 |
Description  |
|
|
BACKGROUND OF THE INVENTION
This invention relates to computer memories and more particularly to a DRAM
memory subsystem capable of minimum wait state accessing by a high
performance device outputting a series of row and column addresses for
data to be accessed comprising, a DRAM comprised of rows and columns and
operable in a static column mode in response to address requests asserted
at RAS and CAS inputs thereof; and, access optimization logic means
operably connected for receiving a first row and column address as output
by the device, for asserting the first row and column address at the RAS
and CAS inputs, for receiving a second row and column address as output by
the device, for changing the column address being asserted at the CAS
input to correspond to the second column address, for checking the second
row address against the first row address, and for reasserting the second
row and column address at the RAS and CAS inputs if the second row address
is not the same as the first row address.
As depicted in FIGS. 1 and 2, it is known in the art to employ either a
static RAM 10 or a dynamic RAM 12 to serve as a memory subsystem for a
central processing unit (CPU) 14 which processes instructions contained in
the RAM 10, 12 as well as reading from and writing to that memory.
Commercial designs for instruction processors which are used in
contemporary general purpose, high performance computing systems, however,
require fast random access memory subsystems to sustain maximum
performance. The instruction processor provides the memory subsystem with
an address which identifies a location in the memory subsystem where data
is stored. The memory subsystem must then provide quick access to the
location so that the data stored in the location can be read from or
written to by the processor. If the memory subsystem is incapable of
accessing data at the rate which the processor requires for peak
performance, the processor must wait for the memory access operation to
complete. The time spent by the processor waiting for data directly
decreases the performance of the computing system. This is a classic
example of the old adage that a chain is only as strong as its weakest
link. No matter how fast the computer processor may be, it can only
operate as fast as it can access its data and instructions from memory.
The customary units for specifying the memory access time are clock cycles.
The instruction processor executes instructions in one or more clock
cycles. State-of-the-art designs such as so-called reduced instruction set
computers (RISCs) strive to execute one instruction per clock cycle. Since
each load and store instruction requires an access to the memory
subsystem, these operations must also complete in one clock cycle if this
design objective is to be achieved. For the same reasons, instructions
from the memory subsystem must be delivered to the processor for execution
at an average rate of one instruction per clock cycle.
A typical value for the clock period of a state-of-the-art design is 50
nanoseconds (ns). For example, the Motorola model 68030 instruction
processor chip has a 50 ns clock cycle and the memory cycle time is less
than two clock cycles or 100 ns. For the Motorola 68030 to operate without
waiting for data, the Motorola company states, "When the MC68030 is
operating at high clock frequency, a no-wait-state external memory
subsystem will of necessity be composed of static RAMs." (MC68030 User's
Manual, C. 1987, Sections 12-14). In other words, the Motorola company,
like everyone else skilled in the art, is of the opinion that high
performance instruction processors (and other high performance devices
having to access memory subsystems) will not operate at their maximum
performance capability (i.e. with no waiting for memory accesses) with
dynamic random access memory (DRAM) as the memory.
The static RAM (SRAM) is a random access memory which provides access times
from as low as 12 ns (but typically 45 ns) for state-of-the-art
components. SRAM is built employing memory cells which each require six
transistors. In contrast, DRAM is built of cells each requiring a single
transistor; but, has a typical access time of 100 ns for state-of-the-art
components with a cycle time of 200 ns. The cycle time indicates the
maximum rate at which the DRAM can respond to memory access requests. A
peculiarity of DRAM memory chips is that the devices require a significant
"precharge" time when the row address is changed. Thus, as evidenced by
the figures quoted above, DRAM cycle time is typically twice the time
required to access the datum. In comparison, SRAM has a cycle time which
is only slightly longer than its access time. The primary advantages of
the DRAM over the SRAM are density and price; that is, more memory can be
placed into the same space with DRAM because of the 6:1 reduction in the
number of transistors for each cell of the memory. Obviously, the simpler
design also results in a substantial cost reduction as well. Generally
speaking, DRAM affords a 4:1 advantage in density and a 5:1 advantage in
price. Such an advantage makes the use of DRAM over SRAM very desirable
when possible. But, as we have seen, those skilled in the art have
considered DRAM as unsuitable for use with high performance devices.
The DRAM market is highly competitive and the manufacturers of DRAM chips
have produced novel variations on the customary organization of the DRAM
to gain customer acceptance. One common variation is called "static column
mode". Internally, as depicted in FIG. 3, a DRAM 12 is organized as a
two-dimensional array of rows 16 and columns 18. The memory address
employed for reading and writing the DRAM 12 is partitioned into a row
address and a column address. With a static RAM 10, the row and column
addresses are input in parallel as depicted in FIG. 4. As depicted in FIG.
5, however, in the dynamic RAM 12 the row address is first strobed into
the memory device followed by the column address. As depicted in FIG. 6,
this is accomplished by the CPU 14 providing its request to address
forming logic 20 which, in turn, transmits the address information over
the unmultiplexed address line or address bus 22 to a multiplexer (MUX)
24. The MUX 24 transmits the row address and the column address to the
DRAM 12. To access another memory location requires both strobes to be
removed for a preset time period and then be reapplied in the same
sequence, as depicted in FIG. 7. For static column mode operation, the
first step remains the same; once the row address has been strobed in,
however, any data within the column can be randomly accessed without
changing the row and column strobes. The net effect is that any data
within the column, called a "page", can be accessed as if the data were
stored in SRAM.
The principal application for static column mode operation in DRAMs is to
provide copying of data from one page to another page, i.e. with the row
address remaining the same, only the column address need be changed. This
mode is particularly useful for disk controllers and other peripherals
which employ direct memory to memory copy operations. The access and cycle
times for DRAM operating in static column mode are competitive with purely
static RAM parts.
Wherefore, it is the main object of the present invention to provide an
interface and method of operation which will permit DRAM to be employed in
place of SRAM with high performance devices without high risk of imposing
wait conditions on the devices thereby.
It is another object of the present invention to employ the static column
mode of operation in DRAMs in a manner which will provide high performance
devices with a high density, low cost memory subsystem having a
statistically low probability of requiring the DRAM to change pages.
Other objects and benefits of the present invention will be recognized from
the description which follows hereinafter when taken in conjunction with
the drawing figures which accompany it.
SUMMARY
The foregoing objects have been achieved in a DRAM memory subsystem having
a DRAM comprised of rows and columns and operable in a static column mode
in response to address requests from a high performance device outputting
a series of row and column addresses for data to be accessed, by the
method of operation of the present invention to provide minimum wait state
accessing comprising the steps of, beginning each new access request from
the device in a continuing static column mode access by only changing the
column address; checking the row address being employed in the continuing
static column mode access against the row address of the new access
request; continuing with the current continuing static column mode access
if the row addresses are the same; and, terminating the current continuing
static column mode access and beginning a new static column mode access
with the row and column addresses of the new access request if the row
addresses are not the same.
In the preferred embodiment, the row and column addresses of each access
request are asserted at RAS and CAS inputs of a DRAM and the steps
comprise, receiving a first row and column address as output by the
device; asserting the first row and column address at the RAS and CAS
inputs; receiving a second row and column address as output by the device;
changing the column address being asserted at the CAS input to correspond
to the second column address; checking the second row address against the
first row address; and, reasserting the second row and column address at
the RAS and CAS inputs if the second row address is not the same as the
first row address.
As necessary in the particular application the method also includes
advising the device of aborting and retrying an access to the DRAM if the
second row address is not the same as the first row address.
The method also is applicable where there are a plurality of the DRAMs each
identifiable by a bank number and the device includes a bank address as
part of each row and column address. In such case, it additionally
comprises the steps of, storing the presently being asserted row address
for each bank in a memory table; storing the presently being asserted row
address for each bank in the memory table; changing the column address
being asserted at the CAS input on the DRAM indicated by the bank address
portion of second address from the device to correspond to the second
column address; and, checking the second row address against the first row
address as stored in the memory table for the bank indicated by the bank
address portion of second address from the device.
DESCRIPTION OF THE DRAWINGS
FIG. 1 is a simplified block diagram of a CPU connected to a static RAM as
is known in the prior art.
FIG. 2 is a simplified block diagram of a CPU connected to a dynamic RAM as
is known in the prior art.
FIG. 3 is a simplified drawing depicting the row and column two-dimensional
addressing environment of a RAM as is known in the prior art.
FIG. 4 is a simplified drawing depicting how a static RAM receives its row
and column address information in parallel resulting in faster access to
the memory as is known in the prior art.
FIG. 5 is a simplified drawing depicting how a dynamic RAM first receives
its row address information and then its column address information
resulting in slower access to the memory as is known in the prior art.
FIG. 6 is a functional block diagram of the typical apparatus and method
employed in the prior art to access a dynamic RAM.
FIG. 7 is a drawing showing the relationship of the access addressing
strobe signals in the apparatus and method of FIG. 6.
FIG. 8 is a functional block diagram in the manner of FIG. 6 of the
apparatus and method employed in the present invention to access a dynamic
RAM.
FIG. 9 is a drawing in the manner of FIG. 7 showing the relationship of the
access addressing strobe signals in the apparatus and method of the
present invention.
FIG. 10 is a flow chart showing the basic logic accomplished by the access
optimizing logic of the present invention.
FIG. 11 is a simplified block diagram depicting a prior art approach to a
paged memory wherein a single page at a time can be accessed via a page
register indicating the base page address.
FIG. 12 is a functional block diagram of the preferred multi-bank
configuration of multiple dynamic RAMs of the present invention wherein
multiple "pages" of memory can be opened simultaneously for increased
access speed.
FIG. 13 is a flow chart showing the basic logic accomplished by the access
optimizing logic of the present invention in its preferred embodiment
employing the multi-bank configuration of multiple dynamic RAMs of FIG. 12
.
DESCRIPTION OF THE PREFERRED EMBODIMENT:
Before beginning the description of the present invention, it should be
noted that the description focuses on memory as accessed by an instruction
processor or CPU. As those skilled in the art will readily recognize and
appreciate, the methods and apparatus of the present invention can be
applied equally to any device which accesses memory and desires high
performance in conjunction therewith. Therefore, it is applicants' intent
that the teachings hereof and the claims appended hereto be given a
breadth of interpretation in keeping with the breadth of application of
the present invention despite the focus on instruction processors and
CPUs.
Broadly stated, the present invention is a mechanism, and associated method
of operation, which exploits the fast static RAM aspects of static column
mode in DRAM for application in high performance computing systems. The
use of the DRAM does not significantly slow the instruction processor. The
processor provides the memory subsystem with an address which is
partitioned into a row and column address. The row address is strobed into
the DRAM chip. Subsequent memory accesses within the same page will result
in zero-wait-state accesses. If the memory location requested does not lie
within the current page, then the row address is changed and the process
is repeated. The preferred implementation uses multiple DRAM chips
organized into "banks". For each bank there is one page that is currently
accessible. By adding additional banks, more pages are "open" to the
processor at any given time such that, statistically, the number of
zero-wait-state accesses can only increase.
Whenever the instruction processor accesses data from a page that is not
open, the full dynamic RAM cycle must be activated. This activation, of
course, will cause wait states to be inserted. The memory subsystem
according to the present invention, therefore, does not ensure
no-wait-state performance unless the entire program presently operating
fits onto the current set of open pages. It should be noted that the
static RAM approach suffers from a similar problem, however.
Zero-wait-state performance can be achieved only if the entire program
fits into the SRAM. If the SRAM is overflowed, then the program must be
operated from slower RAM, or the program must be copied into SRAM in
overlay portions. To operate out of slower RAM introduces wait states;
and, to copy data into SRAM introduces wait states. Those skilled in the
art will recognize, therefore, that the present invention operates best in
instances where, because of the programs and/or data organization, access
to open pages will occur over long sequences and the changing of pages
will occur only periodically such that the wait states which occur at the
time of page changing will impact the overall operation of the device
attached thereto minimally. When organized thus, the overall savings in
space and cost far outweigh the statistically minimum amount of wait state
time added to the operation of the accessing device as a result.
Furthermore, the present invention has an inherent advantage over the
static RAM approach in that no copying is involved when one page is closed
and another is opened. As a further consideration, statistical evidence
known in the art indicates that almost all programs exhibit substantial
"locality of reference", i.e. over some time interval some segments of
memory are accessed intensively while others are not. In particular, code
execution is usually quite sequential and stack operations are extremely
localized. References to blocks of data tend to be localized as well since
blocks of data are often relatively small. All this is to say that when
using the present invention with the "typical" program, stack, and data
regions, very little changing of pages is required such that the overall
effect of DRAM operating according to the present invention is
substantially identical to SRAM which, as will be recalled, costs much
more in both dollars and space. In this regard, the freely accessible,
multiple pages of the present invention operate substantially identically
to so-call "cache memory" and, for that reason, can be thought of as
"pseudo cache memory". By way of example, in a tested embodiment of the
present invention, the main memory is organized as multiple 1MB banks,
e.g. four banks for a 4MB computer. At any moment, each bank of memory may
have one 2KB page accessible without requiring a change of the row address
of that bank. Overall, a 4MB memory may have four 2KB pages or 8KB of
preferentially accessible, i.e. "pseudo cache", DRAM. Of course, the
particular pages which are "current" may be changed by providing new row
addresses to the banks, but this takes some time. When used with a
multiprogramming operating system, the system may preferentially locate
codes, stack, and data segments in different banks so as to increase the
probability of zero-wait-state operation.
In fairness, it should also be pointed out that while the discussion of the
present invention herein talks of keeping addresses asserted until they
need to be changed, as those skilled in the art are aware, the static
column mode of DRAM can only be maintained for a certain amount of time,
typically 100 microseconds. At that time, the address assertions must be
remade in the normal manner. At typical contemporary access times, this is
adequate for a few hundred DRAM accesses to be made before reassertion is
required. Again, the benefits attained far outweigh the limitation so that
this is not considered to be a problem.
The mechanism of the present invention in its preferred embodiment is
comprised of:
1. A table for storing the set of open pages.
2. Logic which determines whether there is a match between the row address
stored in the table and the row address emitted from the instruction
processor.
3. Logic to replace an old row address with a new row address.
When the instruction processor emits an address, the address is partitioned
into three fields: bank number, row address, and column address. The bank
number is used as an index into the table of open pages. The output of the
table is compared with the row address. If the two addresses match, then
the memory access continues in the static column mode. If the two
addresses do not match, then the replacement logic is used to overwrite
the old address in the table with the new row address. The memory access
is then continued and is guaranteed to succeed because of the replacement
step. Having thus looked at the present invention on an overview basis,
both the apparatus employed and its manner of operation will now be
addressed with even greater particularity. As depicted in FIG. 8, the CPU
14 makes its requests for data and instruction transfers from and to the
DRAM 12 through address forming logic 20 in the manner of the prior art
approach of FIG. 6. In the present invention, however, the output of the
address forming logic 20 goes to access optimizing logic 26. The access
optimizing logic 26 then makes the request to the DRAM 12 through the MUX
24 as in the prior art. It should be noted, and will be readily recognized
and appreciated by those skilled in the art, that while the access
optimizing logic 26 of the present invention is shown as a separate entity
disposed between the address forming logic 20 and the MUX 24, that logic
could be located within the address forming logic 20, in the DRAM,
anywhere inbetween, or distributed in portions therebetween, as
appropriate and convenient to the implementation. All that is required is
that the logical steps being described herein be accomplished somewhere
between the CPU 12 (or other using device) and the DRAM 12. Also, it will
be noted that the logic to be described is in the form of basic flow
diagrams which can be implemented in any number of ways well known to
those skilled in the art including hardware, software, firmware, etc. In
the interest of simplicity and to avoid redundancy, therefore, no details
of specific implementations have been included.
As depicted in FIG. 9, the strobe signals as applied in the present
invention occur in the same order as in the prior art of FIG. 7; but, in a
different manner. The row access strobe (RAS) is applied first followed by
the column access strobe (CAS) after the normal delay required by the DRAM
12 being employed. With the present invention, however, the assertion of
the RAS and CAS strobes is maintained--only the contents of that
assertion, i.e. the column address, changes. It is characteristic of the
DRAMs that when an address appears at the address input, the stored data
appears at the data output some fixed time later. Thus, when the column
address enabled by the continuously asserted CAS strobe is changed, the
new data corresponding to the new address will appear at the output at
that fixed time later. That, of course, describes a "read" cycle. The
static column mode works substantially the same way for a "write" from the
CPU to DRAM.
The basic logic of the access optimizing logic 26 of the present invention
is depicted in flowchart form in FIG. 10. As can be seen therein, to begin
a memory access sequence, the RAS and row address are output and the row
address indicated by the RAS is saved. Thereafter, CAS is output and the
next column address is multiplexed onto the DRAM address inputs. That, of
course, means that the address indicated by the continuously asserted CAS
is changed. Thus, as described above, unless there is a problem, the data
for the new column address will appear at the DRAM output a short time
later. In the meantime, however, having thus begun the access on the
assumption that, in all likelihood, the row address has remained the same
(i.e. the access is in the same page of memory), the logic compares the
row address of the current request to the row address of the previous
request as saved, i.e. the "current" row/page of the selected bank. If
they are the same, as they will be most of the time in accesses where
returns to output, i.e. change, the next CAS address. In the rare
instances where there is a page change and the row addresses do not agree,
the logic "aborts" the access and returns to the point in the logic where
the RAS is negated and reasserted with a new row address. Note that the
concept of "abort" within the logic 26 can take many forms depending upon
the memory accessing device with which the present invention and its
associated DRAM is being used. In some cases, the cycle can be easily
delayed while reasserting a new RAS and CAS. In others, the logic 26 may
have to take some affirmative action such as setting a "bad data
transfer-retrying" flag, or the like, to the requesting CPU or other
device.
Before continuing, it should be pointed out that, as depicted in FIG. 11,
it is known in the art to have a page-oriented memory 28 wherein the
memory 28 can be divided into a plurality of pages 30. This approach was
employed, for example, some twenty years ago in the Control Data model
1700 computer, which was a relative-addressed, sixteen bit word length
computer, which made it impossible to designate all address locations
within the computer's main memory 28 in a single computer word. In that
computer, the memory 28 was dividable into pages 30 of a size that could
be addressed in total with the available bits and a page register 32 was
employed to point to the currently active or "open" page. Actually, the
contents of the page register 32 pointed to a center address of an area of
the memory 28 to be considered as a page for current accessing purposes.
Subsequent accesses to the memory 28 (read, write, and next instruction
address) were to the address contained in the page register 32 plus a
relative offset in the range of plus or minus 7FFF (hexadecimal) as could
be designated in one sixteen bit computer word.
The above-described very simplified prior art concept of paging is greatly
restructured, improved and enhanced by the preferred embodiment of the
present invention as depicted in FIG. 12. As shown therein, a plurality of
DRAMs 12 are employed wherein each DRAM 12 is referred to as a "bank" of
memory. The addresses into the overall memory, generally indicated as 34,
comprising the multiple DRAMs 12, therefore, include a bank indicator
along with the row and column indicators. Additionally, the saving of the
current row address is accomplished by the access optimizing logic 26
through the use of a current row table 36 in volatile memory containing
the currently open page (i.e. row number) within each "bank" of DRAM 12.
The flowchart for the logic 26 to implement this approach is contained in
FIG. 13. As depicted therein, the logic outputs the bank identifier, the
row address and RAS and then saves the row address in the current row
table 36 indexed by the bank number. It thereafter outputs the column
address and CAS and then the next column address as in the previous
embodiment of the present invention described in detail above. Again in
similar manner, it compares the row address as output to the "current" row
as last output. In this case, however, the logic uses the bank number as
an index into the current row table 36 so as to pick up the current row
for the bank being accessed. As before, it either continues (if a match)
or aborts and reaccesses (if no match).
Thus, it can be seen that, for most applications, the present invention has
met its stated objectives by providing a method and associated apparatus
for allowing DRAM to be employed in high performance application with
little or no degradation of performance (through waiting for memory
accesses) by the associated device.
* * * * *
|
|
 |
|
 |
Description |
|
