Method and Apparatus for Transferring Information Between a Processor and a Memory System

The present invention relates to the field of computer systems; more particularly, the present invention relates to the field of transferring information between an agent and a memory unit.

Throughout the development of computer systems, a primary emphasis has been on increasing the speed of such systems and their ability to handle larger and more complicated programs while reducing their cost. In order to increase the ability of a computer system, it is necessary to both increase the size of the random access memory (RAM) so its larger programs may be utilized by the computer system and to increase the speed at which access to that RAM is afforded. The straight forward method of increasing access speed is to use components which operate more quickly. However, such rapidly-operating components are more expensive than slower memory components.

With the cost involved in providing high speed RAM, advanced computer systems have used high-speed cache memory arrangements to increase the operational speed of the memory system. A cache memory arrangement provides a small portion of a specially fast memory in addition to the regular RAM. As commands am issued and data is utilized, the information is called from the RAM and stored in this cache memory. As each new read and write command is issued, the system looks to the fast memory cache to determine if the information is stored in the cache. If the information is available in the cache memory, access to the RAM is not required and the command may be processed or the data accessed much more readily. If the information is not available in the cache memory, the new data can be copied from the main memory and stored in the cache memory where it remains and can be accessed for later use by the system. In well-designed memory systems, the information slot lies in the cache memory over 90% of the time, on average. Consequently, use of the cache memory substantially speeds the overall operation of the memory utilized in the computer system.

In order to further enhance the speed of operation of the computer system, it has been found desirable to directly associate a small portion of extremely rapid cache memory directly on a processor chip. For example, it may be useful to provide such a small fast cache memory consisting of 8 kilobytes of memory directly on the chip with the other elements of a CPU. Such an arrangement is capable of greatly increasing the speed of the operation of the system to a great degree for information which is used repeatedly by various processes.

Today, cache memories are commonly designed at two levels: a first level (L1) cache and a second level (L2) cache. An L1 cache is a single layer of high speed memory between a microprocessor and main system dynamic RAM (DRAM) memory. L1 caches hold copies of code and data most frequently requested by the microprocessor and are typically small ranging from 4 kilobytes to 64 kilobytes in size. The L2 cache, on the other hand, is a second layer of high speed memory between the L1 cache and the main system DRAM memory. L2 caches also hold copies of code and data frequently requested by the microprocessor. The L2 cache handles the more random memory request that the L1 cache misses. In order to simplify the handling of requests that the L1 cache misses, the L2 cache typically includes all the data of the L1 cache and more. As a result, a L2 cache is almost always larger than a L1 cache ranging in size typically from 64 kilobytes to 1024 kilobytes.

The performance of a cache is affected by the organization of the cache. Typically, them are three types of organizations that are most commonly used. These are fully associative, set associative and direct mapped (one-way set associative). In a fully associative cache memory, each item of information from the main memory system is stored as a unique cache entry. Them is no relationship between the location of the information in the data cache RAM memory and its original location in the main system memory. If there are x storage locations in the cache, the cache will remember the last x main system memory locations accessed by the microprocessor. With a fully associative cache, the location of each store can hold information from any location in the main system memory. As a result, the cache requires complex tag entries (to map the complete main memory system memory space), resulting in very complex and expensive cache comparison logic. Set associative cache organizations divide the data cache RAM into banks of memory, or "ways". A 2-way set associative cache divides the data cache RAM into two ways, a 4-way set associative cache into four ways, and so on. The set associative cache separates main system memory into pages, where each page is equal in size to the size of a way. For example, a 64 k-byte 2-way set associative cache would logically see main memory as a collection of 32 k-bytes pages, equal in size to each way. Each location in a memory page can map only to the same location in a cache way. For example, in a 2-way set associative cache memory, each location in the main system memory page can map in the same location of either of the two cache way locations in the cache. When the microprocessor makes a memory request, the set associative cache compares the memory request with the tag entry at the page location in each of its ways to determine if the information is in the cache (i.e., a hit). This means the cache has to do one comparison for each way, for a total number of comparisons equal to the number of ways. For example, in a 2-way set associative cache memory, the cache would only have to make two parallel comparisons to determine if the information requested is stored in the cache.

A direct mapped (1-way set associative cache organization) uses the entire data cache RAM as one bank of memory or way. The main system memory of the cache is logically separated into pages, where each page is the size of a data cache RAM. Each location in any main system memory page directly maps only into the same location in the data cache RAM.

In prior art, a separate cache controller is used to provide access to the L2 cache. The cache controller is separate from the processor in the computer system, usually as a separate computer chip. The cache controller is very complicated logic. Most processors systems contain two such controllers, one to control the L1 cache within the processor and the other to control the L2 cache in the system. The design of these two controllers is a compromise between performance and complexity of state that must be shared between them. The system of such hierarchical caches would provide the highest overall performance if the two cache controllers had access to information of both the caches and the processor and bus accesses. This is clearly not possible when the cache controller for the L2 cache lies in a separate package.

Another problem with the prior art is that the L2 cache is on the system bus and access to the L2 cache is limited to the speed of the system bus. For instance, if the system bus is running at 10 MHz, an access to the L2 cache can not be performed faster than 10 MHz. It would be beneficial for the processor to be able to access the L2 cache at a rate faster than that of the system bus in order to increase the overall speed of the system.

Thus, it is desirable to have a processor that may operate with multiple types of cache organizations, including the option of operating without a cache (if so desired). Therefore, as the different organizations are upgraded in the future, the microprocessor may not have to undergo any changes itself.

An additional problem with the prior art is that many processors utilize information from memory in quantities larger than the size of the bus. For example, a processor may read and write information in 16-byte quantities over a 4-byte bus. Thus, at least four clock cycles are needed to transfer the entire 16-byte quantity over the bus. With some transactions, however, a pause by the processor of four clock cycles while the 16-byte quantity is transferred unnecessarily degrades processor performance. For example, for a write request where the processor has finished using the data being written back to memory, a pause while the data is transferred over the bus prevents the processor from executing further instructions. This pause is not necessary for the processor's operation because the processor does not need the data being transferred. Thus, it would be advantageous for the processor to be able to continue executing instructions without pausing in such situations. The present invention provides for this and other advantageous results.

A method and apparatus for transferring information between a processor and a memory system are described herein. Read and write requests to the L2 cache memory are controlled by the processor. The cache line associated with each such request is larger than the interface coupling the L2 cache memory and the processor. Read requests are returned from the L2 cache memory to the processor in burst fashion. Write requests are transferred from the processor to the L2 cache memory during clock cycles in which the processor does not require the interface for a read request. Write requests need not be transferred in burst fashion; rather, a portion of the write request corresponding to the size of the interface, referred to as a chunk, is transferred from the processor to the L2 cache memory and stored temporarily in the chunk write buffer. When the processor has transferred the entire cache line to the L2 cache memory, the processor signals the L2 cache memory to transfer the contents of the chunk write buffer into the data array of the cache memory.

The present invention will be understood more fully from the detailed description given below and from the accompanying drawings of the preferred embodiments of the invention, which, however, should not be taken to limit the invention to the specific embodiments, but are for explanation and understanding only.

FIG. 1 is a block diagram of one embodiment of the cache memory system of the present invention.

FIG. 2 is a block diagram of the second level (L2) cache memory of the present invention.

FIG. 3 is a block diagram of one embodiment of the memory subsystem of a computer system implemented according to the present invention.

FIG. 4 is a diagram of one embodiment of the computer system of the present invention.

FIGS. 5A, 5B, 5C and 5D illustrate embodiments of an integrated circuit package containing multiple integrated circuit dice according to the present invention.

FIG. 6 is a flowchart showing the steps followed by one embodiment of the present invention in transferring a write request to an L2 cache memory.

A cache memory subsystem for use in a computer system is described. In the following detailed description of the present invention numerous specific details are set forth, such as specific numbers of bits, command and signal names, etc., in order to provide a thorough understanding of the present invention. However, it will be understood to one skilled in the art that the present invention may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form, rather than in detail, in order to avoid unnecessarily obscuring the present invention.

The present invention provides for single integrated circuit (IC) die control of multiple IC dice. The multiple dice may be contained within a single integrated circuit package, such that the single die controls the operations of the dice within the package, including its own control. The single die is able to provide control for other dice due to the partitioning of functionality between the dice. The control function of the dice is partitioned, such that the control necessary to monitor and start operations on other dice is placed on a single die, while the control needed to perform a specific function on a die remains. In this manner, a single die is able to know what all of the other dice are doing, while the other dice are performing the function themselves.

The present invention may provide for single die control of a two die integration. In one embodiment, the present invention provides for a second level (L2) cache memory and a processor multi-die arrangement in a computer system.

FIG. 1 illustrates the processor and L2 cache memory arrangement of the present invention. Referring to FIG. 1, processor 101 is coupled to the L2 cache memory 102 by bus 103. Bus 103 is referred to as the backside bus. Processor 101 is also coupled to a system bus (not shown) via bus 104. In one embodiment, L2 cache memory 102 may also be coupled to the system bus via its own dedicated bus. Processor 101 and L2 cache memory 102 may be contained in a dual cavity integrated circuit package. Such a package, along with the coupling of processor 101 to cache memory 102, is described later below.

In one embodiment, the processor of the present invention is an Intel Architecture Microprocessor such as manufactured by Intel Corporation of Santa Clara, Calif., the corporate assignee of the present invention. In the present invention, processor 101 may include a level one (L1) cache memory. In the present invention, processor 101 contains the functionality for controlling both the L1 and L2 cache memories.

In one embodiment, L2 cache memory 102 comprises cache management logic and a static random access memory (SRAM) that includes data and a cache directory. The data is stored in a data storage array. The cache directory may include a tag array, tag status and least recently used (LRU) bits. The cache management logic includes the logic to perform tag matching, a replacement algorithm and data routing.

In one embodiment, L2 cache memory 102 is a 4-way set associative cache memory. L2 cache memory 102 is block oriented with a 32-byte line size divided into four 8-byte "chunks". In one embodiment, L2 cache memory 102 stores 512 Kbytes of information. L2 cache memory 102 may be designed to store other amounts of information, such as 256 Kbytes of information.

The L2 cache memory 102 communicates with processor 101 using bus 103. In one embodiment, bus 103 includes an address bus for sending addresses from processor 101 to L2 cache memory 102, a bi-directional data bus, a command bus for sending commands from processor 101 to the L2 cache memory 102 and a bi-directional status and state bus. Bus 103 is internal to the package when processor 101 and cache memory 102 are contained in the same package. In one embodiment, bus 103 is a high frequency (125-225 MHz), non-multiplexed bus with a 64-bit wide data bus with error correction coding (ECC), a 33-bit address bus (supporting a 36-bit address), 24 control pins, 12 data integrity pins plus a reset signal and a clock signal. Each 8-byte quantity on the data bus is referred to herein as a "chunk".

In the present invention, L2 cache memory 102 and processor 101 may operate at different speeds. In one embodiment, L2 cache memory 102 operates at 66 MHz, while processor 101 operates at 133 MHz. In one embodiment, bus 103 provides for communication between L2 cache memory 102 and processor 101 and operates at 150 MHz.

Included along with each L2 data array 32-byte line are 32 data integrity bits. (Eight bits are stored with each set of 64 data bits). The L2 cache memory regards these extra bits as extensions of the data array line, and does not interpret, or generate these bits.

One Embodiment of the L2 Cache Memory

A block diagram of one embodiment of the L2 cache memory is shown in FIG. 2. Referring to FIG. 2, L2 cache memory 200 includes address latch 201, tag array 202, tag comparison logic 203, control logic 204, chunk write buffer 205, data array 206, multiplexer (MUX) select 207, data output register 208, and chunk MUX select logic 209. Control logic 204 is coupled to components in L2 cache memory 200 to control their operation, partially in response to control signals 212 received on its inputs and signals by L2 cache memory 200. Control logic 204 also provides signals 213 to the processor.

The address latch 201 receives an address from bus 103. Bits ›15:5! of the address select one of the sets in the tag array 202. The contents of the set are compared with bits ›35:16! of the address in the comparison logic 203 and the result is used to select the appropriate way in the multiplexer select logic 207. Bits ›4:3! of the address select which chunk of the data gets sent to the data bus 220, via the chunk multiplexer logic 209. The status signals from the tag array 202 are also included with signals 213 from the control logic 204.

In one embodiment, cache memory 200 is a 256 Kbyte four-way set associative cache, such that tag array 202 includes four separate ways (shown as way 0-3). In one implementation, tag array 202 includes 2048 sets, labeled set 0 to set 2047. In an alternate embodiment, cache memory 200 is a 512 Kbyte four-way set associative cache. In this alternate embodiment, tag array 202 includes 4096 sets, bits ›16:5! of the address in address latch 201 select one of the sets in the tag array 202, and the contents of the set are compared with bits ›35:17! of the address in the comparison logic 203.

Each set location in each way includes tag bits and status bits and additional parity bits. Two additional parity bits are included. The size of the tag is dependent on the size of the L2 cache memory. The status bits reflect the state of the L2 cache memory line: Modified, Exclusive, Share, or Invalid (MESI). These states are well-known to those skilled in the art.

It should be noted that the L2 cache memory regards the Modified/Exclusive/Shared states as valid, and does not distinguish between any of the valid states. In the present invention, the processor distinguishes between the valid states and, thus, implements the cache coherency protocol. That is, the processor of the present invention directly controls the status (MESI) of the stored data lines within the L2 cache memory, and the target associative way information on line writes. Each set of four tags corresponding to the cache lines in all four ways in the tag array contains one least recently used (LRU) field for determining victim line during a read miss. The L2 cache memory implements a pseudo-LRU algorithm for its line replacement policy. In one embodiment, tag array 202 includes an LRU array having an entry for each set in order to replace the least recently used lines in the L2 cache memory. Therefore, each set is associated with one LRU entry, four tag entries and four state entries.

The L2 cache memory supports data integrity through parity. The data integrity is intended to cover transmission errors on the backside bus and storage errors in tag array 202 and data array 206.

In one embodiment, data array 206 is a 288 Kbyte, 4-way set associative SRAM that is logically organized as a 4.times.2K.times.288. The 288 bit word length is comprised of four 72 bit wide chunks, where each chunk contains 64 data bits plus 8 ECC bits. In an alternate embodiment, data array 206 is a 576 Kbyte, 4-way set associative SRAM that is logically organized as a 4.times.4K.times.288.

The L2 cache memory allows pipelined read accesses for producing a "seamless data" stream along with mechanisms to allow data line read capability and tag array access capability. If a memory request is made to the L2 cache memory, a read operation occurs. During a read operation to the L2 cache memory, an address is driven on the address bus portion of the backside bus. The address represents a portion of the memory request. The address is latched into address latch 201. The set number portion of the address from bits ›15:5! in address latch 201 is used to select a set in the four separate ways. The tag bits for the set in each of the ways are sent to tag comparison logic 203 with the tag bit portion of the address from bits ›35:16! in address latch 201. Thus, the memory request is compared with the tag entry at the page location in each of the ways to determine if the information is in the cache. In the currently preferred embodiment, since them are four ways, them am four separate comparisons made. These comparisons occur in parallel to determine if the information requested is stored in the cache.

If one of the tags matches the tag from the address, then them is a hit. If the results of the comparison determine there is a hit, the tag comparison logic 203 outputs a signal to MUX select logic 207 causing MUX select logic 207 to select the page and corresponding cache line in data array 206. The cache line data selected is output to the data output register 208 which latches the data for output onto the data bus portion 220 of the backside bus. Using the critical chunk portion of the address from bits ›4:3! in address latch 201, the chunk MUX selection logic 209 selects the desired chunk of data for output onto data bus 220, which is output to the processor. In one embodiment, the L2 cache memory transfers a 32-byte cache line in burst fashion in four chunks, with the critical chunk first. On the other hand, if none of the tags from tag array 202 match the tag portion of the address of the memory request, then them is a miss, which is indicated to the processor using one of the signals 213.

When data is being written into the L2 cache memory due to the L1 cache of the processor writing data back to the L2 cache memory or due to information from system memory being input into the processor, a cache line is sent to the L2 cache memory. In one embodiment, the cache line is 32-bytes. The data is received by chunk write buffer 205 from data bus 220. Since data bus 220 is only 8-bytes wide and the cache line in 32-bytes wide, the data is received into the chunk write buffer one chunk at a time. The write control signals 210 cause the data from data bus 220 to be directed to its designated chunk location in chunk write buffer 205. Once chunk write buffer 205 is filled, write control signals 210 cause the data to be transferred from chunk write buffer 205 into data array 206. Thus, chunks of data can be re-ordered when written into the chunk write buffer and therefore need not be received by the L2 cache memory in the same order as they occur in the cache line.

FIG. 3 is a block diagram of a memory subsystem of the computer system of the present invention. Boundary lines are shown to define the portion of the memory subsystem that is contained within the processor and that portion which is external to the processor. Referring to FIG. 3, an L2 cache memory 301 is coupled to a backside bus 302. Backside bus 302 is also coupled to backside bus logic (BBL) 303. Backside bus logic 303 is also coupled to external bus logic 304, instruction fetch unit (IFU) 305 and the data cache (DCU) 306. The external bus logic 304 is also coupled to the instruction fetch unit 305, data cache 306 and the system bus 307. Also shown coupled to system bus 307 is system memory 308.

Instruction fetch unit (IFU) 305 includes instruction fetch logic and the instruction cache and fetches instructions for execution in the processor. When the instruction cache of IFU 305 desires to fetch more instructions, it sends a request on signal 310 to external bus logic 304. The external bus logic 304 sends the request on request lines 315 to backside bus logic 303 which forwards the request to the L2 cache memory 301. In response, the L2 cache memory 301 performs a read operation. If there is a hit in L2 cache memory 301, the instructions are returned to instruction fetch unit 305 on the data return signal path 311 from backside bus logic 303 and backside bus 302.

The DCU 306 holds temporary copies of data for use by the processor in executing the instructions. In one embodiment, DCU 306 comprises an L1 cache memory. In executing instructions, if the data is not in the DCU 306 (that is, the L1 cache), a request is made on request signal path 312 to external bus logic 304. External bus logic 304 sends a request on request lines 315 to backside bus logic 303 which forwards the request onto L2 cache memory 301. If the data is in L2 cache memory 301 (if L2 cache memory 301 hits), then the data is forwarded and returned to DCU 306 on data return signal path 313. When data is written back from DCU 306 or if there is no longer room in DCU 306, data is sent out to backside bus 302 via backside bus logic 303 to L2 cache memory 301.

In one embodiment of the present invention, data return signal path 311 and data return signal path 313 comprise a single signal path. In this embodiment, backside bus logic 303 utilizes the same signal path to return data to both the IFU 305 and the DCU 306, thereby reducing the number of signal paths necessary within the processor chip.

Any time data from L2 cache memory 301 is to be written back to system memory 308, a request is made to external bus logic 304 via request lines 316 from backside bus logic 303. The data is written from the backside bus 302 through backside bus logic 303 to external bus logic 304 via data path 327. External bus logic 304 controls the data by writing the data on external bus 317 to system memory 308 via system bus 307. Any time them is a miss to L2 cache memory 301, external bus logic 304 sends the request to system memory 308 using external bus 317 and system bus 307. The data returned is received by external bus logic 304 and sent to backside bus logic 303 via data path 327 for storage in L2 cache memory 301. The data is then written to DCU 306 via data return lines 313 and to L2 cache memory 301 via backside bus 302.

Processor accesses from the instruction fetch unit 305 are always fetch operations with no intent to modify the data. Similarly read operations from the data cache 306 that are generated by the processor executing a read operation are always data fetch operations. These requests are forwarded to the system bus 307, if they miss the L2 cache memory 301, as read operations from the memory. When the data is returned by the system memory 308, a signal on the system bus 307 indicates if copies of this data also exist in other cache memories in a multiple processor system. If they do, then the data is placed in the L2 cache memory 301 and data cache memory 306 and marked with a state of Shared (S). On the other hand, if no other cache memory has a copy of the data as indicated on the system bus 307, the data can be placed in the L2 cache memory 301 and the data cache memory 306 and marked Exclusive (E). In one embodiment of the present invention, instructions which are returned by the system memory 308 are placed in the L2 cache memory 301 and the instruction cache of IFU 305 and marked with a state of Shared (S).

Whenever the processor intends to modify data, it issues a request to the data cache memory 306 to obtain a copy of the data with complete ownership. If the data is not present in the data cache memory 306, a similar request is sent to the L2 cache memory 301. If the request also misses the L2 cache memory 301, it is then forwarded to the system bus 307 as a data read request with a command modifier indicating to all other processors in the system bus 307 that they must relinquish ownership of this cache line of data and invalidate all copies of the line in their cache memories. The data is then returned to the requesting processor granting it exclusive ownership and is placed in the L2 cache memory 301 and data cache memory 306 in the Modified (M) state. The processor is then free to update any portion of the cache line it chooses. This results in a write command being issued to the data cache memory 306 with the modified data and the data cache memory 306 updates its copy of the cache line to reflect the change. If the processor chooses to modify the data in that cache line yet again, it can do so within the data cache memory 306 without issuing further request to the L2 cache memory 301 and the system bus 307.

In one embodiment of the present invention, the L2 cache memory 301 and the data cache 306 and IFU cache 305 memories do not rely on any special inclusion properties. That is, a cache line may be valid in the L2 cache memory 301 and may not be in either the data cache memory 306 or the instruction cache of the IFU 305. In an alternate embodiment, the processor may follow a protocol by which the state of a cache line in the data cache memory 306 is always lower than or equal to the state of the same cache line in the L2 cache memory 301, wherein, for example, the hierarchy for a MESI based cache coherency is Modified, Exclusive, Shared, and then Invalid (listed from highest priority to lowest). This is referred to herein as a "dirty" or "modified" inclusion. Implementing such a priority, the processor would be able to determine that if the cache line in the data cache memory 306 is shown to be Modified, then the processor knows that the state of the same cache line in the L2 cache memory 301 was changed to Modified upon the cache in state in the data cache memory 306 and that the L2 cache memory 301 still reflects such a state for the cache line. In another embodiment, if a subsequent change to the same cache line is to occur, the processor could request the state of the cache line as currently indicated by the L2 cache memory 301. Upon seeing that the state is Modified, the processor knows that the data in the data cache memory 306 may be changed. Another mechanism by which the processor is able to identify whether a change in state needs to be made to a cache line in the L2 cache memory 301 upon the occurrence of a subsequent write to the same cache line in the data cache memory 306 is to include a "more modified" indication (e.g., a bit location) for the cache line in the data cache memory 306 that indicates when the data in the data cache memory 306 has been subsequently modified. In such an implementation, when such "more modified" data is to be written back to the external bus (main memory), then the processor is able to determine which data should be written, specifically whether the data in the L2 cache memory 301 suffices or only the data in the data cache memory 306 will do.

Several variations of this scenario can exist regarding modifications to data in the cache memories. For example, the data the processor intends to modify could be in the data cache memory 306 or the L2 cache memory 301, or both, because of a previous data fetch transaction. However, the data is marked Shared (S) indicating that although the cache memory has the latest copy of the data, other cache memories in the system also have copies of it. Under such a circumstance, the data cache memory 306 or the L2 cache memory 301 must make a request on the system bus 307 requesting ownership of the line but not necessarily perform a data fetch operation. This request causes all other cache memories in the system to mark their copies of the cache line as Invalid (I), granting exclusive ownership of the cache line to the requesting processor. This operation needs to be performed in order to ensure that no more than one processor at a time attempts to modify the data in a cache line at a time.

If two processors were to attempt to do so, then the request for ownership for the cache line from one of them will reach the system bus 307 before the other and it will get the opportunity to makes its modifications to the cached data first. The integration of the control for the L2 cache memory 301 into the processor permits considerable optimization of the control logic of these complex operations. Moreover, the interaction between sequential transactions targeted at the same cache line can now be resolved optimally without holding them up and waiting for the entire previous transaction to complete. To this end, multiple operations may be merged into single transaction by the processor. This has a dramatic effect upon the performance of multiple processor systems that are sharing the same data out of the system memory 308.

By integrating the control for the L2 cache memory 301 into the processor, the present invention is able to better coordinate the activities and transactions that occur in the computer system. The single control of the processor is functionally positioned with respect to the data cache memory 306, the L2 cache memory 301 and the external bus to allow the controller to obtain information and respond, if necessary, to any transaction on the three ports (data cache memory 306, the L2 cache memory 301 and the external bus) that is currently in progress. The controller of the present invention can merge transactions sourced from one or more of the data cache memory 306, the L2 cache memory 301 and the external bus into a single transaction directed to the processor core or a port.

For example, if the processor core desires to modify a cache line in the data cache memory 306 which is marked as Exclusive (E) in the data cache memory 306 and the L2 cache memory 301, when the data is modified in the data cache memory 306, the L2 cache memory 301 must be set to Modified (M) (if following the protocol discussed above). Then, if there is no Modified state written into the L2 cache memory 301, a later modification to different data in the same cache line occurs, and the cache line is to be evicted from the data cache memory 306, then the state must be changed to Modified (M) in the L2 cache memory 301 and the L2 cache memory 301 must be updated with the modified data. The controller of the present invention is able to merge the operations into a single operation. In a similar fashion, the controller of the present invention is able to recognize conflicts between transactions with the processor core and the L2 cache memory 301 and merge the transactions, such that multiple transactions, as a whole, are optimized into lesser operations, when possible.

Information is transferred from backside bus logic 303 to L2 cache memory 301 in chunks. The size of each chunk is the size of the data bus of backside bus 302. In one embodiment of the present invention, a cache line is 32 bytes and the width of the data bus of backside bus 302 is 8 bytes. Thus, transferring information from backside bus logic 303 to L2 cache memory 301 comprises transferring four chunks over backside bus 302.

Backside bus logic 303 operates to optimize the bandwidth of backside bus 302. By optimizing the bandwidth of backside bus 302, backside bus logic 303 reduces IFU 305 idle time because IFU 305 is not required to wait for instructions to be returned from L2 cache memory 301.

Backside bus logic 303 optimizes the bandwidth of backside bus 302 by giving read requests priority over write requests. A read request may be received from either IFU 305 requesting an instruction, or DCU 306 requesting data. A write request is received from DCU 306 either requesting to write data back to system memory or requesting to write data to L2 cache memory 301.

Read requests are immediately forwarded to L2 cache memory 301. If there is a hit in L2 cache memory 301, the information, either instructions or data, is returned to backside bus logic 303 via backside bus 302. In one embodiment of the present invention, this information is burst over backside bus 302. Thus, the entire cache line is transferred over backside bus 302 in four consecutive clock cycles. Backside bus logic 303 then returns the information to the requesting agent, IFU 305 or DCU 306.

Backside bus logic 303 does not, however, immediately burst a cache line for a write request over backside bus 302 to L2 cache memory 301. Rather, a write request to L2 cache memory 301 is buffered in backside bus logic 303 and is transferred to L2 cache memory 301 in chunks when backside bus 302 is available. Chunks received by L2 cache memory 301 are stored in the chunk write buffer. Backside bus 302 is available when instructions or data from a read request are not being returned to backside bus logic 303 over backside bus 302 from L2 cache memory 301.

Backside bus 302 is frequently available for a period of clock cycles subsequent to a read request. This is because although IFU 305 retrieves instructions from L2 cache memory 301 in 32-byte lines, IFU 305 typically cannot execute those instructions in a single clock cycle, Thus, one or two clock cycles are available immediately subsequent to the read request in which IFU 305 is not performing another read request, thereby leaving backside bus 302 unused. These one or two clock cycles are long enough to transfer one or two chunks to L2 cache memory 301, however this period is typically not of sufficient duration to allow all four chunks of a write request to be transferred to L2 cache memory 301. It should be noted that under certain circumstances, four or more consecutive clock cycles may exist in which IFU 305 is not performing a request which requires the use of backside bus 302. In these situations, backside bus logic 303 transfers all four chunks to L2 cache memory 301 in four consecutive clock cycles.

In addition, many of the write requests serviced by backside bus logic 303 are a result of read requests which missed L2 cache memory 301. For these requests, backside bus logic 303 waits for external bus logic 304 to fetch the instructions from system memory 308 via processor system bus 307. Processor system bus 307, however, typically operates at a speed which is a fraction of that which IFU 305, backside bus logic 303, and backside bus 302 all operate. Thus, if backside bus logic 303 were to stall IFU 305 while waiting for the entire 32-byte cache line to be fetched from system memory 308, the operation of IFU 305 would be significantly delayed. For example, if processor system bus 307 is operating at one-fourth the speed of IFU 305, then the fetch operation to system memory 308 would result in a delay of 16 clock cycles for IFU 305 because external bus logic 304 would be retrieving a single 8-byte chunk from system memory 308 every four clock cycles.

Backside bus logic 303 prevents such a long delay by writing single chunks to L2 cache memory 301 as the chunks and backside bus 302 become available. Thus, IFU 305 can continue to issue requests to L2 cache memory 301 via backside bus logic 303 and backside bus 302 because backside bus logic 303 is not waiting for an entire cache line. Thus, in the example above, backside bus logic 303 transfers a single chunk to L2 cache memory 301 when it receives the chunk from external bus logic 304, assuming backside bus 302 is available, thereby utilizing backside bus 302 once every four clock cycles rather than utilizing it for the entire 16 clock cycle period.

In one embodiment of the present invention, numerous micro-operations may be utilized to control the L2 cache memory behavior. In one implementation, these micro-operations include control beyond reading and writing an entire cache line. In this implementation, backside bus logic 303 gives priority to any read operations from IFU 305, where these read operations involve the transferring of data or commands from L2 cache memory 301 to backside bus logic 303. Thus, any write operations (that is, those operations requiring data or commands to be transferred to L2 cache memory 301 from backside bus logic 303) are not performed until backside bus 302 is available.

Thus, backside bus logic 303 services read requests immediately and services write requests when backside bus 302 is available. Therefore, IFU 305 idle time is reduced because requests whose instructions or data are required for the operation of IFU 305 are serviced immediately, while requests whose data has already been utilized by IFU 305 are not necessarily serviced immediately.

Controlling the L2 Cache Memory

The processor of the present invention controls the functionality of the L2 cache memory, such that the L2 cache memory is a slave on the backside bus. That is, the control of the L2 cache memory is partitioned between the processor and the L2 cache memory in such a way that the processor of the present invention controls and knows what the L2 cache memory is doing (e.g., whether the L2 cache memory is transferring data). In the present invention, processor 101 controls the L2 cache memory using opcodes sent by the processor to the L2 cache memory.

The processor controls the L2 cache memory behavior through micro-operations which provide the functionality needed for all processor micro-operations requiring the L2 cache memory. In one embodiment, these micro-operations are used to perform all write operations to the L2 cache memory. Also, the processor uses these micro-operations to maintain the cache consistency. These micro-operations function within the confines of the interface provided between the processor and the L2 cache memory in order to implement single die control in a multi-die package. A brief description of each micro-operation performed by the L2 cache memory is given below.

L2 Read Inquire: The processor uses this micro-operation to read an entire line from L2 cache memory if the line is resident in the L2 cache memory. The processor sends a complete address to the L2 cache memory. The L2 cache memory examines the specified set number to see if it has the specified tag. The L2 cache memory signals a hit to the processor if a tag match is found and the status of the line is modified, exclusive or shared. On a hit, the L2 cache memory updates the appropriate LRU field and returns the way information, the value of the status, and the data line in a burst fashion over four cycles to the processor. If the tag did not match, the L2 cache memory returns the way information, status and full line address (tag and set) of the cache line to be replaced. If a miss occurred, the LRU field is not updated, but will be updated when the replacement line is written using the Write Line Update micro-operation.

State Write: The processor may change the status of any given line in the L2 cache memory with this micro-operation. The processor indicates the set address, the way, and the status value to be stored. No tag matching is done on the address and the L2 cache memory LRU field is not modified.

State Read: The processor reads the status of any given line in the L2 cache memory with this micro-operation. The processor indicates the set address and the way to be read. The L2 cache memory returns the status of the specified line. No tag matching is done on the address and the L2 cache memory LRU field is not modified. This operation may be used for testing purposes.

Tag Inquire: This operation prompts the L2 cache memory to determine if it contains a given line. The L2 cache memory examines the specified set number to determine if it has the specified tag. The L2 cache memory then signals a hit to the processor if a tag match is found and the status of the line is something other than invalid (an illegal status encoding will all signal a hit, but will also cause a BERR). The L2 cache memory data array is not accessed and the L2 cache memory LRU field is not modified. This operation may be used in response to a request from a device on the external bus made to determine if data that is desired is currently stored in the cache memory.

Tag Inquire Update: This micro-operation is identical to Tag Inquire, except that the LRU field is updated upon a match.

Tag Array Read: This micro-operation is used to read an entry in the tag array without accessing the data array. The processor provides the set address and the way information. The L2 cache memory returns the tag status and the full address. The address is returned over the data lines using the natural alignment of the tag. This scheme avoids bi-directionality on the address lines. The L2 cache memory LRU field is not modified.

Data Array Read: This micro-operation is used to read an entry in the data array without accessing the tag array. The processor provides the set address and the way information. The data line is sent back in a burst fashion over a four cycle burst. The L2 cache memory LRU field is not modified. This operation may be used to test the array.

Write Chunk: The processor uses this micro-operation to load a data chunk into the Chunk Write Buffer on the L2 cache memory chip by specifying the chunk number, driving a data word onto the data bus, and signaling the L2 cache memory to write the chunk. This micro-operation can occur independently of any other L2 cache memory operation not involving the data bus. It is the responsibility of the processor to prevent data bus collisions when issuing Write Chunk micro-operations. No data array or tag array accesses are made during this operation and the LRU field is not modified. This operation is used to write a single chunk into the L2 Chunk Write Buffer, which, at a later time, transfers the contents of the Chunk Write Buffer to the L2 cache memory.

Write Line: This micro-operation is used as the final step in writing an entire line into the L2 cache memory. The line to be written is first stored in the L2 cache memory Chunk Write Buffer using the Write Chunk micro-operation described above. The processor then uses the Write Line micro-operation along with specifying the tag and set address, the way information, and the status of the line to be written. The entire line contained in the Chunk Write Buffer is then written into the L2 cache memory data array. The tag and the status are written into the tag array and status array, respectively. The Write Line micro-operation may be issued concurrently with the last of the four Write Chunk commands. The LRU field is not modified with this micro-operation.

Write Line Update: This micro-operation is identical to the Write Line with the addition that the LRU field is updated.

The processor sends these micro-operations to the control logic of the L2 cache memory which decodes the micro-operations to control the sequence of outputs and to provide data integrity. The decoding performed produces signals that initiate access cycles. The control block of the L2 cache memory also performs output control. The control logic also includes a logic for performing data integrity operations.

The micro-operations operate on the L2 cache memory using the interface provided by the backside bus. One embodiment of the backside bus provides the following L2 cache memory external interface:

Through the use of the above signals, the present invention allows access to the L2 cache memory by the processor. However, the L2 cache memory is able to complete the access in any manner. The processor does not need to know how the access is being completed. Because of this partitioning, any L2 cache memory may be designed for use in the computer system of the present invention. Other signals for bond control and testing, as well as for testing the processor, have been omitted to avoid obscuring the present invention.

Note that the processor die includes the control logic for generating the micro-operations and using the interface provided between the L2 cache memory die and the processor die, the processor provides a unilateral control path between the two dice, with only the bi-directional data bus as an exception. In this manner, the partitioning between the two dice allows one of the dice to include the necessary functionality to control the other die(s).

The sequence of operations that are sent to the L2 cache are controlled by the processor. This control logic can make optimal use of the L2 as it has access to the L1 cache, the processor requests and the system bus requests. The logic is designed to provide the highest priority to all processor accesses. These accesses use the address bus and the arrays for a short period of time but then use the data bus to read out the data chunks over four cycles. The control logic can recognize this condition and perform tag inquire instructions to operate in parallel and not require the data bus.

In the present invention, the processor performs an L1 cache memory look-up before sending a memory request out to the L2 cache memory. In another embodiment, the L1 and L2 cache memory look-ups are sent at the same time. In this case, the information may be obtained one or two cycles earlier than the prior art.

By having a single point of control in the processor, the complexity of the memory subsystem is reduced by having to use only a simple interface between the processor and the L2 cache memory. This interface only includes an address bus, a data bus and some status signals sent from the processor to the L2 cache memory for providing information to the L2 cache memory in one direction. Only the data bus is bi-directional.

Due to the partitioning in the present invention, multiple types of cache memory organizations may be utilized with the present invention. The cache memory must only be able to satisfy the memory request from the processor according to its organization. This allows the L2 cache memory to be upgraded to a different organization in the future without impacting the particular site in the circuit board. The only requirement is that the signals from the IC package remain the same. In the present invention, the directory and data of the L2 cache memory operate independently. Because of this, it can be upgraded independently. It is possible for the data to be transferring data over on the data bus while the control is receiving a command to look up the directory to perform a tag match.

In order to accommodate multiple requests in a single cycle, a duplicate of the directory could be utilized. In this case, the processor can examine the directory as well as another item on the bus. In another embodiment, a handshake signal may be used to indicate that the cache memory is not ready, and at a proper time, a signal could be used to indicate that the cache memory is ready.

By interleaving a data transfer cycle with a directory look-up, the data transfer and the look-up can be totally hidden under the command of the data transfer such that it can be done inexpensively.

The processor and the L2 cache memory in the present invention may be integrated in a computer system such as that in FIG. 4. Referring first to FIG. 4, an overview of a computer system of the present invention is shown in block diagram form. It will be understood that while FIG. 4 is useful for providing an overall description of the computer system of the present invention, a number of details of the system are not shown. As necessary for disclosure of the present invention, further detail is set forth with reference to the other figures provided with this specification. Further, the present invention is described with reference to its preferred embodiment.

As illustrated in FIG. 4, a computer system, as may be utilized by the present invention, generally comprises a processor-system bus or other communication means 401 for communicating information and a processor 402 coupled with processor-system bus 401 for processing information. In the present invention, processor-system bus 401 includes address, data and control buses. In the currently preferred embodiment, processor 402 includes an internal cache memory, commonly referred to as a level one (L1) cache memory for temporarily storing data and instructions on-chip. A level two (L2) cache memory 404 is coupled to processor 402 for temporarily storing data and instructions for use by processor 402. In the currently preferred embodiment, cache memory 404 is included in the same chip package as processor 402.

Also coupled to processor-system bus 401 is processor 403 for processing information in conjunction with processor 402. Processor 403 may comprise a parallel processor, such as a processor similar to or the same as processor 402, or may comprise a co-processor, such as a digital signal processor. A level three (L3) cache memory 411 for temporarily storing data and instructions for use by other devices in the computer system (e.g., processor 402, processor 403, etc.) and a L3 cache controller 410 for controlling access to L3 cache memory 411 may also be coupled to processor-system bus 401. The L3 cache controller 410 is also coupled to memory-system bus 415.

A memory-system bus or other communication means 415 for communicating information is coupled to processor 402 for providing processor 402 and other devices in the computer system access to the memory and input/output (I/O) subsystems. A memory controller 422 is coupled with memory-system bus 415 for controlling access to a random access memory (RAM) or other dynamic storage device 421 (commonly referred to as a main memory) for storing information and instructions for processor 402 (and processor 403). A mass data storage device 425, such as a magnetic disk and disk drive, for storing information and instructions, and a display device 423, such as a cathode ray tube (CRT), liquid crystal display (LCD), etc., for displaying information to the computer user are coupled to memory-system bus 415.

An input/output (I/O) bridge 424 is coupled to memory-system bus 415 and I/O bus 431 to provide a communication path or gateway for devices on either memory-system bus 415 or I/O bus 431 to access or transfer data between devices on the other bus. Specifically, bridge 424 turns the byte/word/dword data transfer traffic from I/O bus 431 into line size traffic on memory-system bus 415.

I/O bus 431 communicates information between devices in the computer system. Devices that may be coupled to system bus 431 include a display device 432, such as a cathode ray tube, liquid crystal display, etc., an alphanumeric input device 433 including alphanumeric and other keys, etc., for communicating information and command selections to other devices in the computer system (e.g., processor 402) and a cursor control device 434 for controlling cursor movement. Moreover, a hard copy device 435, such as a plotter or printer, for providing a visual representation of the computer images and a mass storage device 436, such as a magnetic disk and disk drive, for storing information and instructions may also be coupled to system bus 431.

Of course, certain implementations and uses of the present invention may not require nor include all of the above components. For example, in certain implementations, the L3 cache controller and L3 cache memory may not be required. In such implementations processors (402) and (403) will reside directly on a memory system bus 415. In other implementations, it may not be required to provide a display device for displaying information. Certain implementations of the present invention may include other components.

In one embodiment, the L2 cache memory and the processor of the present invention are contained within the same integrated circuit package. FIG. 5A illustrates a section view of one embodiment of the dual die package of the present invention. Referring to FIG. 5A, package 500 includes chip carrier 501 for encapsulating and protecting multiple integrated circuit (IC) dice, leads or pins 502 for electrically coupling dice in package 500 to the circuit board (not shown), wirebond 505, processor die 503 and L2 cache memory die 504. The coupling between the circuit board and package 500 has not been shown to avoid obscuring the present invention. Die 503 is also coupled to leads 502, such that electrical signals may be transferred from die 503 (and die 504) to off-chip (i.e., to the printed wiring board). In one embodiment, die 503 is wire bonded to leads 502 of package 500 in a manner well-known to those skilled in the art. Although only six leads are shown, the actual number of leads for the package is generally much larger, numbering in the hundreds.

The leads 502 and the chip carrier 501 are comprised of materials standard in the art. For instance, leads 502 may comprise copper leads. Similarly, chip carrier 501 may be a ceramic carrier, a plastic carrier or other material suitable for encasing the multiple dice of package 500.

In one embodiment, processor die 503 and L2 cache memory die 504 are integrated circuits fabricated using well-known metal-oxide semiconductor (MOS) techniques. Although only die 503 and die 504 are shown, other embodiments may include more than two dice integrated into the same package. Processor die 503 and cache memory die 504 are electrically coupled to each other by wirebond 505. Moreover, die 503 and die 504 are coupled to chip carrier 501 using, for example, adhesive, such a epoxy.

Die 503 and die 504 may be electrically coupled to each other in numerous other ways. In one embodiment, L2 cache memory die 504 is coupled directly on top of die 503 using solder bumps (e.g., solder bumps 506) in a flip-chip type configuration, such as shown in FIG. 5B. Processor die 503 and L2 cache memory die 504 may be coupled to a trace 507 using wirebonds 509, such as shown in FIG. 5C. In this case, trace 507 acts as part of the electrical connection between the dice. In yet another embodiment, die 503 and die 504 may be solder bumped to a common silicon substrate 511 using solder bumps 510 as shown in FIG. 5D. In this case, silicon substrate 511 provides a communication path between the dice.

FIG. 6 is a flowchart showing the steps followed by one embodiment of the present invention in transferring a write request to an L2 cache memory. The write request is first received and buffered in the backside bus logic, step 610. The backside bus logic then checks whether a read request is currently utilizing the backside bus, step 620. The read request could be utilizing the backside bus by, for example, backside bus logic transferring the request to the L2 cache memory or by the L2 cache memory transferring the data corresponding to the read request to the backside bus logic. If a read request is currently utilizing the backside bus, then the backside bus logic allows the read request to continue, step 630.

However, if a read request is not utilizing the backside bus, then a chunk of the data for the write request is transferred to the chunk write buffer in the L2 cache memory, step 640. The L2 cache memory then checks whether additional chunks remain to be transferred, step 650. In one embodiment, the L2 cache memory makes this determination based on receipt of a signal from the backside bus logic which indicates all chunks of the write request have been transferred. If no additional chunks remain to be transferred, then the L2 cache memory stores the contents of the chunk write buffer into the appropriate L2 cache line, step 660. Otherwise, the backside bus logic continues to transfer additional chunks of the write request to the chunk write buffer until the entire write request is transferred (steps 620 through 650).

It should be noted that additional requests may be received by the backside bus logic from the instruction fetch unit or data cache unit between chunk transfers. Thus, subsequent read requests may prevent transfer of the current write request to the L2 cache memory, and subsequent write requests may be buffered in the backside bus logic until transfer of the current write request is complete.

The present invention provides for a more reliable connection because the distance the signals travel between the two dice is reduced. This distance is substantially shorter than the distance of traveling over the printed wiring board and the pins. Moreover, the fact that both dice are contained in one package prevents signals from having to cross the packaging material (e.g., the ceramic) travel through the printing circuit board and then continue into another chip. Therefore, the present invention provides excellent electrical characteristics for high frequencies.

The present invention also reduces the total number of pins required for the microprocessor and the cache memory. In the prior art, a certain number of pins are used to transfer information between the cache memory and the microprocessor. This transfer occurs over the printed wiring board. However, because the cache memory is contained in the same package, the pins used to transfer data between the cache memory and the microprocessor. Thus, the overall number of pins is reduced.

The present invention allows multiple dice developed using different technologies to be integrated into the same package. Each of the dice may be developed and optimized independently. In other words, a processor may be designed and manufactured based on technology that optimizes the processor while a cache memory may be designed and manufactured based on technology optimized for the cache memory (e.g., SRAM technology). Therefore, even though the two have different design complexities both may be integrated into the same package.

The present invention also provides a single chip replacement upgrade capability when a new processor and cache memory pair is to be placed in the system. In this case, all that is required to obtain an upgrade is the extraction of one chip from the printed wiring board. Moreover, the cache memory can be upgraded independently because of the partitioning. Normally, the printed wiring board is designed to include a cache memory having a predetermined number of pins and particular electrical characteristics. Therefore, to change the cache memory requires a change in the printed wiring board. Thus, using the present invention, the cache memory may be upgraded without having to upgrade the entire printed wiring board.

Furthermore, the present invention provides reduced electromagnetic interference (EMI). The package, in effect, shields radiation between the two dice due to the shorter connection. Thus, the physical proximity of the two multiple dice together provides for a much more uniform terminal environment. This environment can also reduce the effects of drifts because of the characteristic process variations due to temperature are lowered because both dice are at approximately the same temperature.

Whereas many alterations and modifications of the present invention will no doubt become apparent to a person of ordinary skill in the art after having read the foregoing description, it is to be understood that the particular embodiment shown and described by way of illustration is in no way intended to be considered limiting. Therefore, references to details of a preferred embodiment are not intended to limit the scope of the claims which in themselves recite only those features regarded as essential to the invention.

Thus, a method and apparatus for transferring information between a processor and a memory system has been described.