The disclosure relates to memory devices. More specifically, the disclosure relates to memory devices formed using three dimensional die stacks.
Stacking memory chips is conventionally used to increase the capacity of a memory device while reducing its silicon footprint. Two conventional methods of stacking are Package-on-Package (PoP) and System-in-Package (SiP). In a PoP system, discrete logic and memory ball grid array (BGA) are vertically combined in a package. The two packages rest on top of one another and are connected with a standard interface that routes signals between the two packages. In a SiP implementation, a number of dies are vertically stacked and are connected using conventional off-chip wire bonds or solder bumps.
Recently 3D integrated circuits (ICs) that utilize through silicon vias (TSVs) for interconnection have been developed as an improved alternative to the PoP and SiP packages. TSV technology utilizes vertical vias in the silicon (or other dielectric material) wafers that are used to interconnect each of the chips. Using through silicon vias results in shortened interconnect length, improved electrical performance, and reduced power consumption by the memory device.
TSV technology has been incorporated into memory storage devices that conform to conventional standards such as DDR2 and DDR3 SDRAM. To create a 1 gigabit DRAM, eight 128 Mb chips may be stacked on top of one another and are connected using through silicon vias. Although stacked vertically, a 3D IC memory device reads and writes data in accordance with the conventional memory standards, e.g., DDR2 and DDR3. For example a DDR2 SDRAM circuit has a prefetch buffer that is 4 bits deep and accesses data storage locations using a multiplexer. For DDR2 SDRAM, DDR2 memory cells transfer data both on the rising and falling edge of the system clock to enable four bits of data to be transferred per memory cell cycle. DDR3 SDRAM has a higher bandwidth than DDR2 and has the ability to transfer data at eight times the speed of the memory cells using an 8-bit prefetch buffer.
Although TSV technology has been used to increase the data storage capacity of a memory device, the speed at which the device may be read from and written to is constrained by the specification to which they conform, e.g., DDR2 and DDR3, and the bandwidth of the devices remains unchanged.
In one embodiment, a system comprises a central processing unit (CPU), a memory device in communication with the CPU, and a direct memory access (DMA) controller in communication with the CPU and the memory device. The memory device includes a plurality of vertically stacked chips and a plurality of input/output (I/O) ports. Each of the I/O ports is connected to at least one of the plurality of chips through a through substrate via. The DMA controller is configured to manage to transfer of data to and from the memory device.
In one embodiment, a memory system includes a storage device and a controller connected to the storage device. The storage device comprises first and second integrated circuit chips. Each of the first and second integrated circuit chips comprises a plurality of memory locations and through silicon vias. Each of the through silicon vias corresponds to a respective input/output (I/O) port. The controller is configured to manage writing data to and reading data from each of the memory locations of the first and second integrated circuit chips.
As used herein, the terms “through-silicon via” (TSV) and “though-substrate via are used interchangeably to refer to a configuration including a through-via penetrating an integrated circuit (IC) semiconductor substrate, and are not limited to ICs formed on substrates of silicon material. Thus, the term TSV as used herein can also encompass a through-substrate via that is formed through a different semiconductor IC substrate material, such as a III-V compound substrate, a silicon/germanium (SiGe) substrate, a gallium arsenide (GaAs) substrate, a silicon-on-insulator (SOI) substrate, or the like.
A novel approach for high bandwidth memory die stacking is now disclosed. In the following detailed description, reference is made to the accompanying drawings which form a part of the detailed description.
CPU 102 may be any processor that may performs computing functions. Examples of such processors include, but are not limited to, “INTEL® CORE”™, “PENTIUM”®, “CELERON”®, or “XEON”® processors available from Intel of Santa Clara, Calif., as well as an AMD “PHENOM”™, “ATHLON”™, or “SEMPRON”™ processors from AMD of Sunnyvale, Calif. The CPU 102 is connected to ROM 104, main memory 200, I/O devices 108, and DMA controller 300 through system bus 106.
System bus 106 may include a data bus used to transfer data from one of the memory devices 104, 200 to the CPU 102 or to an I/O device, an address bus used to transfer the source and destination address of the data, and a control bus used to transmit signals controlling how the data are being transmitted. System bus 106 may also include a power bus and I/O bus. The busses that comprise the system bus 106 are not shown to simplify the figures. In one embodiment, the width of system bus 106 may be 64-bits; however, other bus widths may be used.
ROM 104 may be any read only memory including, but not limited to, programmable read only memory (“PROM”), erasable programmable read only memory (“EPROM”), electrically erasable programmable read only memory (“EEPROM”), and flash memory.
Memory chips 202 are connected to one another using through-silicon via (“TSV”) technology. For example, each of the chips 202 may be connected to each other using laser-cut holes which are filled with a conductive metal as shown in
As shown in
The stack of IC memory chips 202a-202d are connected to the semiconductor substrate 212 by one or more solder bumps 210. Solder bumps 210 may be formed from lead or lead-free alloys. Examples of lead-free alloys include, but are not limited to, tin/silver, tin/copper/silver, copper, copper alloys, and the like.
The stacked IC memory chips 202a-202d may be separated from adjacent chips by spacers 206a-206d. For example, as shown in
A through-substrate via 216 is formed through each of the IC memory chips 202a-202d and spacers 206a-206d. The spacers are filled with a conductive material such as copper to form interconnects 208. Vertically stacking and connecting the chips using TSV technology improves the electrical performance and power consumption of the chips by reducing the length of the leads. Additionally, vertical stacking of the chips and interconnecting them using through-substrate vias enables the number of input/output (“I/O”) ports to be increased. The number of I/O ports may be increased because instead of using wire bonding that requires horizontal spacing on the package boards that is hundreds of microns wide to connect the chips, the chips may be connected using laser-cut holes on the order of one micron in width. Thus, connecting the chips using TSV eliminates the need for gaps of extra space for connecting the chips. Increasing the number of I/O ports on a chip increases enables the bandwidth of the chip to be increased as more I/O ports may be used. In some embodiments, each of the through-substrate vias 216 corresponds to an I/O port of the memory device 200.
Data are stored in the memory locations 206 on each of the memory chips 202. The data in memory chips 202 may be read from or written to the memory locations 206 using direct memory access (“DMA”) controlled by controller 300. Utilizing DMA enables the data to be accessed independently of the system clock enabling higher data transfer rates than may be accomplished in conventional DDR2 or DDR3 systems. Additionally, using DMA to access data stored in the memory 200 enlarges the bandwidth through which the data may be accessed as larger amounts of data may be transferred than through conventional memory systems, e.g., DDR2, DDR3, etc.
In some embodiments, controller 300 may receive a request signal from another device or from the CPU 102 to perform a data transfer. Upon receipt of the signal, controller 300 acquires control of the system bus 106 and performs a data transfer. The controller 300 manages the data transfer which may occur in only a few bus read/write cycles. Since the DMA controller 300 is managing-the transfer, the CPU 102 is available to perform other functions during the data transfer. In other embodiments, the controller 300 may be accessed by the CPU 102, which programs the data registers 304 and address registers 306 of controller 300 to perform a data transfer.
The data stored in memory 200 may be transferred in one of several ways during a DMA data transfer. For example, data stored in memory 200 may be transferred in a single bus operation where the data is simultaneously read from the source and written to the destination. This transfer is typically performed by the controller taking control of the system bus 106 from CPU 102 and signaling for the data to be latched onto or off of the system bus 106.
Another way in which data may be transferred is a fetch-and-deposit transfer where the controller 300 fetches or reads the data from one memory address and deposits or writes the data to another address. The fetch-and-deposit manner of transferring data requires two memory cycles, a first cycle to read the data and a second cycle to write the data.
Using DMA to access the data stored in the vertically stacked chips 202 of memory 200 enables the bandwidth of the memory to be increased as the full width of the bus may be used to transfer data. For example, if a bus has a width of 64-bits, then 64-bits of data may be transferred using DMA, which is eight times the amount of data that may be transferred using DDR3, which has an 8-bit prefetch buffer. Additionally, data transferred using DMA does not require the CPU to allocate resources and is transferred using a DMA clock (not shown) independent of the system clock, so the data are transferred faster than data transferred in accordance with conventional memory specification, e.g., DDR2 and DDR3. By connecting the memory chips 202 with through silicon vias, the number of I/O ports of a memory device 200 may be increased which enables a wider bus to be used which also enables the bandwidth of the memory device to be increased.
Although the invention has been described in terms of exemplary embodiments, it is not limited thereto. Rather, the appended claims should be construed broadly, to include other variants and embodiments of the invention, which may be made by those skilled in the art without departing from the scope and range of equivalents of the invention.