US 5,829,026 AGrant
Method and Structure for Implementing a Cache Memory Using a DRAM Array
Issue Date:1998-10-27
•47 Claims
•11 Drawing Sheets
Abstract
A method and structure for implementing a DRAM memory array as a second level cache memory in a computer system. The computer system includes a central processing unit (CPU), a first level SRAM cache memory, a CPU bus coupled to the CPU, and a second level cache memory which includes a DRAM array coupled to the CPU bus. When accessing the DRAM array, row access and column decoding operations are performed in a self-timed asynchronous manner. Predetermined sequences of column select operations are then performed in a synchronous manner with respect to a clock signal. The DRAM array is operated at a higher frequency than the frequency of the CPU bus clock signal, thereby reducing the access latency of the DRAM array. By operating the DRAM array at a higher frequency than the CPU bus, additional time is provided for precharging the DRAM array. As a result, the precharging of the DRAM array is transparent to the CPU bus.
Metadata
Assignee
- Monolithic System Technology, Inc.
Inventors
- Wingyu Leung
- Fu-Chieh Hsu
Application Information
Application Number:US 8120005
Filing Date:1997-03-05
Priority Date:1994-11-22
Art Unit:271
Classifications
IPC:
G06F 1208
Field of Search:
395711445;432105;118;119;122
Patent Drawings (11 sheets)
Description
Background of the Invention
1. Field of the Invention
The present invention relates to a method and structure for implementing a memory system. More specifically, the invention relates to a second level cache memory.
2. Description of the Prior Art
High-speed computer systems frequently use fast, small-capacity cache (buffer) memory to transmit signals between a fast processor and a slow (and low cost), large-capacity main memory. Cache memory is typically used to temporarily store data which has a high probability of being selected next by the processor. By storing this high probability data in a fast cache memory, the average speed of data access for the computer system is increased. Thus, cache memory is a cost effective way to boost system performance (as compared to using all high speed, expensive memories). In more advanced computer systems, there are multiple levels (usually two levels) of cache memory. The first level cache memory, typically having a storage of 4 Kbytes to 32 Kbytes, is ultra-fast and is usually integrated on the same chip with the processor. The first level cache is faster because it is integrated with the processor and therefore avoids any delay associated with transmitting signals to and receiving signals from an external chip. The second level cache is usually located on a different chip than the processor, and has a larger capacity, usually from 64 Kbytes to 1024 Kbytes.
FIG. 1 is a block diagram of a prior art computer system 100 using an SRAM second level cache configuration. The CPU or microprocessor 101 incorporates on-chip SRAM first level cache 102 to support the very fast internal CPU operations (typically from 33 Mhz to 150 Mhz).
First level cache 102 typically has a capacity of 4 Kbytes to 32 Kbytes and performs very high speed data and instruction accesses (typically with 5 to 15 ns). For first-level cache miss or other non-cacheable memory accesses, the memory read and write operations must go off-chip through the much slower external CPU bus 104 (typically from 25 Mhz to 60 Mhz) to the SRAM second level (L2) cache 106 (typically with 128 Kbytes to 1024 Kbytes capacity) with the additional latency (access time) penalty of round-trip off-chip delay.
The need for CPU 101 to manage the delay penalty of off-chip operation dictates that in almost all modern microprocessors, the fastest access cycle (read or write) through the CPU bus 104 is 2-1-1-1. That is, the first external access will consume at least 2 clock cycles, and each subsequent external access will consume a single clock cycle. At higher CPU bus frequencies, the fastest first external access may take 3 or more clock cycles. A burst cycle having 4 accesses is mentioned here for purposes of illustration only. Some processors allow shorter (e.g., 2) or longer (e.g., 8 or more) burst cycles. Pipelined operation, where the parameters of the first external access of the second burst cycle are latched into CPU bus devices while the first burst cycle is still in progress, may hide the longer access latency for the first external access of the second burst cycle. Thus, the first and second access cycles may be 2-1-1-1, 1-1-1-1, respectively.
The cache tag memory 108 is usually relative small (from 8 Kbytes to 32 Kbytes) and fast (typically from 10 to 15 ns) and is implemented using SRAM cells. Cache memory 108 stores the addresses of the cache lines of second level cache 106 and compares these addresses with an access address on CPU bus 104 to determine if a cache hit has occurred. This small cache tag memory 108 can be integrated with the system logic controller chip 110 for better speed and lower cost. An integrated cache tag memory operates in the same manner as an external cache tag memory. Intel's 82430 PCI set for the Pentium processor is one example of a logic controller chip 110 which utilizes an SRAM integrated cache tag memory.
One reason for the slower operating frequency of CPU bus 104 is the significant loading caused by the devices attached to CPU bus 104. Second level (L2) SRAM cache memory 106 provides loading on the data and address buses (through latch 112) of CPU bus 104. Cache tag memory 108 provides loading on the address bus, system logic controller chip 110 provides loading on the control, data and address buses, and main memory DRAM 114 provides loading on the data bus (through latch 116).
In prior art computer system 100, the system logic chip 110 provides an interface to a system (local) bus 118 having a typical operating frequency of 25 Mhz to 33 Mhz. System bus 118 may be attached to a variety of relatively fast devices 120 (such as graphics, video, communication, or fast disk drive subsystems). System bus 118 can also be connected to a bridge or buffer device 122 for connecting to a general purpose (slower) extension bus 124 (at 4 Mhz to 16 Mhz operating frequency) that may have many peripheral devices (not shown) attached to it.
Traditional high speed cache systems, whether first level or second level, are implemented using static random access memories (SRAMs) because the SRAMs are fast (with access times ranging from 7 to 25 nanoseconds (ns) and cycle times equal to access times). SRAMs are suitable for storing and retrieving data from high-speed microprocessors having bus speeds of 25 to 100 megahertz. Traditional dynamic random access memories (DRAMs), are less expensive than SRAMs on a per bit basis because DRAM has a much smaller cell size. For example, a DRAM cell is typically one quarter of the size of an SRAM cell using comparable lithography rules. DRAMs are generally not considered to be suitable for high speed operation because DRAM accesses inherently require a two-step process having access times ranging from 50 to 120 ns and cycle times ranging from 90 to 200 ns.
Access speed is a relative measurement. That is, while DRAMs are slower than SRAMs, they are much faster than other earlier-era memory devices such as ferrite core and charge-coupled devices (CCD). As a result, DRAM could theoretically be used as a "cache" memory in systems which use these slower memory devices as a "main memory." The operation modes and access methods, however, are different from the operation modes and access methods disclosed herein.
In most computer systems, the second level cache operates in a fixed and rigid mode. That is, any access read or write to the second level cache is of a few constant sizes (line sizes of the first and second level caches) and is usually in a burst sequence of 4 or 8 words (i.e., consecutive reads or writes of 4 or 8 words) or in a single access (i.e., one word). These types of accesses allow standard SRAMs to be modified to allow these SRAMs to meet the timing requirements of very high speed processor buses. One such example is the burst or synchronous SRAM, which incorporates an internal counter and a memory clock to increment an initial access address. External addresses are not required after the first access, thereby allowing the SRAM to operate faster after the first access is performed. The synchronous SRAM may also have special logic to provide preset address sequences, such as Intel's interleaved address sequence. Such performance enhancement, however, does not reduce the cost of using SRAM cells to store memory bits.
Synchronous DRAMs (SDRAM) have adopted similar burst-mode operation. Video RAMs (VRAM) have adopted the serial port operation of dual-port DRAMs. These new DRAMs are still not suitable for second level cache operation, however, because their initial access time and random access cycle time remain much slower than necessary.
It would therefore be desirable to have a structure and method which enables DRAM memory to be used as a second level cache memory.
Prior art computer systems have also included multiple levels of SRAM cache memory integrated on the same chip as the CPU. For example, DEC's Alpha 21164 processor integrates 16 Kbytes of first level SRAM cache memory and 96 Kbytes of second level SRAM memory on the same chip. In such cases, a third level SRAM cache is typically used between the processor and a DRAM main memory. In such a computer system, it would be desirable to use a DRAM memory to replace the third level SRAM cache memory.
Summary of the Invention
In accordance with the present invention, a structure and method for configuring a DRAM array, or a plurality of DRAM arrays, as a second level cache memory is provided. A structure in accordance with the invention includes a computer system having a central processing unit (CPU), a SRAM cache memory integrated with the CPU, a CPU bus coupled to the CPU, and a second level cache memory comprising a DRAM array coupled to the CPU bus. The second level cache memory is configured and integrated with system logic on a monolithic integrated circuit (IC) in one embodiment. For high pin count microprocessors such as Intel's Pentium, the companion system logic controller may be partitioned into multiple chips (e.g., Intel's 82430 PCI set). In such a system, the second level cache DRAM array of the present invention may be integrated with one of the system logic chips, preferably the system logic chip(s) for the data path. In another configuration, the second level cache memory can be integrated with the CPU itself.
When accessing the DRAM array, row access and column decoding operations are performed in a self-timed asynchronous manner. Predetermined sequences of column select operations are then performed, wherein the column select operations are synchronous with respect to a clock signal. This asynchronous-synchronous accessing scheme reduces the access latency of the DRAM array.
In one embodiment, the DRAM array is operated in a dual-edge transfer mode in response to the CPU bus clock signal. Consequently, the DRAM array performs access operations at a frequency which is twice as fast as the frequency of the CPU bus clock signal. DRAM access therefore occurs twice as fast as operations on the CPU bus.
In another embodiment, the second level cache memory includes a phase locked loop (PLL) circuit coupled to the CPU bus. The PLL circuit generates a fast clock signal having a frequency greater than the frequency of a CPU bus clock signal. The fast clock signal is provided to the DRAM array to control read and write operations. In one embodiment, the fast clock signal has a frequency equal to twice the frequency of the CPU bus clock signal. Again, DRAM access occurs twice as fast as the operations on the CPU bus.
Data values can be read from the DRAM array to the CPU bus through a read first in first out (FIFO) memory having a data input port coupled to the DRAM array and a data output port coupled to the CPU bus. The data input port is clocked by the fast clock signal and the data output port is clocked by the CPU bus clock signal. Because data is read out of the DRAM array faster than the data is read out to the CPU bus, additional time is available during which the DRAM array can be precharged. The precharge time is thereby "hidden" from the CPU bus during a read operation from the second level cache memory.
Data values can also be written from the CPU bus to the DRAM array through a write FIFO memory having a data output port coupled to the DRAM array and a data input port coupled to the CPU bus. The output port of the write FIFO memory is clocked by the fast clock signal and the input port of the write FIFO memory is clocked by the CPU bus clock signal. A first set of data values is written and stored in the write FIFO memory until a second set of data values is written to the write FIFO memory. At this time, the first set of data values is written to the DRAM array at the frequency of the fast clock signal. Because the first set of data values is written to the DRAM array faster than the second set of data values is written to the write FIFO memory, a DRAM precharge operation can be performed during the time the second set of data values are written to the write FIFO memory. Therefore, the DRAM precharge operation is effectively "hidden" from the CPU bus during a write operation to the second level cache memory.
In an alternate embodiment, data values to and from the DRAM array are routed through a sense amplifier circuit, a data amplifier circuit and a column selector coupled between the sense amplifier circuit and the data amplifier circuit. Writing data values to the DRAM array then involves the steps of (1) opening the column selector to isolate the data amplifier circuit from the sense amplifier circuit, (2) writing the data values from the write FIFO memory to the data amplifier circuit substantially in parallel with performing a row access operation in the DRAM array, and (3) closing the column selector to connect the data amplifier circuit to the sense amplifier circuit, thereby causing the data values to be provided to the DRAM array through the sense amplifier circuit. By writing data values to the write FIFO memory in parallel with the row access operation, more time is available to precharge the DRAM array.
The column selector can also be used during a DRAM read operation to provide additional time for a DRAM precharge operation. To do this, data values are read from the DRAM array to the sense amplifier circuit. The column selector is then closed to connect the sense amplifier circuit to the data amplifier circuit. After the data values have been written to the data amplifier circuit, the column selector is opened, thereby isolating the sense amplifier circuit from the data amplifier circuit. The data values can then be read out of the data amplifiers while the DRAM array is being precharged.
By operating the DRAM array with a faster clock signal than the CPU bus clock, a DRAM memory array can be used to satisfy the speed and operational requirements of a second level cache memory. Such a DRAM memory array can be used at a lower cost, typically 25%, than traditional SRAM implementations.
The present invention will be more fully understood in light of the following detailed description taken together with the drawings.
Brief Description of the Drawings
FIG. 1 is a block diagram of a prior art computer system having an SRAM second level cache memory;
FIG. 2 is a block diagram of a computer system having a DRAM second level cache memory in accordance with the invention;
FIGS. 3(a) and 3(b) illustrate a schematic diagram and a timing diagram, respectively, of a self-timed RAS/CAS/burst accessing sequencer;
FIG. 4 is a schematic diagram of a fast column accessing circuit;
FIG. 5 is a schematic diagram of circuitry which provides for operation of a DRAM second level cache memory at twice the frequency of the CPU bus clock;
FIG. 6 is a timing diagram of a 2-1-1-1 DRAM second level cache read operation;
FIG. 7 is a timing diagram of a 2-1-1-1 DRAM second level cache write operation;
FIG. 8 is a timing diagram of a 2-1-1-1 DRAM second level cache read operation in accordance with an alternate embodiment of the invention;
FIG. 9 is a timing diagram of a 2-1-1-1 DRAM second level cache write operation in accordance with an alternate embodiment of the invention; and
FIG. 10 is a schematic diagram of a refresh management controller.
Detailed Description of the Invention
FIG. 2 is a block diagram of computer system 200 which includes second level DRAM cache memory 213 in accordance with the invention. In addition to second level DRAM cache memory 213, computer system 200 includes CPU 201, first level SRAM cache 202, CPU bus 204, latches 212 and 216, second level SRAM cache tag memory 208, system logic controller 211, main DRAM memory 214, system bus 218, bridge buffer circuit 222, system bus devices 220 and extension bus 224.
Although DRAM cache memory 213 is referred to as a "second level" cache memory, it is understood that the present invention can also be applicable to other "levels" of cache memory higher than the second level (e.g., third level or fourth level). In general, the present invention is applicable to the "next level" cache memory, where the "next level" cache memory is defined as the cache memory which is connected between the processor and a large-capacity main memory (where the main memory is typically DRAM). Using this definition, the term "second level cache memory" is interchangeable with the term "next level cache memory" in the following.
CPU 201, which is typically fabricated on the same integrated circuit chip as first level SRAM cache memory 202, is coupled to the control, address and data lines of the CPU bus 204. Second level SRAM cache tag memory 208 receives address signals from the address lines of CPU bus 204 through latch 212. System logic controller 211 (which controls second level DRAM cache memory 213 and main memory 214) is coupled to the control, address and data lines of CPU bus 204. In one embodiment, main DRAM memory 214 receives data signals from the data lines of CPU bus 204 through latch 216. In another embodiment, main DRAM memory 214 receives data signals from system logic controller 211 through alternate data path 226. System logic controller 211 interacts with cache tag memory 208 and main memory 214 in a conventional manner. In one embodiment, SRAM cache tag memory 208, latches 212, system logic controller 211 and second level cache memory 213 are fabricated on the same integrated circuit chip 210. In another embodiment, second level DRAM cache memory 213 and system logic controller 211 are fabricated on separate chips. In yet another embodiment, second level DRAM cache memory 213 and system logic controller 211 are fabricated on the same chip as CPU 201 and first level SRAM cache memory 202.
Because the data paths from CPU bus 204, SRAM cache tag memory 208, system bus 218 and main memory 214 feed into system logic chip 211, the system logic chip 211 can manage most of the data traffic locally without tying up CPU bus 204. For example, system logic chip 211 controls data traffic between main memory 214 and second level DRAM cache memory 213 or between system bus 218 and main memory 214. Consequently, the loading of CPU bus 204 is lighter and the physical layout of CPU bus 204 is more compact, thereby allowing for faster operations on CPU bus 204.
Embodiments of the present invention overcome many obstacles to successfully use second level DRAM cache memory 213 as a second level cache memory which matches the performance of an SRAM-cell based second level cache. These obstacles include: (1) slow access latency, (2) precharge time, (3) refresh arbitration and control. These obstacles are overcome as described below.
(1) Slow Access Latency
The access time of a DRAM cell array consists of a row access (RAS) time (i.e., the time to decode a row address, select a row word-line, and load a row of data bits from the DRAM cell array into sense amplifiers in the column area) and a column access (CAS) latency (i.e., the time to decode a column address, select a column, read the data from the sense amplifiers into data amplifiers, and then propagate the data signal to the chip input/output area). The sum of the row and column access latencies is relatively long (45 to 60 ns) compared to SRAM access latency (7 to 25 ns). The DRAM access is longer because of the two-step sequential access, as well as the relatively long column decoding and access time (17 to 25 ns).
Some prior art DRAM devices such as pseudo-SRAMs and BiCMOS fast DRAM (e.g., Hitachi's 35 ns 1 Mb BiCMOS DRAM with demultiplexed addressing), use an SRAM-like interface in which the full row and column addresses are provided to the chip at one time and internal two-step access is performed in a fully asynchronous fashion. However, these pseudo-SRAM devices combine the RAS and CAS data operations, precharge time and the array refresh operation into each access cycle to emulate standard asynchronous SRAM operations. As a result, these pseudo-SRAM devices are very slow and not suitable for cache memory applications.
The BiCMOS fast DRAM has a fast initial access time because demultiplexed addresses allow row and column addresses to be loaded at the beginning of each access cycle using separate address pins. However, the BiCMOS fast DRAM still operates with the constraints of traditional asynchronous DRAM (i.e., long access cycle times and relatively slow subsequent accesses).
In one embodiment of the invention, accesses to second level DRAM cache memory 213 are made faster by including a self-timed RAS/CAS/burst sequencer within second level DRAM cache memory 213. The burst sequencer merges asynchronous and synchronous operations of the DRAM accesses in a seamless fashion as described below. FIG. 3a is a block diagram of a self-timed RAS/CAS burst sequencer 300 in accordance with the invention. Burst sequencer 300 represents a portion of the accessing circuit included within second level DRAM cache memory 213. Burst sequencer 300 includes control circuit 301, row address register 302, column address register 303, row decoder 304, row selector 305, sense amplifier control circuit 306, delay circuits 307-308, sense amplifier circuit 306, column decoder 310, column selector 311, data amplifier circuit 321 and burst sequence controller 313. Burst sequencer 300 is used to access an array of DRAM memory cells 317.
FIG. 3b is a waveform diagram illustrating the operation of burst sequencer 300. To access second level DRAM cache memory 213, CPU 201 transmits a control signal through CPU bus 204 to second level DRAM cache 213. Second level DRAM cache 213 converts this control signal to an address strobe signal (See, e.g., timing control circuit 502, FIG. 5) which is provided to control circuit 301. CPU 201 also transmits an address through CPU bus 204 to second level DRAM cache memory 213. Second level DRAM cache 213 converts this control signal to row and column addresses (See, e.g., address buffer 503, FIG. 5) which are provided to row address register 302 and column address register 303, respectively. In response to the address strobe signal, control circuit 301 generates a signal which causes the full row and column addresses (and bank address, not shown here, to simplify the schematic diagram) to be latched into registers 302 and 303, respectively. It is not necessary to simultaneously latch in the row and column addresses because the column address is not needed until the RAS operation is completed. As long as the column address is latched into register 303 before the completion of the RAS operation, there is no speed penalty.
The row address stored in row address register 302 and a row decode signal (FIG. 3b) generated by control circuit 301 are transmitted to row decoder 304. In response, row decoder 304 decodes the row address and transmits this decoded address to row selector 305. Row selector 305, turns on the appropriate word line (FIG. 3b) (i.e., performs a row selection operation) within DRAM array 317.
A sense amplifier enable signal (FIG. 3b) is then generated by delay circuit 307 and transmitted to sense amplifier circuit 306. In response, the sense amplifiers in sense amplifier circuit 306 turn on to receive the data values of the selected row within DRAM array 317. The asynchronous delay introduced by delay circuit 307 is selected in view of the delays inherent in row decoder 304, row selector 305 and DRAM array 317, such that sense amplifier circuit 306 is enabled as soon as the data values from the selected row of DRAM array 317 are available (i.e., as soon as the row access operation is completed). Delay circuit 307 can be realized in a number of different ways, such as an inverter chain or an RC circuit.
At the same time the row access operation is being performed, the column address can be provided from column address register 303 to column decoder 310, and column decoder 310 can perform the column decode operation. At the same time, burst sequence controller 313 can be set up to supply a special address scrambling sequence based on the initial column address received from column address register 303. For example, Intel's 486 and Pentium microprocessors modula-4 sequences are 0-1-2-3, 1-0-3-2, 2-3-0-1 & 3-2-1-0.
After sense amplifier circuit 306 is turned on and the column decode operation has been performed, the column decoder 310 receives a column decode enable signal from delay circuit 308. This column decode enable signal causes column decoder 310 to provide the decoded column address to the column selector 311. After the initial decoded column address is provided to column selector 311, burst sequence controller 313 causes column decoder 310 to sequentially change the decoded column address provided to the column selector 311 once during each half clock cycle. As a result, sequential data (with appropriate burst sequence scrambling determined by burst sequence controller 313) are read into data amplifiers 312 synchronous to the clock signal. In one embodiment, the clock signal is the CPU bus clock signal (i.e., a buffered copy of the clock signal provided by CPU bus 204).
The clock signal is also provided to data amplifier circuit 312. Data is read from data amplifier circuit 312 to data line 330 at both the rising and falling edges of the clock signal (i.e., dual edge transfer). In this specification, when an operation is said to occur "at a clock edge", it is understood that the operation occurs immediately after the occurrence of the clock edge.
A burst mode write operation is very similar to the read operation, except that data is coming from the chip input/output circuitry and is synchronously loaded into data amplifiers 312 and through column selector 311 into the appropriate sense amplifiers 306 with appropriate burst address sequence. The asynchronous self-timed RAS and CAS operations allow very tight access timing independent of clock frequency (that is, the RAS and CAS access time is constant and not a function of the clock frequency) while at the same time employing fully synchronous operation for the burst read/write operation that scales with clock frequency.
FIG. 4 is a schematic diagram of a portion of column selector 311 and data amplifier circuit 312 of FIG. 3a. The following discussion describes the manner in which the circuitry of FIG. 4 provides for dual-edge data transfer.
Column selector 311 includes tree decoders 311a and 311b. Tree decoders 311a and 311b are coupled to sense amplifier circuit 306 through a predetermined number (e.g., 32) of complementary signal lines. Tree decoders 311a and 311b are also coupled to column decoder 310. In the embodiment illustrated, column decoder 310 provides control signals Sa7:0! and Sb3:0!, which cause tree decoders 311a and 311b to selectively couple one of the sense amplifiers in sense amplifier circuit 306 to data amplifier circuit 312.
Data amplifier circuit 312 includes data amplifiers 312a and 312b, multiplexer 907, read data latch 914, write buffers 903 and 913, tri-state buffer 905 and clock generation circuit 918. The circuitry illustrated in FIG. 4 services 64 of the sense amplifiers in sense amplifier circuit 306. The circuitry of FIG. 4 is repeated for each additional 64 sense amplifiers in sense amplifier circuit 306. In one embodiment, the total number of data amplifiers in data amplifier circuit 312 is equal to the number of bits in each data word read from or written to DRAM array 317.
The complementary outputs of tree decoders 311a and 311b are provided to data amplifiers 312a and 312b, respectively. Data amplifiers 312a and 312b are regenerative latches which include cross-coupled transistors 970-973 and transistors 974-977. These regenerative latches are controlled by a locally generated, single phase clock signal D.sub.sense.
A local self-timed clock circuit 918 generates the control signals used to control data amplifiers 312a and 312b and multiplexer 907. A column precharge signal, PC, and the D.sub.sense signal are generated in response to the clock signal, a column-access (CAS) signal and a bus precharge signal, WE (for write operation). The CAS and WE signals are generated by control circuit 301. The clock signal is the same clock signal illustrated and described in connection with FIGS. 3a and 3b. The PC and D.sub.sense signals are local signals which are not used to drive any circuitry outside data amplifier pair 312a and 312b. Thus, timing skew in the control signals is minimized.
Read operation
To perform a read operation, the WE signal is deasserted high. As a result, transistors 950-953 of write buffers 903 and 913 are turned off and tri-state buffer 905 is placed in a low impedance state. The CAS signal is asserted high. During a first half cycle of the clock signal, the clock signal is in a logic high state, thereby forcing both the D.sub.sense and PC signals to a logic high state. Under these conditions, the complementary outputs of tree decoders 311a and 311b are latched in data amplifiers 312a and 312b, respectively.
For example, a logic low signal on lead 925 and a logic high signal on lead 926 cause transistors 971 and 972 to turn on and transistors 970 and 973 to turn off. The high D.sub.sense signal causes transistor 961 to turn on. As a result, node 991 is pulled down to ground through transistors 972 and 961 and node 992 is pulled up to Vdd through transistor 971. In a similar manner, a logic low signal on lead 926 and logic high signal on lead 925 results in node 992 being pulled to ground through transistors 973 and 961 and node 991 being pulled to Vdd through transistor 970.
Data amplifier 312b operates in the same manner as data amplifier 312a to latch the signals present on leads 927 and 928. Thus, a logic high signal on lead 927 and logic low signal on lead 928 results in node 993 being pulled up to Vdd through transistor 974 and node 994 being pulled down to ground through transistors 977 and 962. Similarly, a logic low signal on lead 927 and logic high signal on lead 928 results in node 993 being pulled to ground through transistor 976 and 962 and node 994 being pulled to Vdd through transistor 975.
Within multiplexer 907, the high D.sub.sense signal causes transmission gates 995 and 997 to close (i.e., be placed in a conducting state) and transmission gate 996 to open (i.e., be placed in a non-conducting state). As a result, the voltage on node 992 is transmitted through transmission gate 995 and tri-state buffer 905 to data line 330. Data line 330 connects tri-state buffer 905 directly to the bus transceivers in the input/output circuit. This connection results in little loading other than the routing capacitance because there is no other signal multiplexed on this line. Loading of data line 330 is thus substantially smaller than that present in prior art schemes. Consequently, the data lines of the present invention are capable of operating at much higher frequency (up to 250 Mhz).
In addition, the voltage on node 933 is transmitted through transmission gate 997 and is stored in read data latch 914.
During the second half cycle of the clock signal, the clock signal transitions low, thereby forcing both the D.sub.sense and PC signals low. In response to the low PC signal, transistors 920-923 are turned on. As a result, leads 925-928 are coupled to Vdd (i.e., leads 925-928 are precharged). In addition, the low D.sub.sense signal opens transmission gates 995 and 997 and closes transmission gate 996. As a result, the voltage stored in read data latch 914 is read out through transmission gate 996 and tri-state buffer 905 to data line 330 during the second half cycle. In the foregoing manner, dual-edge transfer of data from sense amplifier circuit 306 to data line 330 is facilitated.
Write operation
To perform a write operation, the WE signal is asserted low, thereby placing tri-state buffer 905 in a high-impedance state and applying a logic low signal to an input of each of NOR gates 954-957 in write buffers 903 and 913. During a first half cycle of the clock signal, the clock signal is in a logic low state, thereby closing transmission gate 906 and opening transmission gate 916. The signal on the data line 330 is therefore routed to an input of NOR gate 955. For example, a high signal on the data line 330 causes NOR gate 955 to provide a logic low signal to transistor 951, thereby turning off this transistor. The low output of NOR gate 954 is also provided to an input of NOR gate 954, causing NOR gate 954 to output a logic high signal which turns on transistor 950.
The low WE signal also causes the D.sub.sense and PC signals to go high, thereby turning off p-channel transistors 961-962. As a result, p-channel transistor 971 and n-channel transistor 972 are turned on. Consequently, tree decoder 311a receives supply voltage Vdd on lead 926 and the ground supply voltage on lead 925, thereby writing a high data value to the selected column of sense amplifier.
If the input from data line 330 is a logic low signal (as opposed to a logic high signal as previously discussed), tree decoder 311a receives ground supply voltage on lead 926 and supply voltage Vdd on lead 925 in a manner similar to that previously described. In this manner, data is written from data line 330 to the sense amplifiers during each half cycle of the clock signal. The demultiplexing performed by transmission gates 906 and 916 is necessary because the address selected by tree decoders 311a and 311b changes only once every clock cycle.
Tree decoders 311a and 311b limit the multiplexing loading to approximately 12 lines (8+4) (as opposed to 512 lines in a typical conventional scheme). The decreased capacitive loading together with the higher drive signal provided by data amplifier circuit increase the data bandwidth.
Delay matching
Delay matching in the column circuitry is minimized by routing the lines carrying the clock signal, the pre-decoded column select signals Sa7:0! and Sb3:0!, and the data signals in the same manner through the column area of the memory array.
Alternate Embodiment
FIG. 5 illustrates an alternate embodiment, in which the frequency of the clock signal is doubled and the clock generation circuit 918 (FIG. 4) is modified such that data values are read from (or written to) data amplifier circuit 312 at each rising (or falling) edge of the clock signal. Thus, data values are transferred on data line 330 at the same rate as the previously described embodiment (i.e., at a frequency equal to twice the CPU bus clock frequency). The clock generation circuit 918 can be modified by transmitting the doubled clock signal through a 1-bit counter (not shown) before the clock signal is applied to clock generation circuit 918. Other modifications to clock generation circuit 918 to allow the circuitry of FIG. 4 to operate in a single edge transfer mode with respect to the doubled clock signal would be apparent to one of ordinary skill.
Phase-locked loop (PLL) circuit 501 buffers the CPU bus clock signal (from CPU bus 204, FIG. 2) and generates a clock signal having the same frequency as the CPU bus clock signal (hereinafter referred to as a 1X clock signal) and a clock signal having twice the frequency of the CPU bus clock signal (hereinafter referred to as a 2X clock signal). The 1X and 2X clock signals have fixed phase relationships with respect to the incoming CPU bus clock signal. These phase relationships are selected to provide data set-up and hold times appropriate for proper data transfer.
Address buffer 503 latches the CPU bus address and decodes it into row and column addresses (and a bank address if there are multiple DRAM arrays, not shown here for purposes of simplicity). Timing control circuit 502 derives an internal address strobe signal from the CPU bus address (received from address buffer 503) and a control signal received from CPU bus 204. The address strobe, row address, column address and 2X clock signals are provided to burst sequencer 300 and DRAM array 317. Burst sequencer 300 and DRAM array 317 operate substantially as previously described in response to these signals (See, FIG. 3a).
Address buffer 503 may have additional latches (pre-fetch buffers) to store the address of the next access operation while the current access is still in progress. The pre-fetch buffers enable pipelined operation so that back to back operations may be partially overlapped to reduce the latency cycles between operations.
The remainder of the circuitry illustrated in FIG. 5 is directed toward solving the problems introduced by the precharge time required for the DRAM array. Accordingly, this circuitry is discussed below.
(2) Precharge Time
The operation of a DRAM cell array requires that after a normal read or write access (RAS+CAS access), the selected row be de-selected and the sense-amplifiers be turned off and equalized before any subsequent RAS operations are initiated. This operation is referred to as a precharge operation. The time period required to perform the precharge operation is referred to as the precharge (PRE) time. The PRE time is sufficiently long to fully equalize the sense amplifiers and the relatively high capacitance bitlines, so that the very small signal provided by the cell capacitor to the sense amplifier in connection with the next RAS operation can be read correctly and reliably. The PRE time requirement prevents DRAMs from executing back to back accesses which SRAMs can easily support. Thus, the access cycle time of DRAM is much longer (typically 1.5X to 2X) than its access latency, while SRAM's access cycle time is approximately equal to its access latency.
To be able to use DRAM with SRAM performance, the PRE time must be substantially "hidden" from the access operations of the CPU bus. Read data FIFO 504, write data buffer 505 and write data FIFO 506, illustrated in FIG. 5, operate to allow the PRE time to be hidden from the access operations.
As described in more detail below, the 2X clock signal is used to clock DRAM array 317, the data input terminal of read data FIFO 504, and the data output terminal of write data FIFO 506. The 1X clock signal is used to clock the data output terminal of read data FIFO 504 and the data input terminal of write data buffer 505.
Data values are read from DRAM array 317 to CPU bus 204 through read data FIFO 504. Data values are read into read data FIFO 504 at the frequency of the 2X clock signal. Data is then read out of read data FIFO 504 at the frequency of the CPU bus clock signal. In this manner, read data FIFO 504 performs clock resynchronization.
Conversely, data values are written to DRAM array 317 from CPU bus 204 through write data buffer 505 and write data FIFO 506. Data values are read into write data buffer 505 at the frequency of the CPU clock signal and read out of write data FIFO 506 at the frequency of the 2X clock signal.
To minimize clock-to-data skew, DRAM array 317 can alternatively provide a 2X clock signal to read data FIFO 504 along with the data in a source-synchronous fashion. The alternative 2X clock signal is a return clock signal which travels along a path which is selected such that the 2X clock signal exiting DRAM array 317 has a preselected delay and phase relationship with respect to the data values exiting DRAM array 317.
FIG. 6 is a timing diagram for a 2-1-1-1 data read burst operation performed by the circuitry of FIG. 5. After the address strobe signal is asserted low, the RAS and CAS operations are initiated in a self-timed, asynchronous fashion (See, FIGS. 3a and 3b). Two rising clock edges after the address strobe signal is asserted, the RAS and CAS operations are completed and a burst read operation is performed by DRAM array 317 in a fully synchronous fashion with respect to the 2X clock signal. The read burst data from DRAM array 317 is clocked into read FIFO 504 by the 2X clock signal. The read burst data is clocked out of read FIFO 504 to the CPU bus by the 1X clock signal. As soon as the read data burst is completed, a precharge operation is initiated to DRAM array 317, thereby preparing DRAM array 317 for the next operation. This next operation can be either a normal back-to-back access or a pipelined access. Because the read data burst is written to read FIFO 504 by the 2X clock signal, there is time left to perform the precharge operation before the data is read out of read FIFO 504 by the 1X clock signal. Thus, the precharge time is hidden from CPU bus 204. If the precharge time is short enough, DRAM array 317 may be ready for a subsequent operation at a time which would allow for pipelined operation.
FIG. 7 is a waveform diagram illustrating the timing diagram for a 2-1-1-1 data write burst operation. Because the data lines of the CPU bus 204 receive write burst data at a rate equal to the frequency of the CPU bus clock signal (i.e., the frequency of the 1X clock signal,) a full write burst is not completed until the end of the 5th clock cycle. Thus, no time remains in a 2-1-1-1 data write burst operation to perform a precharge operation.
Therefore, a separate write data buffer 505 (FIG. 5) is used to latch in a first group of write burst data values (e.g., D1-D4) from CPU bus 204. When a second group of write burst data values (e.g., D1'-D4') arrives from CPU bus 204 (there may be multiple intervening read bursts), the first group of write burst data values D1-D4 is forwarded to DRAM array 317 through write data FIFO 506. The second group of write burst data values D1'-D4' is then stored in write data buffer 505. The address strobe signal initiates the RAS and CAS operations in a self-timed, asynchronous fashion (See, e.g., FIGS. 3a and 3b). The first group of write burst data values D1-D4 is clocked from write data FIFO 506 by the 2X clock signal in a fully synchronous fashion. A precharge operation is initiated after the data values D1-D4 are written to DRAM array 317. Because the write data burst is written to DRAM array 317 by the 2X clock signal, there is time left to perform the precharge operation before data values D1'-D4' are written to write data buffer 505 by the 1X clock signal. Thus, the precharge time is hidden from CPU bus 204. Again, if the precharge time is short enough, DRAM array 317 may be ready for a subsequent operation at a time which would allow for pipelined operation.
FIG. 5 also illustrates a data bypass path 510 from write data FIFO 506 to read data FIFO 504. Data bypass path 510 allows for the special case where a CPU bus 204 requires access to a group of write burst data stored in write data buffer 505 or write data FIFO 506, but not yet sent to DRAM array 317. In this case, the write burst data is transmitted from write data FIFO 506 to read data FIFO 504 at the same time that the write burst data is sent from write data FIFO 506 to DRAM array 317.
In alternate embodiments, additional write data buffers can be connected between write data buffer 505 and CPU bus 204 to add depth to the multi-depth write data FIFO created by write data buffer 505 and write data FIFO 506.
Older microprocessors (i.e., CPU's) may not support the write data burst access to second level DRAM cache memory 213. In these microprocessors, single write accesses are individually transmitted. However, write data buffer 505 and write data FIFO 506 can still operate as previously described to "hide" the precharge time of DRAM array 317.
In another embodiment, read data FIFO 504, write data buffer 505 and write data FIFO 506 can be used in conjunction with the circuitry illustrated in FIGS. 3a and 4. As previously discussed, this circuitry causes data to be transferred on data line 330 at both the rising and falling edges of the 1X clock signal (i.e., dual-edge transfer). When performing dual-edge transfer with the 1X clock signal, PLL 501 is not necessary because the two edges of the incoming CPU bus clock signal provide the necessary timing references used for data transfer. In such an embodiment, the input port of read data FIFO 504 and the output port of write data FIFO 506 are modified such that they are clocked by both the rising and falling edges of the 1X clock signal.
In yet other embodiments, the burst methods previously described can be performed at other clock frequencies (e.g., 4X clock frequency), depending on the timing requirements of DRAM array 317.
The precharge time of DRAM array 317 (FIG. 2) can alternatively be hidden using tree decoders 311a and 311b (FIG. 4) of column selector 311 (FIG. 3a) as isolation switches between sense amplifier circuit 306 and data amplifier circuit 312. To electrically isolate data amplifier circuit 312 from sense amplifier circuit 306, a disconnect control signal is provided to column decoder 310. In response, column decoder 310 disables all control signals Sa7:0! and Sb3:0! such that all of the switches in tree decoders 311a and 311b (FIG. 4) are opened, thereby isolating data amplifiers 312a and 312b from sense amplifiers SA63:0! and SA63:0!. The switches of column selector 311 are opened as soon as the data amplifiers in data amplifier circuit 312 have settled. The precharge operation can be initiated as soon as data amplifier circuit 312 and sense amplifier circuit 306 are isolated.
FIG. 8 illustrates the timing of a read burst operation in accordance with this embodiment of the invention. Isolation occurs after the RAS/CAS access operations are performed. Thus, the burst read and precharge operations can be performed simultaneously. Although FIG. 8 indicates that the burst read operation is performed at the frequency of the 2X clock signal, the burst read operation can also be performed at the frequency of the 1X clock signal because the burst read operation can be performed at the same time as the precharge operation. Performing the burst read operation at the frequency of the 1X clock signal advantageously reduces read errors and power consumption.
Once the data amplifier circuit 312 is disconnected from the sense amplifier circuit 306, data values can only be accessed from data amplifier circuit 312. Consequently, to support a burst access, the number of data amplifiers in data amplifier circuit 312 must be sufficient to store all of the data values required during the burst access. Thus, to support a burst access of 4 words, there must be enough data amplifiers in data amplifier circuit 312 to simultaneously store all of the bits which make up the 4 words. In such an embodiment, multiple sense amplifier data values are read into multiple data amplifiers simultaneously by using multiple data amplifiers. This is in contrast to the previously described embodiments in which data amplifier circuit 312 only needs to have a data amplifier for each of the bits in a single word.
In a variation of the embodiment which uses column selector 310 as an isolation switch, multiple DRAM arrays can be simultaneously accessed to provide a burst access. Thus, four DRAM arrays can be used to provide a burst access of 4 words. To accomplish this, a data word is simultaneously stored in the data amplifier circuit of each of the four DRAM arrays. The data amplifier circuits of these DRAM arrays are then disconnected from their associated sense amplifier circuits. The four words can then be read from the data amplifier circuits of the DRAM arrays in the desired order while the DRAM arrays are simultaneously being precharged.
FIG. 9 illustrates the timing of a write burst operation in accordance with this embodiment of the invention. Write data buffer 505 stores a first group of write burst data values D1-D4. Upon receiving a second group of write burst data values D1'-D4', data amplifier circuit 312 is isolated from the sense amplifier circuit 306 and the first group of write burst data values D1-D4 is transmitted through write data FIFO 506 to data amplifier circuit 312 at the frequency of the 2X clock signal. At the same time, the RAS/CAS access operations are performed. After the RAS/CAS access operations are complete, data amplifier circuit 312 is connected to sense amplifier circuit 306, thereby providing the first group of write burst data values D1-D4 to sense amplifier circuit 306. The precharge operation is then initiated. Because the write data is burst at the frequency of the 2X clock signal, more time is provided to perform the precharge operation. As a result, the precharge operation can be performed before a subsequent write burst operation is to be performed with the second group of write burst data values D1'-D4'.
(3) Refresh Management and Arbitration
Second level DRAM cache memory 213, with passive charge storage, requires refresh operations to periodically (typically every 4 to 64 ms) replenish the charges stored in each cell capacitor. This is because junction, transistor and dielectric leakage currents may cause the stored charge to leak out. FIG. 10 illustrates one embodiment of a refresh management circuit 800 which can be used in connection with the present invention. Refresh management circuit 800 consists of address buffer 801, refresh counters 802, in-progress pointer 803, comparator 804, cache tag comparator 805 and CPU access delay circuit 806. Refresh management circuit 800 is used in connection with an embodiment which uses multiple DRAM arrays (similar to DRAM array 317) within second level DRAM cache 213.
Refresh counters 802 keep track of the addresses of the DRAM arrays and rows to be refreshed next. Refresh counters 802 periodically initiate a RAS/PRECHARGE operation to the DRAM arrays and rows indicated by the refresh counters 802 by transmitting signals selecting these arrays and rows to the appropriate DRAM array(s). In-progress pointer 803 indicates the address of the DRAM array currently being refreshed. Each refresh operation typically lasts 40 to 60 ns.
Any CPU bus access request initiates a comparison to see if there is cache hit in the second level DRAM cache memory 213. At the same time, if there is a refresh operation in progress, the contents of in-progress pointer 803 (the DRAM array address for a single array refresh, or the high order bits of the DRAM array address for a group array refresh) are compared by comparator 804 to the address of the requested DRAM array. Any collision (match of array addresses) will cause CPU access delay circuit 806 to delay CPU access until the refresh operation is completed, in-progress pointer 803 is cleared, and the appropriate ready signal is sent from CPU access delay circuit 806 to the CPU bus. The delay of CPU access delay circuit 806 is set to a predetermined time based on the known timing of the refresh operation. By partitioning second level DRAM cache memory 213 into multiple banks, the probability of a collision during a refresh operation is proportionately reduced.
Although the invention has been described in connection with several embodiments, it is understood that this invention is not limited to the embodiments disclosed, but is capable of various modifications which would be apparent to one of ordinary skill in the art. Thus, the invention is limited only by the following claims.
Claims
What is claimed is:
1. A computer system comprising: a central processing unit (CPU); a first level cache memory integrated with the CPU; a bus coupled to the CPU wherein data is transferred on the bus in response to a first clock signal having a first clock frequency, wherein data is transferred on the bus at a first data rate; and a next level cache memory coupled to the bus, wherein the next level cache memory comprises at least one dynamic random access memory (DRAM) array and a control circuit for controlling data transfer between the bus and the DRAM array, wherein the control circuit derives a second clock signal from the first clock signal, and causes data to be transferred to and from the DRAM array in response to a second clock signal at a second data rate which is faster than the first data rate, wherein the control circuit comprises a read buffer having a data input port coupled to the DRAM array and a data output port coupled to the bus, wherein the data input port is clocked by the second clock signal and the data output port is clocked by the first clock signal, and data is transferred from the DRAM array to the read buffer via the data input port faster than data is transferred from the read buffer to the bus via the data output port.
2. The computer system of claim 1, further comprising an SRAM cache tag memory coupled between the bus and the next level cache memory.
3. The computer system of claim 1, wherein the next level cache memory is integrated with the CPU.
4. The computer system of claim 1, further comprising a system logic controller coupled to the bus, wherein the system logic controller and the next level cache memory are fabricated on the same integrated circuit chip.
5. The computer system of claim 1, wherein the control circuit comprises a circuit path for routing the second clock signal, wherein the circuit path is located such that the second clock signal is synchronized with data values written to or read from the DRAM array.
6. The computer system of claim 1, wherein the frequency of the second clock signal is higher than the frequency of the first clock signal.
7. The computer system of claim 1, further comprising a DRAM main memory coupled to the bus.
8. The computer system of claim 7, further comprising a system controller, wherein the DRAM main memory is coupled to the bus through a direct data port of the system controller.
9. The computer system of claim 1, wherein the control circuit further comprises: a phase locked loop (PLL) circuit coupled to the bus, wherein the PLL circuit generates the second clock signal in response to the first clock signal.
10. The computer system of claim 9, wherein the second clock signal has a frequency higher than the frequency of the first clock signal.
11. The computer system of claim 1, wherein the data input port is clocked twice during each cycle of the second clock signal and the data output port is clocked once during each cycle of the first clock signal.
12. The computer system of claim 11, wherein the frequency of the first clock signal is equal to the frequency of the second clock signal.
13. The computer system of claim 1, wherein the next level cache memory further comprises a refresh counter which stores the address of a DRAM array which is currently being refreshed.
14. The computer system of claim 13, wherein the next level cache memory further comprises a refresh collision comparator which compares the address of the DRAM array currently being refreshed with an address of a DRAM array requested by the CPU.
15. The computer system of claim 14, wherein the next level cache memory further comprises a CPU access delay circuit coupled to the refresh collision comparator, wherein the CPU access delay circuit transmits a ready handshake signal to the bus when the current refresh operation is completed.
16. The computer system of claim 1, wherein the control circuit comprises a burst sequencer which accesses the DRAM array in an asynchronous, self-timed manner with respect to the first clock signal.
17. The computer system of claim 16, wherein the burst sequencer and the DRAM array are fabricated on the same integrated circuit chip.
18. The computer system of claim 16, wherein the burst sequencer includes circuitry to access the DRAM array at both edges of the second clock signal.
19. The computer system of claim 16, wherein the burst sequencer further comprises a data amplifier circuit coupled to the DRAM array, wherein the data amplifier circuit has a plurality of synchronous regenerative data amplifier latches.
20. The computer system of claim 16, wherein the burst sequencer further comprises: a row address register coupled to the bus, wherein the row address register receives a row address through the bus; a column address register coupled to the bus, wherein the column address register receives a column address through the bus; and a controller coupled to the bus and to the row and column address registers, wherein the controller provides a signal to the row and column address registers to cause the row and address registers to simultaneously latch the row and column addresses.
21. The computer system of claim 20, wherein the control circuit further comprises an address buffer coupled between the bus and the row and column address registers, wherein the address buffer stores row and column addresses required for a subsequent access while a current access is still in progress.
22. The computer system of claim 16, wherein the burst sequencer further comprises: a sense amplifier circuit coupled to the DRAM array; a data amplifier circuit having a plurality of data amplifiers; and a column selector coupled between the sense amplifier circuit and the data amplifier circuit, wherein the column selector can be selected to isolate the sense amplifier circuit from the data amplifier circuit.
23. The computer system of claim 22, wherein the number of data amplifiers in the data amplifier circuit is selected to allow the data amplifier circuit to simultaneously store all of the data values required for a burst access of the DRAM array.
24. The computer system of claim 22, wherein the next level cache memory comprises a plurality of DRAM arrays operating in parallel.
25. The computer system of claim 22, wherein the burst sequencer further comprises: a first delay circuit coupled to the sense amplifier circuit, wherein the first delay circuit enables the sense amplifier circuit in an asynchronous manner when data values are available from the DRAM array.
26. A computer system comprising: a central processing unit (CPU); a first level cache memory integrated with the CPU; a bus coupled to the CPU wherein data is transferred on the bus in response to a first clock signal having a first clock frequency, wherein data is transferred on the bus at a first data rate; and a next level cache memory coupled to the bus, wherein the next level cache memory comprises at least one dynamic random access memory (DRAM) array and a control circuit for controlling data transfer between the bus and the DRAM array, wherein the control circuit derives a second clock signal from the first clock signal, and causes data to be transferred to and from the DRAM array in response to a second clock signal at a second data rate which is faster than the first data rate, wherein the control circuit comprises a write buffer having a data output port coupled to the DRAM array and a data input port coupled to the bus, wherein the output port of the write buffer is clocked by the second clock signal and the input port of the write buffer is clocked by the first clock signal; wherein a first set of data values is transferred from the bus to the write buffer via the data input port and stored in the write buffer, until a second set of data values is transferred from the bus to the write buffer via the data input port, after which the first set of data values is transferred to the DRAM array via the data output port.
27. The computer system of claim 26, wherein the write buffer is multi-depth buffer capable of storing at least one set of write burst data.
28. The computer system of claim 26, wherein the frequency of the second clock signal is higher than the frequency of the first clock signal.
29. The computer system of claim 26, wherein the data input port is clocked once during each cycle of the first clock signal and the data output port is clocked twice during each cycle of the second clock signal.
30. The computer system of claim 29, wherein the frequency of the first clock signal is the same as the frequency of the second clock signal.
31. A method of using a DRAM array as a next level cache memory from a bus, the method comprising: transferring data on the bus in response to a first clock signal having a first clock frequency, wherein data is transferred on the bus at a first data rate; generating a second clock signal in response to the first clock signal; operating the DRAM array in response to the second clock signal, wherein data is transferred to and from the DRAM array at a second data rate, the second data rate being greater than the first data rate; reading a plurality of data values from the DRAM array to a plurality of sense amplifiers; reading the data values from the sense amplifiers to a read buffer memory using the second clock signal to clock the data values into the buffer memory; reading the data values from the read buffer memory to the bus using the first clock signal to clock the data values out of the read buffer memory so that the data values are read to the read buffer memory faster than the data values are read from the read buffer memory; and precharging bit lines of the DRAM array during the time that the data values are read from the read buffer memory to the bus.
32. The method of claim 31, wherein the second clock signal has a higher frequency than the first clock signal.
33. The method of claim 31, further comprising the steps of: transmitting data signals on the bus at either the rising edges or the falling edges of the first clock signal; and performing read operations within the DRAM array at both the rising and falling edges of the second clock signal.
34. The method of claim 31, wherein the frequency of the second clock signal is the same as the frequency of the first clock signal.
35. The method of claim 31, further comprising the steps of: performing a row access operation in the DRAM array in a self-timed asynchronous manner with respect to the second clock signal; performing a column decode operation in the DRAM array in a self-timed asynchronous manner with respect to the second clock signal; and performing a predetermined sequence of column select operations in the DRAM array, wherein the column select operations are synchronous with respect to the second clock signal.
36. The method of claim 31, wherein the frequency of the second clock signal is twice the frequency of the first clock signal.
37. The method of claim 31, wherein the step of precharging the DRAM array is completed before the step of reading the data values from the read buffer memory is completed.
38. The method of claim 31, further comprising the steps of: transmitting data signals on the bus at either the rising edges or the falling edges of the first clock signal; and performing write operations within the DRAM array at both the rising and falling edges of the second clock signal.
39. The method of claim 38, wherein the frequency of the second clock signal is the same as the frequency of the first clock signal.
40. The method of claim 31, further comprising the steps of: reading a plurality of data values from the DRAM array to a plurality of sense amplifiers; closing a plurality of column selector switches, thereby connecting the sense amplifiers to a plurality of data amplifiers, thereby causing the data values to be written to the data amplifiers; reading the data values out of the data amplifiers; opening the column selector switches, thereby disconnecting the sense amplifiers from the data amplifiers; and precharging a plurality of DRAM cells in the DRAM array when the column selector switches are open.
41. The method of claim 40, wherein the step of reading the data values out of the data amplifiers further comprises the steps of: writing the data values from the data amplifiers to the read buffer memory using the second clock signal to clock the data values into the read buffer; and reading the data values form the read buffer memory to the bus using the first clock signal to clock the data values onto the bus.
42. The method of claim 41, wherein the step of precharging the DRAM array is completed before the step of reading the data values from the read buffer to the bus is completed.
43. A method of using a DRAM array as a next level cache memory from a bus, the method comprising: transferring data on the bus in response to a first clock signal having a first clock frequency, wherein data is transferred on the bus at a first data rate; generating a second clock signal in response to the first clock signal; operating the DRAM array in response to the second clock signal, wherein data is transferred to and from the DRAM array at a second data rate, the second data rate being greater than the first data rate; writing a first set of data values from the bus into a write buffer using the first clock signal to clock the first set of data values into the write buffer; writing a second set of data values from the bus into the write buffer using the first clock signal to clock the second set of data values into the write buffer; writing the first set of data values to the DRAM array after the time the second set of data values are written into the write buffer; and precharging bit lines of the DRAM array during the time the second set of data values are written into the write buffer.
44. The method of claim 43, wherein the step of precharging the DRAM array is completed before the step of writing the second set of data values into the write buffer is completed.
45. The method of claim 43, wherein the step of synchronously writing the first set of data values to the DRAM array further comprises the steps of: providing a data amplifier circuit and a sense amplifier circuit between the write buffer and the DRAM array; disconnecting the data amplifier circuit from the sense amplifier circuit; writing the first set of data values from the write buffer to the data amplifier circuit at the same time as performing a row access operation in the DRAM array; and connecting the data amplifier circuit to the sense amplifier circuit, thereby causing the first set of data values to be provided to the DRAM array through the sense amplifier circuit.
46. A computer system comprising: a central processing unit (CPU); a first level cache memory integrated with the CPU; a bus coupled to the CPU wherein data is transferred on the bus in response to a first clock signal having a first clock frequency, wherein data is transferred on the bus at a first data rate; and a next level cache memory coupled to the bus, wherein the next level cache memory comprises at least one dynamic random access memory (DRAM) array and a control circuit for controlling data transfer between the bus and the DRAM array, wherein the control circuit derives a second clock signal from the first clock signal, and causes data to be transferred to and from the DRAM array in response to a second clock signal at a second data rate which is faster than the first data rate, the control circuit comprising a phase locked loop (PLL) circuit coupled to the bus, wherein the PLL circuit generates the second clock signal in response to the first clock signal, and wherein the second clock signal has the same frequency as the first clock signal.
47. A computer system comprising: a central processing unit (CPU); a first level cache memory integrated with the CPU; a bus coupled to the CPU wherein data is transferred on the bus in response to a first clock signal having a first clock frequency, wherein data is transferred on the bus at a first data rate; and a next level cache memory coupled to the bus, wherein the next level cache memory comprises at least one dynamic random access memory (DRAM) array and a control circuit for controlling data transfer between the bus and the DRAM array, wherein the control circuit derives a second clock signal from the first clock signal, and causes data to be transferred to and from the DRAM array in response to a second clock signal at a second data rate which is faster than the first data rate, wherein the control circuit comprises a burst sequencer which accesses the DRAM array in an asynchronous, self-timed manner with respect to the first clock signal; wherein the burst sequencer further comprises: a sense amplifier circuit coupled to the DRAM array; a data amplifier circuit having a plurality of data amplifiers and a column selector coupled between the sense amplifier circuit and the data amplifier circuit, wherein the column selector can be selected to isolate the sense amplifier circuit from the data amplifier circuit; wherein the burst sequencer further comprises: a first delay circuit coupled to the sense amplifier circuit, wherein the first delay circuit enables the sense amplifier circuit in an asynchronous manner when the data values are available from the DRAM array; and a second delay circuit coupled to the column selector, wherein the second delay circuit enables the column selector in an asynchronous manner when the data values are available in the send amplifier circuit.
Patent Citations (10)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US4464715 | 1984-08-01 | Stamm | |
| US5113506 | 1992-05-01 | Moussouris et al. | |
| US5164621 | 1992-11-01 | Miyamoto | |
| US5249282 | 1993-09-01 | Segers | |
| US5265102 | 1993-11-01 | Saito | |
| US5293603 | 1994-03-01 | Mac Williams et al. | |
| US5357479 | 1994-10-01 | Matsui | |
| US5422858 | 1995-06-01 | Mizukami et al. | |
| US5463584 | 1995-10-01 | Hoshino | |
| US5511031 | 1996-04-01 | Grover et al. |
Non-Patent Literature (9)
- INTEL, Pentium.sup.198 Processor 3.3V Pipelined BSRAM Specification, Oct. 5, 1994, pp. 1-39.
- IDT, Advanced Information IDT71V432, "32K.times.32 CacheRAM.sup..TM. Pipelined/Flow Through Outputs Burst Counter, & Self-Timed Write-for Pentium.sup..TM. /PowerPC.sup..TM. Processors," May 1994, pp. 1-2.
- IDT, Preliminary IDT71420, "32K.times.18 CacheRAM.sup..TM. Burst Counter & Self-Timed Write-For the Pentium.sup..TM. Processor", May 1994, pp. 1-8.
- Przybylski, Steven A., "New DRAM Technologies, A Comprehensive Analysis of the New Architectures," MicroDesign Resources, pp. 111-216, 357-578.
- Preliminary Data Sheet, KM741006J CMOS SRAM, "262,144 Words.times.4-Bit Synchronous Static Random Access Memory," Samsung Electronics, pp. 348-354 (date and author unknown).
- OPTi 82C803/82C822 PCI Chipset Overview, OPTi, Inc., Jun. 3, 1993, pp. 4-8.
- OPTi 486SXWB Overview (82C493/82C392/82C206 Chipset), OPTi, Inc., pp. 5-11.
- INTEL 82420 PCIset, Advance Information, Intel Corporation, Apr. 1993, pp. 1, 12-15, 27-32.
- INTEL 82430 PCIset for the Pentium.sup..TM. Processor, Advance Information, Intel Corporation, Mar., 1993, pp. 1-7.