Background Art Server devices in a network typically have more stringent memory processing characteristics than desktop personal computers. In particular, it is desireable that servers reliably store a great deal of information and quickly distribute that information in response to requests from other devices in the network. There are numerous transactions, such as for bank accounts, etc., in which it is desireable that the data be correctly stored and that an error be covered or corrected as software is running. For example, if the data is a credit card account number, corruption of the data could result in the wrong account being charged, etc. Consequently, server devices should be able to either: correct erroneous data stored in its memory and continue processing or, if the data cannot be corrected, stop the transaction and provide an error notice.

Corruption sometimes occurs in the storage medium of the memory itself. Therefore, servers typically have error correction capability support for the stored data in the memory interface. This error correction may include, for example, Single Bit Correct/Double Bit Detect ("SBCDBD") and Double Bit Correct/Triple Bit Detect ("DBCTBD"). Some server memory arrays also have a"chip kill"feature-the ability to detect the complete or substantial failure of a single memory device in the array.

Some dynamic random access memory ("DRAM") arrays in servers are specifically designed to use error correction codes ("ECC"), which are additional memory bits stored along

with the data, to detect and correct errors of the data stored in the memory. Full error correction codes employ at least one or two extra bits for each 8-bit byte of data. ECC memories in server devices storing 16-bit data frequently have 3 additional bits used for error correction.

Some memory arrays specifically designed for these servers use 72 bit data words (to provide eight additional bits for error correction) instead of the 64 bit data path width of the standard memory and memory interface used in desktop systems without any error detection circuitry. But such specialized memory arrays are less available and higher in cost than the standard desktop memory.

Furthermore, some memory devices such as RDRAMTM brand dynamic random access memory (available from Rambus, Inc., of Mountain View, Calif.) transfer data over a narrow data path having less bits than the data words transferred into and out of the processor. These narrow data path memory devices are more expensive and do not easily accommodate the full error correction desired in some environments such as in servers.

Consequently, it is desireable to devise a manner of organizing such narrow data path memory devices to accommodate full error correction.

Disclosure of the Invention The present invention is directed to a method of organizing memory devices into a memory array having error correction. In a first aspect, a memory array has Rambus Direct Random Access Memories (RDRAMTMS) coupled to respective first RDRAMTM channels, at least one of which stores and transfers a respective mutually exclusive group of the bits of a data word over its respective first RDRAMTM channel in parallel with the other first RDRAMTMS. There is also a second RDRAM coupled to a respective second RDRAMTM channel, the second RDRAMTM storing and transferring error correction data used in detecting and correcting errors in the data stored in the first RDRAMTMS.

Brief Description of the Drawings Fig. 1 is a generalized block diagram providing a basic illustration of RDRAMTM installation in a computer device.

Fig. 2 is a block diagram illustrating the memory array arrangement of a server according to a first example embodiment of the invention.

Fig. 3 is a block diagram illustrating the memory array arrangement of a server according to a second example embodiment of the invention.

Fig. 4 is a block diagram illustrating the memory array arrangement of a server according to a third example embodiment of the invention.

Best Mode for Carrying Out the Invention An example application of the invention is in the memory array of a server device storing and transferring error sensitive data. In particular, the example embodiments of the invention seek to utilize and organize widely available desktop memory devices into a memory array in such a manner so as to facilitate error correction and thereby make such devices suitable for use as the building blocks of a memory array in a server. The example embodiment of the invention is implemented with RDRAM memory from Rambus, Inc. of Mountain View, California. However, the invention may of course receive application in memory intensive devices other than servers.

Fig. 1 illustrates an example installation of RDRAMTM memories. As shown, they may be mounted on a number of Rambus In-line Memory Module ("RIMM") packaging units 101-1,101-2, etc, which are electrically coupled in daisy chain fashion to memory interface 104 via Rambus channel 103.

RIMMs 101 are substantially similar to DIMMs except, of course, that they have RDRAMTM memories 102-1,102-2, etc., rather than DRAMs, and also have different sizes. Each RIMM 102 has two connectors instead of one, so that they can be coupled in sequence in daisy chain fashion as shown in Fig.

1. There are electrical performance advantages to such an arrangement. But another advantage is that the size of the memory array can be easily changed by just adding or deleting a memory component.

While the architecture shown in Fig. 1 makes it easy to expand capacity by adding another RDRAMTM memory to the daisy chain, the Rambus channel for each daisy chain in this embodiment is restricted to an 8-bit or 16-bit data word (9 or 18 bits if a parity bit is used)-making it unsuitable for server devices which employ a larger data word to accommodate strong error correction. The parity bits in 9- bit and 18-bit RDRAMTM memories are insufficient to support error correction codes sometimes desired for servers. In Double Bit Correct/Triple Bit Detect ("DBCTBD") code, if any two parity bits fail or any two bits fail, the data is corrected before it is transferred out of the memory device.

Also, if one of the memory devices fails entirely, error correction with chip kill can detect the failure and sometimes continue running the memory device in some degraded mode. Potentially, error correction can also be performed upon the detection of three or more failed bits ("triple bit correct"). Error correction frequently involves review of successive parity bits. But of course, the error code supported may be any presently available or later developed error code.

The example embodiments arrange the basic units of RDRAM memory to allow for strong error correcting capabilities. Such ECC support for 16 bit data words employs three additional bits. Such ECC support for 64 bits employs 8 additional bits. This strong error correction and data reliability cannot be accomplished using the limited data path widths of RDRAMTM memories and Rambus channels in a conventional desktop memory configuration.

A block diagram of a first example embodiment of a memory array in a server according to the invention is shown in Fig. 2. The example array has four RDRAM memories 201-1

to 201-4 (each RDRAMTM memory 201-1 to 201-4 may be made up of several devices as shown in Fig. 1 but are referred to in the singular for convenience) on respective RIMMS 101-1 to 101-4, each RDRAMTM having a 16-bit data path and coupled to memory interface chipset 206 via respective Rambus channels 203-1 to 203-4 and channel interfaces 206-1 to 206-4. The RDRAMTM memories 201-1 to 201-4 and RIMMs 101-1 to 101-4 need not have 16-bit data paths. The data path widths could be 8 bits, for example.

RIMMs 101-1 to 101-4 all have the same capacity in this embodiment and are driven by the same clock and control signals such that RDRAMTM memory 201-1 to 201-4 each transfer a respective 16 bits of a 64 bit data word in parallel.

Collectively, RDRAMTM memories 201-1 to 201-4 transfer 64 bits of a data word in parallel.

In addition, 8 more bits are stored and transferred by RDRAMTM 201-5 on RIMM 101-5'via channel interface 206-5 in memory interface chipset 206 and Rambus channel 203- 5 for a total of 72 bits. Memory interface chipset 206 has five separate respective channel interfaces 206-1 to 206-5, one for each Rambus channel 203-1 to 203-5. Channel interface for Rambus channel 203-5 is coupled to ECC logic 205 and provides internal ECC support. As part of each Rambus channel 203-1 to 203-5, identification (ID) bits indicate the data path width and size of RDRAMTMs 201-1 to 201-5. The data path width of RDRAMTM 201-5, RIMM 101-5'and Rambus channel 203-5 is one half of the width of the data path for RDRAMTMs 201-1 to 201-4, RIMMs 101-1 to 101-4 and Rambus channels 203-1 to 203-4. Correspondingly, the capacity of RDRAMTM 201-5 is one half the capacity of each of RDRAMTMs 201-1 to 201-4. For purposes of illustrating the relationship only, RDRAMTM memories 201-1 to 201-4 are each shown as having a 2 megabyte (2M) capacity and RDRAMTM memory 201-5 is shown as having a 1 megabyte (1M) capacity. The memories may, of course, have any capacity so long as, for

this embodiment, the relationship is maintained between RDRAMTM memory 201-5 and RDRAMTM memories 201-1 to 201-4.

The 8 additional bits from RDRAMTM 201-5 are dedicated to error detection and correction. The memory interface chipset 206 decides and controls what data is read, written and corrected with the memory array as known in the art. It also generates the proper error correction code data to be stored in ECC RDRAMTM 201-5.

Each pair of ECC bits stored and transferred by ECC RDRAMTM 201-5 corresponds to sixteen of the bits in the 64-bit data word stored and transferred by RDRAMTMs 201-1 to 201-4.

However, the sixteen bits corresponding to the pair of ECC bits are not stored and transferred by any single one of RIMMs 101-1 to 101-4. Instead, the sixteen bits are spread out among RIMMs 101-1 to 101-4.

Alternatively, in a second example embodiment shown in Fig. 3, ECC RDRAMTM 301-5, RIMM 101-5 and Rambus channel 303-5 could each have a sixteen bit data path and ECC RDRAMTM 301-5 could have a capacity which is one-half that of each of RDRAMTMs 201-1 to 201-4. For example, if the capacity of each of RDRAMTMs 201-1 to 201-4 is 1 megabyte, then the capacity of RDRAMTM 301-5 is 1/2 megabyte. Such a relationship of RDRAMTM capacity could easily be accomplished by the appropriate selection of the number of memory devices on each respective RIMM as discussed above with respect to Fig. 1.

In the second example embodiment, either one of the upper eight bits or lower eight bits of the 16 bits stored and transferred in ECC RDRAMTM 301-5 is arbitrarily accessed by different addressing from memory interface chipset 206 via address signal lines 304-5. This address shifting of one bit permits a similar advantage of the first example embodiment.

However, it allows the server memory to made up entirely of RDRAMTM, RIMM and Rambus channel components of one single data path width so that a server memory may be built with only one data path width in inventory. As in the first example embodiment, in this embodiment, a server memory array with

strong error correction is organized from RDRAMTMs not intended to be able to support such strong error correction.

A third example embodiment is shown in Fig. 4. This embodiment extends the error detection and correction capability of the first two example embodiments to include DBCTBD and chip kill.

The third example embodiment differs from the first example embodiment insofar as RDRAMTM 401-5, RIMM 101-5 and Rambus channel 403-5 have a 16-bit data path width as RDRAMTMs 201-1 to 201-4, RIMMs 101-1 to 101-4 and Rambus channels 203- 1 to 203-4. RDRAMTM 401-5 has the capacity of RDRAMTMs 201-1 to 201-4. RDRAMTM 401-5 stores and transfers 16 bits of error correction data. It supports double bit correct and triple bit detect chip kill, with proper encoding, in conjunction with memory interface chipset 206.

Of course, the third example embodiment also is not limited to 16 bit data path widths and RIMMS 101-1 to 101-5 may contain any amount of memory capacity. Other error correction schemes which access memory devices and save some portion of the memories for ECC data may increase the amount of memory that is employed.

One of the advantages of the example embodiments is that one RIMM can be designated and reserved for error correction regardless of the capacity of the memory array. Another advantage is that, since all three example embodiments of the invention utilize a fifth Rambus channel, the same identical memory interface chipset 206 could be used to support ECC correction in all three of the example embodiments as well as other embodiments. In this way, either simple single bit error correction or complicated, more expensive, multi-bit correction could be supported with the same or a similar memory interface chipset and memory organization.

An advantage of the example embodiment comes from the use of RDRAMTM memory devices. While these devices have high performance, the Rambus channel interconnect technology for the devices can only have 8,9,16 or 18 bits. The example

embodiments according to the invention allow a memory array to utilize Rambus devices of 8-bit widths or 16-bit widths to obtain the high performance characteristics thereof and still support strong error correction.

Of course, the example embodiments of the invention are not limited to personal computers. Indeed, the invention is particularly useful for any computing device employing the high memory performance of Rambus and strong error correction. The invention may be used in any device in which a high degree of data storage reliability is desired.

Other features of the invention may be apparent to those skilled in the art from the detailed description of the exemplary embodiments and claims when read in connection with the accompanying drawings. While the foregoing and following written and illustrated disclosure focuses on disclosing exemplary embodiments of the invention, it should be understood that the same is by way of illustration and example only, is not to be taken by way of limitation and may be modified in learned practice of the invention. While the foregoing has described what are considered to be exemplary embodiments of the invention, it is understood that various modifications may be made therein and that the invention may be implemented in various forms and embodiments, and that it may be applied in numerous applications, only some of which have been described herein. It is intended by the following claims to claim all such modifications and variations.