Method and Apparatus for Delaying Write Operations
An apparatus for controlling a dynamic random access memory (DRAM), the apparatus comprising an interface to transmit to the DRAM a first code to indicate that first data is to be written to the DRAM. The first code is to be sampled by the DRAM and held by the DRAM for a first period of time before it is issued inside the DRAM. The interface is further to transmit the first data that is to be sampled by the DRAM after a second period of time has elapsed from when the first code is sampled by the DRAM. The interface is further to transmit a second code, different from the first code, to indicate that second data is to be read from the DRAM. The second code is to be sampled by the DRAM on one or more edges of the external clock signal.
This application is a continuation of U.S. application Ser. No. 13/230,741, filed Sep. 12, 2011, which is a continuation of U.S. application Ser. No. 12/875,483, filed Sep. 3, 2010, now U.S. Pat. No. 8,019,958, which is a continuation of U.S. application Ser. No. 12/349,485, filed Jan. 6, 2009, now U.S. Pat. No. 7,793,039, which is a continuation of U.S. patent application Ser. No. 11/953,803, filed Dec. 10, 2007, now U.S. Pat. No. 7,496,709, which is a continuation of U.S. patent application Ser. No. 11/692,159, filed Mar. 27, 2007, now U.S. Pat. No. 7,330,952, which is a continuation of U.S. patent application Ser. No. 11/059,216, filed Feb. 15, 2005, now U.S. Pat. No. 7,197,611, which is a continuation of U.S. patent application Ser. No. 10/128,167, filed Apr. 22, 2002, now U.S. Pat. No. 6,868,474, which is a divisional of U.S. patent application Ser. No. 09/169,206, filed Oct. 9, 1998, now U.S. Pat. No. 6,401,167, which claims priority to U.S. Provisional Patent Application No. 60/061,770, filed Oct. 10, 1997, all of which are herein incorporated by referenced in their entirety.
FIELD OF THE INVENTIONThe present invention relates generally to electronic systems for data storage and retrieval. More particularly, the invention is directed toward improved methods and structures for memory devices.
BACKGROUND OF THE INVENTIONIn any engineered design there are compromises between cost and performance. The present invention introduces novel methods and structures for reducing the cost of memory devices while minimally compromising their performance. The description of the invention requires a significant amount of background including: application requirements, memory device physical construction, and memory device logical operation.
Memory device application requirements can be most easily understood with respect to memory device operation.
For purposes of illustrating the invention a conventional DRAM core will be described.
A conventional DRAM core 202 mainly comprises storage banks 211 and 221, row decoder and control circuitry 210, and column data path circuit comprising column amplifiers 260 and column decoder and control circuitry 230. Each of the storage banks comprises storage arrays 213 and 223 and sense amplifiers 212 and 222.
There may be many banks, rather than just the two illustrated. Physically the row and column decoders may be replicated in order to form the logical decoder shown in
The operation of a conventional DRAM core is divided between row and column operations. Row operations control the storage array word lines 241 and the sense amplifiers via line 242. These operations control the movement of data from the selected row of the selected storage array to the selected sense amplifier via the bit lines 251 and 252. Column operations control the movement of data from the selected sense amplifiers to and from the external data connections 204d and 204e.
Device selection is generally accomplished by one of the following choices:
-
- matching an externally presented device address against an internally stored device address;
- requiring separate operation control lines, such as RAS and CAS, for each set of memory devices that are to be operated in parallel; and
- providing at least one chip select control on the memory device.
-
- precharge(device, bank)—prepare the selected bank of the selected device for sensing; and
- sense(device, bank, row)—sense the selected row of the selected bank of the selected device.
The operations and device selection arguments are presented to the core via the PRECH and SENSE timing signals while the remaining arguments are presented as signals which have setup and hold relationships to the timing signals. Specifically, as shown in
-
- data=read(device, bank, column)—transfer the data in the subset of the sense amplifiers specified by “column” in the selected “bank” of the selected “device” to the READDATA lines; and
- write (device, bank, column, mask, data)—store the data presented on the WRITEDATA lines into the subset of the sense amplifiers specified by “column” in the selected “bank” of the selected “device”; optionally store only a portion of the information as specified by “mask”.
More recent conventional DRAM cores allow a certain amount of concurrent operation between the functional blocks of the core. For example, it is possible to independently operate the precharge and sense operations or to operate the column path simultaneously with row operations. To take advantage of this concurrency each of the following groups may operate somewhat independently:
-
- PRECH and PRECHBANK on lines 204a;
- SENSE, SENSEBANK, and SENSEROW on lines 204b;
- COLCYC 204f on line, COLLAT and COLADDR on lines 204g, WRITE and WMASK one lines 204c, READDATA on line 204d, and WRITEDATA on line 204.
There are some restrictions on this independence. For example, as shown in
The present invention, while not limited by such values, has been optimized to typical values as shown in Table 1.
The series of memory operations needed to satisfy any application request can be covered by the nominal and transitional operation sequences described in Table 2 and Table 3. These sequences are characterized by the initial and final bank states as shown in
The sequence of memory operations is relatively limited. In particular, there is a universal sequence:
-
- precharge,
- sense,
- transfer (read or write), and
- close.
In this sequence, close is an alternative timing of precharge but is otherwise functionally identical. This universal sequence allows any sequence of operations needed by an application to be performed in one pass through it without repeating any step in that sequence. A control mechanism that implements the universal sequence can be said to be conflict free. A conflict free control mechanism permits a new application reference to be started for every minimum data transfer. That is, the control mechanism itself will never introduce a resource restriction that stalls the memory requestor. There may be other reasons to stall the memory requestor, for example references to different rows of the same bank may introduce bank contention, but lack of control resources will not be a reason for stalling the memory requestor
Memory applications may be categorized as follows:
-
- main memory—references generated by a processor, typically with several levels of caches;
- graphics—references generated by rendering and display refresh engines; and
- unified—combining the reference streams of main memory and graphics.
Applications may also be categorized by their reference stream characteristics. According to the application partition mentioned above reference streams can be characterized in the following fashion:
-
- First, main memory traffic can be cached or uncached processor references. Such traffic is latency sensitive since typically a processor will stall when it gets a cache miss or for any other reason needs data fetched from main memory. Addressing granularity requirements are set by the transfer size of the processor cache which connects to main memory. A typical value for the cache transfer size is 32 bytes. Since multiple memory interfaces may run in parallel it is desirable that the memory system perform well for transfer sizes smaller than this. Main memory traffic is generally not masked; that is, the vast bulk of its references are cache replacements which need not be written at any finer granularity than the cache transfer size.
- Another type of reference stream is for graphics memory. Graphics memory traffic tends to be bandwidth sensitive rather than latency sensitive. This is true because the two basic graphics engines, rendering and display refresh, can both be highly pipelined. Latency is still important since longer latency requires larger buffers in the controller and causes other second order problems. The ability to address small quanta of information is important since typical graphics data structures are manipulated according to the size of the triangle being rendered, which can be quite small. If small quanta cannot be accessed then bandwidth will be wasted transferring information which is not actually used. Traditional graphics rendering algorithms benefit substantially from the ability to mask write data; that is, to merge data sent to the memory with data already in the memory. Typically this is done at the byte level, although finer level, e.g. bit level, masking can sometimes be advantageous.
As stated above, unified applications combine the characteristics of main memory and graphics memory traffic. As electronic systems achieve higher and higher levels of integration the ability to handle these combined reference streams becomes more and more important.
Although the present invention can be understood in light of the previous application classification, it will be appreciated by those skilled in the art that the invention is not limited to the mentioned applications and combinations but has far wider application. In addition to the specific performance and functionality characteristics mentioned above it is generally important to maximize the effective bandwidth of the memory system and minimize the service time. Maximizing effective bandwidth requires achieving a proper balance between control and data transport bandwidth. The control bandwidth is generally dominated by the addressing information delivered to the memory device. The service time is the amount of time required to satisfy a request once it is presented to the memory system. Latency is the service time of a request when the memory system is otherwise devoid of traffic. Resource conflicts, either for the interconnect between the requestor and the memory devices, or for resources internal to the memory devices such as the banks, generally determine the difference between latency and service time. It is desirable to minimize average service time, especially for processor traffic.
The previous section introduced the performance aspects of the cost-performance tradeoff that is the subject of the present invention. In this section the cost aspects are discussed. These aspects generally result from the physical construction of a memory device, including the packaging of the device.
There are many negative aspects to the increase in the length of the package wiring 1640, including the facts that: the overall size of the package increases, which costs more to produce and requires more area and volume when the package is installed in the next level of the packaging hierarchy, such as on a printed circuit board. Also, the stub created by the longer package wiring can affect the speed of the interconnect. In addition, mismatch in package wiring lengths due to the fan-in angle can affect the speed of the interconnect due to mismatched parasitics.
The total number of signal pins has effects throughout the packaging hierarchy. For example, the memory device package requires more material, the next level of interconnect, such as a printed circuit board, requires more area, if connectors are used they will be more expensive, and the package and die area of the master device will grow.
In addition to all these cost concerns based on area and volume of the physical construction another cost concern is power. Each signal pin, especially high speed signal pins, requires additional power to run the transmitters and receivers in both the memory devices as well as the master device. Added power translates to added cost since the power is supplied and then dissipated with heat sinks.
The memory device illustrated in
-
- write empty—sense 1851, write 1853 with mask 1871, data 1881, close (precharge) 1861,
- write miss—precharge 1852, sense 1854, write 1856 with mask 1872, data 1882;
- read hit—read 1857, tristate control 1873, data 1883; and
- transitional write miss—precharge 1855, sense 1858, write 1859, mask 1874, data 1884, close (precharge) 1862.
In
In addition to illustrating a specific type of prior art memory device,
The technique of specifying the burst size in a register makes it difficult to mix transfer sizes unless the burst size is always programmed to be the minimum, which then increases control overhead. The increase in control overhead may be so substantial as to render the minimum burst size impractical in many system designs.
Regardless of the transfer burst size, the technique of a single unified control bus, using various combinations of the command pins 1810, address pins 1820, and mask pins 1830 places limitations on the ability to schedule the primitive operations. A controller which has references in progress that are simultaneously ready to use the control resources must sequentialize them, leading to otherwise unnecessary delay.
Read operations do not require masking information. This leaves the mask pins 1830 available for other functions. Alternately, the mask pins during read operations may specify which bytes should actually be driven across the pins as illustrated by box 1873.
Another technique is an alternative method of specifying that a precharge should occur by linking it to a read or write operation. When this is done the address components of the precharge operation need not be respecified; instead, a single bit can be used to specify that the precharge should occur. One prior art method of coding this bit is to share an address bit not otherwise needed during a read or write operation. This is illustrated by the “A-Prech” boxes, 1861 and 1862.
Another technique makes the delay from write control information to data transfer different from the delay of read control information to data transfer. When writes and reads are mixed, this leads to difficulties in fully utilizing the data pins.
Thus, current memory devices have inadequate control bandwidth for many application reference sequences. Current memory devices are unable to handle minimum size transfers. Further, current memory devices utilize the available control bandwidth in ways that do not support efficient applications. Current memory devices do not schedule the use of the data pins in an efficient manner. In addition, current memory devices inefficiently assign a bonding pad for every pin of the device.
Like reference numerals refer to corresponding parts throughout the drawings.
DESCRIPTION OF EMBODIMENTS-
- in one embodiment according to the present invention, a register within the device specifies whether the mask pins are to be used for masking, tristate control, or precharge control;
- in another embodiment according to the present invention, the encoding of the command pins is extended to specify, on a per operation basis, how the mask pins are to be used; and
- in another embodiment according to the present invention, a register bit indicates whether tristate control is enabled or not and, in the case it is not enabled, an encoding of the command pins indicates if a write is masked or not; in this embodiment all reads and unmasked writes may use the Mask pins to specify a precharge operation while masked writes do not have this capability since the Mask pins are used for mask information
There are many alternatives for how to code the precharge information on the mask pins. In one embodiment in which there are two mask pins and the memory device has two banks, one pin indicates whether an operation should occur and the other pin indicates which bank to precharge. In an alternative embodiment, in which the minimum data transfer requires more than one cycle, more banks are addressed by using the same pins for more than one cycle to extend the size of the bank address field.
Using the mask pins to specify a precharge operation and the associated bank address requires another way of specifying the device argument. In one embodiment the device is specified in some other operation. For example, the precharge specified by the mask pins shares device selection with a chip select pin that also conditions the main command pins. In another embodiment, additional control bandwidth is added to the device. For example, an additional chip select pin is added for sole use by the recoded mask pin precharge. In yet another example of using additional control bandwidth in which the minimum data transfer requires more than one cycle, the device address is coded on the additional bits, the device address being compared to an internal device address register.
In
This change in resource ordering gives rise to resource conflict problems that produce data bubbles when mixing reads and writes. The resource ordering of writes generally leads to the resource timing shown in
The read resource timing of
Note that the data bubble appears regardless of whether the write 2642 and the read 2643 are directed to the same or different memory devices on the channel. Further note that the delay from the control resource to the column i/o resource is identical for reads and writes. In view of this, it is impossible for the data resource timing to be identical for reads and writes.
Matching the timing of the write-use of the data resource to the read-use of the data resource avoids the problem stated above. Since the use of the data pins in a system environment has an intrinsic turnaround time for the external interconnect, the optimal delay for a write does not quite match the delay for a read. Instead, it should be the minimum read delay minus the minimum turnaround time. Since the turnaround delay grows as the read delay grows, there is no need to change the write control to data delay as a function of the memory device position on the channel.
Since write latency is not an important metric for application performance, as long as the write occurs before the expiration of tRAS,MIN (so that it does not extend the time the row occupies the sense amplifiers, which reduces application performance), this configuration does not cause any loss in application performance, as long as the writes and reads are directed to separate column data paths.
Delayed writes help optimize data bandwidth efficiency over a set of bidirectional data pins. One method adds delay between the control and write data packets so that the delay between them is the same or similar as that for read operations. Keeping this “pattern” the same or similar for reads and writes improves pipeline efficiency over a set of bidirectional data pins, but at the expense of added complexity in the interface.
-
- in the sense operation field on the primary control lines 3104, as an alternative to the sense information;
- in the mask field on the secondary control lines, 3105 as an alternative to the mask information; and
- according to the device and bank addresses specified in a read or a write.
The benefit of the present invention according to a specific embodiment is shown in Table 4 and
The operation of this embodiment can be most easily understood through various timing diagrams as shown in
-
FIG. 30 throughFIG. 35 show a basic operation as an embodiment of the present invention, other operations can be thought of as compositions of these basic operations;FIG. 36 throughFIG. 39 show compositions of the basic operations but distinct from notions of the universal sequence;FIG. 40 throughFIG. 43 show operations according to the universal sequence, these figures demonstrate the ability of the embodiment to handle mixed read and write with mixed hit, miss, and empty traffic without control resource conflicts; andFIG. 44 throughFIG. 45 show operations according to the universal sequence demonstrating less control conflicts than the prior art. Other control scheduling algorithms are possible which seek to minimize other metrics, such as service time, with or without compromising effective bandwidth.
The nominal timings for the examples are shown in Table 7.
A description of each of the timing diagrams follows.
The previous figures show how various application references can be decomposed into the memory operations.
The previous figures demonstrate the conditions in which the universal sequence can be scheduled. The ability to schedule the universal sequence guarantees that there will not be any control conflicts which reduce available data transfer bandwidth. However, none of the nominal reference sequences actually requires two precharges to be scheduled. So there is generally adequate control bandwidth for various mixes of miss and empty traffic as shown in
According to an embodiment of the present invention the memory device of
As is shown in Table 10, the memory device is selected by the SDEV field and the SOP field determines the Auxiliary Operation to be performed by the Register Operation Unit 5309 in
According to an embodiment of the present invention the memory device of
In an alternate embodiment, each memory device receives a different CMD signal, one for each device, rather than using the data input field via path 5307 to identify the device for a ExitToNormal or ExitToDrowsy operation.
The foregoing description, for purposes of explanation, used specific nomenclature to provide a thorough understanding of the invention. However, it will be apparent to one skilled in the art that the specific details are not required in order to practice the invention. In other instances, well known circuits and devices are shown in block diagram form in order to avoid unnecessary distraction from the underlying invention. Thus, the foregoing descriptions of specific embodiments of the present invention are presented for purposes of illustration and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed, obviously many modifications and variations are possible in view of the above teachings. The embodiments were chosen and described in order to best explain the principles of the invention and its practical applications, to thereby enable others skilled in the art to best utilize the invention and various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the following claims and their equivalents.
Claims
1. An apparatus for controlling a dynamic random access memory (DRAM), the apparatus comprising:
- an interface to transmit to the DRAM: over a first plurality of wires, a first code to indicate that first data is to be written to the DRAM, wherein the first code is to be sampled by the DRAM on one or more edges of an external clock signal received by the DRAM and then held by the DRAM for a first period of time before it is issued inside the DRAM; over a second plurality of wires, the first data, wherein the first data is to be sampled by the DRAM after a second period of time has elapsed from when the first code is sampled by the DRAM, and wherein the second plurality of wires is separate from the first plurality of wires; and a second code over the first plurality of wires, the second code to indicate that second data is to be read from the DRAM, wherein the second code is to be sampled by the DRAM on one or more edges of the external clock signal, wherein the second code is a different value than the first code.
2. The apparatus of claim 1, wherein the first period of time is selected to equalize an offset from the first code to the first data with an offset from the second code to the second data.
3. The apparatus of claim 2, wherein holding the first code for the first period of time reduces a time gap between the first data and the second data to a predefined time gap in accordance with channel turnaround requirements.
4. The apparatus of claim 1, wherein:
- the interface receives the second data after a read delay time transpires from when the second code was transmitted; and
- the second period of time is based on the read delay time.
5. The apparatus of claim 4, wherein the second period of time is based on the read delay time such that the second period of time matches the read delay time minus a channel turnaround time.
6. A method of controlling a dynamic random access memory (DRAM), comprising:
- transmitting, to the DRAM, a first code over a first plurality of wires, the first code to indicate that first data is to be written to the DRAM, wherein the first code is to be sampled by the DRAM on one or more edges of an external clock signal received by the DRAM held for a first period of time before it is issued inside the DRAM;
- transmitting, to the DRAM, the first data over a second plurality of wires, wherein the first data is to be sampled by the DRAM after a second period of time has elapsed after the first code is sampled by the DRAM, wherein the second plurality of wires is separate from the first plurality of wires; and
- transmitting, to the DRAM, a second code over the first plurality of wires, the second code to indicate that second data is to be read from the DRAM, wherein the second code is to be sampled by the DRAM on one or more edges of the external clock signal, and wherein the second code is a different value than the first code.
7. The method of claim 6, wherein the first period of time is selected to equalize an offset from the first code to the first data with an offset from the second code to the second data.
8. The method of claim 7, wherein holding the first code for the first period of time reduces a time gap between the first data and the second data to a predefined time gap in accordance with channel turnaround requirements.
9. The method of claim 6, wherein:
- the interface receives the second data after a read delay time transpires from when the second code was transmitted; and
- the second period of time is based on the read delay time.
10. The method of claim 9, wherein the second period of time is based on the read delay time such that the second period of time matches the read delay time minus a channel turnaround time.
11. A dynamic random access memory (DRAM) device, comprising:
- an interface to receive:
- a first code to indicate that first data is to be written to the DRAM, wherein the first code is to be sampled by the DRAM on one or more edges of an external clock signal received by the DRAM and then held for a first period of time before it is issued inside the DRAM;
- the first data, wherein the first data is to be sampled by the DRAM after a second period of time has elapsed after the first code is sampled by the DRAM; and
- a second code to indicate that second data is to be read from the DRAM, wherein the second code is to be sampled by the DRAM on one or more edges of the external clock signal, wherein the second code is a different value than the first code.
12. The DRAM device of claim 11, wherein the first period of time is selected to equalize an offset from the first code to the first data with an offset from the second code to the second data.
13. The DRAM device of claim 12, wherein holding the first code for the first period of time reduces a time gap between the first data and the second data to a predefined time gap in accordance with channel turnaround requirements.
14. The DRAM device of claim 11, wherein:
- the interface transmits the second data so that it is received at a controller apparatus after a read delay time transpires from when the second code was transmitted by the controller apparatus; and
- the second period of time is based on the read delay time.
15. The DRAM device of claim 14, wherein the second period of time is based on the read delay time such that the second period of time matches the read delay time minus a channel turnaround time.
16. The DRAM device of claim 11, wherein the interface is to receive the first data and transmit the second data via a set of bidirectional data pins.
17. A method of operating a dynamic random access memory (DRAM), comprising:
- receiving, at an interface of the DRAM, a first code to indicate that first data is to be written to the DRAM, wherein the first code is to be sampled by the DRAM on one or more edges of an external clock signal received by the DRAM and then held for a first period of time before it is issued inside the DRAM;
- receiving, at the interface of the DRAM, the first data, wherein the first data is to be sampled by the DRAM after a second period of time has elapsed after the first code is sampled by the DRAM; and
- receiving, at the interface of the DRAM, a second code to indicate that second data is to be read from the DRAM, wherein the second code is to be sampled by the DRAM on one or more edges of the external clock signal, wherein the second code is a different value than the first code.
18. The method of claim 17, wherein the first period of time is selected to equalize an offset from the first code to the first data with an offset from the second code to the second data.
19. The method of claim 18, wherein holding the first code for the first period of time reduces a time gap between the first data and the second data to a predefined time gap in accordance with channel turnaround requirements.
20. The method of claim 17, wherein:
- the interface transmits the second data so that it is received at a controller apparatus after a read delay time transpires from when the second code was transmitted by the controller apparatus; and
- the second period of time is based on the read delay time.
21. The method of claim 20, wherein the second period of time is based on the read delay time such that the second period of time matches the read delay time minus a channel turnaround time.
22. The method of claim 17, wherein the first data is to be received via a set of bidirectional data pins and the second data is to be transmitted via the set of bidirectional data pins.
Type: Application
Filed: Mar 15, 2012
Publication Date: Jul 5, 2012
Inventors: Richard M. Barth (Palo Alto, CA), Frederick A. Ware (Los Altos Hills, CA), Donald C. Stark (Los Altos, CA), Craig E. Hampel (San Jose, CA), Paul G. Davis (San Jose, CA), Abhijit M. Abhyankar (Sunnyvale, CA), James A. Gasbarro (Mountain View, CA), David Nguyen (San Jose, CA)
Application Number: 13/421,753
International Classification: G06F 12/00 (20060101);