] The format of a downstream packet 150 that can be coupled to the memory system of Figures 1, 2 or 3 or a memory system according to some other embodiment of the invention is shown in Figure 4. The downstream packet 150 may be, as explained above, 32 bits wide, and it contains a first field 152. In the indirect operating mode, the first field 152 includes a 4-bit command 156 ("Cmd 3:0"), and 28 bits of an upper address 158 ("U Address"). The nature of the command 156 and upper address 158 will be described in connection with Figure 6.

[027] As shown in Figure 5, in the direct mode, the first field 152 of the downstream packet 150 may be modified to allow a memory access device to directly access the DRAM die 20, 22, 24, 26. The first bit of the first field 152 may be a row address strobe ("RAS") signal 160, the second bit may be a column address strobe ("CAS") signal 162 and the third bit may be a write enable ("WE") signal 164. The first field 152 also includes a 4-bit column address 166 and a 14-bit row address 168. Finally, the first field 152 includes a four bit vault address 170. The vault address 170 specifies which of the 16 vaults are being accessed.

[028] Returning to Figure 4, the downstream packet 150 also contains a second field 180, which may be used in the indirect operating mode. The second field 180 contains a first group of 8 bits 182 that include 3 -bit command extension ("Cmd Ext") and 5 bits of a lower address ("LAddress"). As subsequently explained, the Cmd Ext bits 182 are used to further define commands designated by the four command bits 156. The next eight bits 184 of the second field 180 are reserved. The next eight bits 186 include 2 reserved bits ("RSV") and 6 header error checking and correcting bits ("HCRC"), which allow errors in the first field 152 to be detected and possibly corrected. A final eight bit ^' s 188 of the second field 180 are tag bits ("Tag") which uniquely identifies each memory request. As explained in greater detail below, these Tag bits 188 are included in upstream packets containing read data so that the memory request to which the read data corresponds can be identified, for example. Also, including these Tag bits 188 in an upstream packet for a write allows the writing of data to be acknowledged in an upstream packet, as will be subsequently explained.

[029] The downstream packet 150 also contains a third field 190, which includes a mask bit 192 that specifies whether a write will be masked, and 31 bits of write data 196. Following the third field 190 are one or more fields of write data 200. A final field contains a set of error checking bits 210, which may be cyclic redundancy check ("CRC") bits, ECC bits or some other type of error checking bits. The error checking bits 210 correspond to the write data to allow the memory system to determine if there were any errors in the transmission of the write data. In the case where the error checking bits are ECC bits and the number of errors is not too great, the bits 210 may allow errors in the write data to be corrected.

[030] Potential commands corresponding to the 4 command bits 156 in the first field 152 are shown in Figure 6 for the indirect mode. For the direct mode, the memory commands are formed by combinations of the WE, CAD and RAS signals shown in Figure 5. As shown in Figure 6, Cmd "0000" is for a no operation ("NOP") command, which does not cause the memory system 10 to perform any memory access. The command "0001" is decoded as a read command, with the number of bytes in the read being designated by the command extension bits 182. The command "0100" is decoded as a write command, with the number of bytes being written again by the command extension bits 182. Finally, the command "0101" is decoded as a masked write command, with the number of bytes also being written by the command extension bits 182. The remaining commands in the Cmd bits 156 are reserved for implementing additional functions.

[031] With further reference to Figure 6, the 28-bit upper address 158 and the

5-bit lower address in the bit group 182 specify the location in the memory system that is the subject of the memory request. ^" The logic die 30 uses these address bits to route a memory request to the corresponding vault and the corresponding row and column address in that vault. As mentioned above, the command extension "Cmd Ext" in the group 182 specifies the number of bytes that are read or written for a read and write or a masked write. If the command 156 in the first field 152 was for a read, the command extensions "Oi l" through "111" designate a read request of 8 through 128 bytes. The remaining command extensions are used for implementing additional functions. If the command 156 in the first field 152 was for a write, the command extensions "011" through "111" similarly designate a write request of 8 through 128 bytes. Finally, if the command 156 in the first field 152 was for a masked write, the command extensions "011" through "111" designate a masked write request of 8 through 128 bytes. The remaining command extensions are used for implementing additional functions.

[032] As also shown in Figure 6, the 6 error checking bits "HCRC" in the group 186 detects whether the data in the first field 152 contains an error. The final 8- bit tag 188 uniquely identifies each memory request, as previously explained.

[033] The format of an upstream packet 250 is shown in Figure 7. A first field

260 of the upstream packet 250 includes a 4-bit command ("CmdO") 262, and 2 error checking bits 264. Next are 2 reserved bits 266 followed by the 8-bit tag 268 ("TagO"), which, as previously explained, corresponds to the tag in the downstream packet 150 to which the read data is responsive. The first field 260 also contains a second set of the above-described bits, namely a 4-bit command ("Cmdl") 272, and 2 error checking bits 274. These error checking bits 274, along with the 2 error checking bits 264, allow detection and possibly correction of errors in the 32 bits of the first field 260. The first field 260 also contains 2 reserved bits 276, and an 8-bit tag 278 ("Tagl"). The upstream packet 250 normally does not include read data for two memory requests. However, the ability to include a second tag 278 and command 272, etc. in the first field 260 allows a write request to be acknowledged in the same upstream packet 250 as an upstream packet containing read data and an associated tag. Following the first field 260 are one or more 32-bit fields 280 of read data and a 32-bit field 290 of error checking bits. These error checking bits allow a memory controller or other memory access device receiving the read data to check for and possibly correct any transmission errors in the read data.

[034] The commands corresponding to the Cmd bits 262, 272 in the upstream packet 250 are shown in Figure 8. The 4-bit command "CmdO" 262 corresponds to a read if the upstream packet 250 is to contain read data. Again, the command "0000" is for a no operation "NOP" command. The next command "0001" is a naked command ("NAK") that acknowledges a read memory request but indicates that the data could not be read because of an error. The command "0100" acknowledges a prior write request, and the command "0101" is a naked command that acknowledges a prior write request but indicates that the write data was in error. The commands "1011" through "1111" indicates the upstream packet 250 contains read data of 8, 16, 32, 64 or 128 bytes, respectively. The remaining commands of "CmdO" are reserved for implementing other features.

[035] The commands corresponding to the Cmdl bits 272 are also shown in

Figure 8. The command "0000" is again for a no operation "NOP" command, and the command "0001" is again a naked command ("NAK") that acknowledges a read memory request but indicates that the data could not be read because of an error. The command "0100" acknowledges a prior write request, and the command "0101" is a naked command that acknowledges a prior write request but indicates that the write data was in error. The remaining commands of "Cmdl" are reserved for implementing other features.

[036] From the foregoing it will be appreciated that, although specific embodiments of the invention have been described herein for purposes of illustration, various modifications may be made without deviating from the spirit and scope of the invention. For example, although the embodiments of the invention are explained in the context of stacked DRAM die, it will be understood that the stacked die may be other types of memory device dice, such as flash memory device dice. Accordingly, the invention is not limited except as by the appended claims.