DESCRIPTION OF SPECIFIC EMBODIMENTS

[0019] One or more specific embodiments of the present invention will be described below. In an effort to provide a concise description of these embodiments, not all features of an actual implementation are described in the specification. It should be appreciated that in the development of any such actual implementation, as in any engineering or design project, numerous implementation-specific decisions must be made to achieve the developers' specific goals, such as compliance with system-related and business-related constraints, which may vary from one implementation to another. Moreover, it should be appreciated that such a development effort might be complex and time consuming, but would nevertheless be a routine undertaking of design, fabrication, and manufacture for those of ordinary skill having the benefit of this disclosure.

[0020] Turning now to the drawings and referring initially to FIG. 1, a block diagram of an exemplary computer system is illustrated and generally designated as reference numeral 10 . The computer system 10 typically includes one or more processors or CPUs 12 A- 12 H. In the exemplary embodiment, the system 10 utilizes eight CPUs 12 A- 12 H. The system 10 utilizes a split bus configuration in which the processors 12 A- 12 D are coupled to a first bus 14 A, whereas the processors 12 E- 12 H are coupled to a second bus 14 B. It should be understood that the processors or CPUs 12 A- 12 H may be of any suitable type, such as a microprocessor available from Intel, AMD, or Motorola, for example. Furthermore, any suitable bus arrangement may be coupled to the CPUs 12 A- 12 H, such as a single bus, a split bus (as illustrated), or individual buses. By way of example, the exemplary system 10 may utilize Intel Pentium III processors and the buses 14 A and 14 B may operate at 100/133 MHz.

[0021] Each of the buses 14 A and 14 B is coupled to a chip set which includes a host controller 16 and a data controller 18 . In this embodiment, the data controller 18 is effectively a data cross bar slave device controlled by the host controller 16 . Therefore, these chips will be referred to together as the host/data controller 16 , 18 . The host/data controller 16 , 18 is further coupled to one or more memory controllers. In this particular example, the host/data controller 16 , 18 is coupled to five memory controllers 20 A- 20 E via five individual bus segments 22 A- 22 E, respectively. These individual bus segments 22 A- 22 E (also referred to herein as MNET) may facilitate the removal of individual memory modules. Each of the memory controllers 20 A- 20 E is further coupled to a segment of main memory designated as 24 A- 24 E, respectively. As discussed in detail below, each of the memory segments or modules 24 A- 24 E is typically comprised of dual inline memory modules (DIMMs).

[0022] As will be appreciated from the discussion herein, the number of memory segments 24 may vary depending upon the type of memory system desired. In general, redundant memory systems will utilize two or more memory segments 24 . Although the five memory segments 24 A- 24 E illustrated in the exemplary embodiment facilitate a “4+1” striping pattern of data and parity information as discussed in detail below, a memory system having two memory segments 24 may be used in which data is mirrored on each segment to provide redundancy. Similarly, a memory system having three or more memory segments may be used to provide various combinations of data and parity striping.

[0023] Each of the memory controllers 20 A- 20 E and its associated main memory segment 24 A- 24 E forms a portion of the main memory array 26 . The five memory controllers 20 A- 20 E operate in lock-step. In this example, each cacheline of data is striped, and each of the memory controllers 20 A- 20 E handle a separate quad-word of each cacheline of data (assuming a 32 byte cacheline) that is being transferred to or from the host and data controllers 16 and 18 . For example, the memory controller 20 A handles the first quad-word of every data read and write transaction, the memory controller 20 B handles the second quad-word, the memory controller 20 C handles the third quad-word, and the memory controller 20 D handles the fourth quad-word. Instead of receiving one of the four quad-words, the memory controller 20 E handles data parity for the four quad-words handled by the memory controllers 20 A- 20 D. Thus, as described below, the memory array 26 forms a “redundant array of inexpensive DIMMs” (RAID) memory structure. The functionality of the exemplary 4+1 striping scheme will be discussed more specifically with reference to FIGS. 2 - 4 , below.

[0024] As will be explained in greater detail below, during a data read operation, the host/data controller 16 , 18 receives four quad-words of data plus parity from the five memory controllers 20 A- 20 E, validates data integrity of each quad-word and parity using ECC codes, and, if necessary, corrects bad data using an exclusive OR (XOR) engine before forwarding the data to its destination. During a data write operation, the host/data controller 16 , 18 uses the XOR engine to calculate data parity and transfers the four quad-words of data and parity to the five respective memory controllers 20 A- 20 E. In this embodiment, all data transfers between the host/data controller 16 , 18 and the five memory controllers 20 A- 20 E are an entire cacheline, and partial writes are translated into read-modify-write operations.

[0025] Each memory controller 20 A- 20 E along with its associated memory segment 24 A- 24 E may be arranged on a removable cartridge 25 A- 25 E. Furthermore, the five MNET bus segments 22 A- 22 E may provide electrical isolation to each of the respective five controllers 20 A- 20 E to facilitate hot-plug removal and/or replacement of each of the five memory cartridges 25 A- 25 E. The RAID functionality described herein allows any one of the five memory segments 25 A- 25 E to be removed while the system 10 continues to operate normally but at degraded performance, (i.e. in a non-redundant mode). Once the removed memory segment is reinstalled, the data is rebuilt from the other four memory segments, and the memory system resumes operation in its redundant, or fault-tolerant, mode.

[0026] In this embodiment, each of the memory segments 24 A- 24 E may include one to eight dual inline memory modules (DIMMs). As described above, such DIMMs are generally organized with an X4 or an X8 device. As discussed below, X8 memory organization may defeat the typical ECC capability to detect a failure in a single device. Thus in current systems, the advantages of using X8 memory devices may be abandoned in favor of the X4 memory organization since current ECC algorithms may provide a more reliable memory subsystem.

[0027] The memory segments 24 A- 24 E may be organized on a single channel or on 2N channels, where N is an integer. In this particular embodiment, each of the memory segments 24 A- 24 E is divided into two channels—a first channel 29 A- 29 E and a second channel 31 A- 31 E, respectively. Since each memory segment 24 A- 24 E in this embodiment is capable of containing up to eight DIMMs, each channel may be adapted to access up to four of the eight DIMMs. Because this embodiment include two channels, each of the memory controllers 20 A- 20 E may essentially includes two independent memory controllers.

[0028] Returning to the exemplary system 10 illustrated in FIG. 1 , the host/data controller 16 , 18 is typically coupled to one or more bridges 28 A- 28 C via a suitable bus 27 . The opposite side of each bridge 28 A- 28 C is coupled to a respective bus 30 A- 30 C, and a plurality of peripheral devices 32 A and 32 B, 34 A and 34 B, and 36 A and 36 B may be coupled to the respective buses 30 A, 30 B, and 30 C. The bridges 28 A- 28 C may be any of a variety of suitable types, such as PCI, PCI-X, EISA, AGP, etc.

[0029] As previously discussed, it may be advantageous to provide one or more mechanisms to improve fault tolerance in the memory array 26 . One mechanism for improving fault tolerance is to provide ECC algorithms corresponding to one or more DIMMs which are able to recreate data for a failed device by incorporating extra storage devices which are used to store parity data (as with the presently described 4+1 parity scheme). For instance, on an X4 DIMM, sixteen of the memory devices may be used to store normal data while two of the memory devices may be used to store parity data to provide fault tolerance. The parity data can be used by the ECC algorithm to reproduce erroneous data bits.

[0030] Current ECC algorithms can detect and correct single-bit errors in X4 memory devices. Disadvantageously, current ECC algorithms cannot reliably detect multi-bit errors in X8 devices, much less correct such errors. The present RAID system, in conjunction with typical ECC algorithms, provides a mechanism for correcting multi-bit errors in X4 devices, as described below. In fact, the presently described system provides a mechanism for overcoming the complete failure of a memory device. “Chipkill” is an industry standard term referring to the ability of a memory system to withstand a complete memory device failure and continue to operate normally. Advantageously, when an error occurs, it effects only a single device. Thus, chipkill eliminates the vast majority of error conditions in a memory system. Disadvantageously, X4 chipkill may not provide sufficient fault tolerance to meet the requirements for all systems.

[0031] Memory devices are gradually transitioning from X4 to X8 devices due to larger capacities per device. However, while DIMMs incorporating X8 devices have become more common place, the ECC algorithm used in industry standard devices has remained the same. The standard ECC algorithm, such as the P6 algorithm defined by Intel, can only provide chipkill detection capabilities for X4 (or narrower) devices. For those systems that incorporate X8 capabilities, it may be advantageous to provide an ECC algorithm capable of reliably detecting multi-bit errors in X8 devices. Further, a mechanism for correcting the multi-bit errors in X8 memory devices would be advantageous. However, it may also be beneficial to provide a system whose fault tolerance is not degraded if X4 memory devices are incorporated.

[0032] The P6 ECC is an encoding scheme which can correct single data bit errors, and detect multi-bit errors, but only up to four adjacent bits within the same nibble or 4 adjacent bits. When a chipkill greater than 4 bits occurs, the P6 ECC logic may inadvertently correct the wrong data bit. In some cases, the P6 ECC cannot detect some multi-bit errors at all. Therefore, the XOR engine (described with reference to FIGS. 2 - 4 below) cannot determine which DIMM is failing. For example, if all data bits in the first byte (bit 0 - 7 ) or multiple bits from two adjacent nibbles are corrupted, the resultant syndrome (which is used to identify errors, as described below) of the P6 ECC is sometimes “zero.” This represents a “no error” condition indicating that no memory error is detected. In the case of a “no error” condition, the P6 ECC does not perform any flagging or correction functions. In other cases, the P6 ECC may wrongly interpret a MBE as an SBE and correct the wrong bit and pass along the data as good. Disadvantageously, in both of these cases, the P6 ECC logic incorrectly interprets the data and cannot pinpoint the bad DIMM. Thus, single bit error correction (SEC) ECC algorithms, such as the PA ECC algorithm, cannot maintain data integrity of an 8-bit or larger device. This may be a serious problem when a system is being used to run mission critical applications, since memory data corruption may occur without the system detecting the error.

[0033] With the implementation of wider memory devices, such as X8 memory devices, it would be desirable for fault-tolerant memory to be able to detect all possible errors in the X8 memory devices, as well as narrower memory devices, such as X4 memory devices. By implementing a new dual mode ECC parity checking algorithm or “matrix” in conjunction with the presently described RAID memory system, the present system is able to provide chipkill detection and correction capabilities for X8 memory devices. The presently described matrix is simply one exemplary embodiment of an algorithm that is capable of detecting multi-bit errors in an 8-bit (byte) segment. Numerous alternate embodiments of the matrix may be utilized in the present system, as can be appreciated by those skilled in the art.

[0034] The dual mode ECC is a dual-purpose error correcting/checking code. When operating in a X4 mode, it acts as a single bit correcting code and can detect errors in the adjacent 4 bits within the same nibble (similar to P6 ECC). When operating in X8 mode, it acts as an 8 adjacent bits within a single byte detection code (no single-bit correction is allowed). Error flag(s) can be used to identify which DIMM/device is bad. In X8 mode, the XOR engine 48 can be used to correct both single bit and multi-bit (in a byte) errors. An exemplary dual-mode ECC parity checking matrix is illustrated with reference to Table 1, in Appendix 1.

[0035] Table 1 defines eight bits that define the syndrome for the error correcting code, wherein each row of Table 1 corresponds to a syndrome bit. The eight syndrome bits are created by eight equations defined by Table 1. Each row of the table represents 72-bits that include 64 data bits and 8 ECC bits. As the 72-bits are read, the data bits are combined in a manner defined by the matrix in Appendix 1. For each row of bits, the data residing at each bit location where a logical 1 is illustrated in Table 1, the bit at that location is combined with all other bits in locations corresponding to the logical is in Table 1. The bits are combined together as logical XORs. Thus, each row of Table 1 defines an equation that is used to combine 64 data bits and 8 check bits to produce a single bit logical result. As indicated, row 1 corresponds to syndrome bit [ 0 ], row 2 corresponds to syndrome bit [ 1 ], etc. The eight syndrome bits are ordered to produce a single HEX number known as the syndrome that can be interpreted using Table 2, illustrated in Appendix 2. The eight syndrome bits are ordered from syndrome bit [ 7 ] to syndrome bit [ 0 ] to form the index used in Table 2.

[0036] An exemplary dual-mode ECC syndrome interpretation table is illustrated with reference to Appendix 2, below. All uncorrectable errors (“UNCER”) will be flagged but will not be corrected by the ECC module. That is to say that the error looks like a multi-bit error and cannot be corrected by the ECC code. If no error is detected, the HEX value (i.e. the syndrome) will be “00.” If a single bit error is detected, the HEX value produced by Table 1 will correspond to an error location identifying the single bit error. For example, if Table 1 produces a value of 68hex, a single bit error was detected at data bit 27 (DB 27 ). If the ECC algorithm is operating in a X4 or normal ECC mode, the single bit error will be corrected by the ECC code. If the ECC algorithm is operating in a X8 mode, any memory error (single or multi-bit) will not be corrected by the ECC code. In fact, if the ECC algorithm is operating in a X8 mode, all syndrome codes except 00hex are considered uncorrectable errors that will not be corrected by the ECC algorithm. This is because in X8 mode some MBEs will map to SBE correctable errors (i.e. the errors are miscorrected or “misaliased”) Check bits, like data bits, can be in error. If the HEX value is 01, then the bit error is the ECC code bit CB 0 .

[0037] Thus, the dual-mode ECC algorithm may be implemented to effectively handle errors in X8 memory devices. The present system provides a memory array with X8 chipkill by implementing the dual-mode ECC algorithm in conjunction with the RAID logic to provide a new level of memory fault tolerance. To implement X8 memory devices in the memory array 26 , the single-bit error correction (SEC) in the dual mode ECC algorithm is disabled and RAID logic is used to correct both single-bit and multi-bit errors detected by the dual mode ECC algorithm. For X4 devices, the SEC in the dual mode ECC algorithm may be enabled.

[0038] The manner in which an exemplary “4+1” RAID architecture functions will now be explained with reference to FIG. 2 . During a memory read operation, a quad-word from each of the first four memory segments 24 A- 24 D and parity bits from the one remaining memory segment 24 E are transmitted to the respective memory controllers 20 A- 20 E. When X4 memory devices are implemented in the memory system 26 and the system is operating in X4 memory mode, each of the memory controllers 20 A- 20 E uses the dual-mode ECC algorithm to detect and correct single bit memory errors and detect multi-bit errors. Each memory controller 20 A- 20 E includes an ECC module 38 A- 38 E to implement the dual mode ECC algorithm and the SEC algorithm.

[0039] When operating in X8 memory mode (i.e. the memory system 26 comprises X8 memory devices), each of the memory controllers 20 A- 20 E uses the dual-mode ECC algorithm in the ECC modules 38 A- 38 E to detect single bit errors, but such errors are not corrected by the ECC modules 38 A- 38 E in the memory controllers 20 A- 20 E. Thus, the SEC which may be provided in typical ECC schemes is disabled. If an error is detected in the memory controller 20 A- 20 E, it may be logged in the memory controller as a single bit error (SBE) or multi-bit error (MBE). The logged information may be retained for trouble shooting and error tracking purposes, for example. Further, an error flag may be set in the data packet to mark the quad-word as including an error. Different flag bits may be set to identify the error as a SBE or a MBE if it is desirable to retain a separate tracking of each type of error. Regardless of the error type, the error is not corrected by the ECC module 38 A- 38 E in the memory controller 20 A- 20 E.

[0040] Once the memory controllers 20 A- 20 E have processed the data as discussed above, the data is transferred via the respective buses 22 A- 22 E to the host/data controller 16 , 18 . The host/data controller 16 , 18 also includes ECC modules 40 A- 40 E implementing the dual-mode ECC algorithm to detect errors in each of the four quad-words and the parity information delivered from the respective memory controllers 20 A- 20 E. Once again, if the system is operating in X4 mode, the ECC module 40 A- 40 E detects SBEs and MBEs. If an SBE is detected, it can be corrected by the SEC algorithm in the ECC module 40 A- 40 E. If a MBE is detected, the error is flagged such that it may be corrected by the RAID mechanism discussed below. If the system is operating in an X8 mode, the SEC is disabled and any errors detected by the ECC modules 40 A- 40 E are flagged but not corrected by the ECC module 40 A- 40 E.

[0041] Each ECC module 40 A- 40 E is coupled to a respective multiplexor 44 A- 44 E, via an error flag path 42 A- 42 E. If an error is detected by one of the ECC modules 40 A- 40 E, a signal is sent to the multiplexor 44 A- 44 E. Based on whether an error flag is sent via the error flag path 42 A- 42 E, each respective multiplexer 44 A- 44 E selects between the original data delivered to the multiplexors 44 A- 44 E on respective buses 46 A- 46 E and the re-created data generated by the exclusive OR (XOR) engine 48 delivered to the multiplexers 44 A- 44 E via the respective buses 50 A- 50 E. Specifically, if one of the ECC modules 40 A- 40 E detects an error, the ECC module 40 A- 40 E switches its respective multiplexer 44 A- 44 E such that the bad data on the respective bus 46 A- 46 E is replaced by the good re-created data available on the respective bus 50 A- 50 E. It should be noted that the ECC module 40 A- 40 E actually corrects single bit errors and transmits the corrected data via the data bus 46 A- 46 E. However, if the system is operating in an X8 memory mode and the SEC correction in the ECC module 40 A- 40 E is disabled or, more accurately, “discarded,” the multiplexor 44 A- 44 E is toggled such that the data on the data bus 46 A- 46 E is ignored. Thus, in one embodiment of the present system, the SEC in the ECC modules 40 A- 40 E is not actually disabled when the system is operating in X8 mode. Instead, the multiplexors 44 A- 44 E are used to effectively disable the SEC by simply selecting the data delivered from the XOR engine 48 on buses 50 A- 50 E (and discarding the data delivered on the buses 46 A- 46 E), even if any error (including a SBE) is detected.

[0042] The data controller 18 also includes parity compare logic 54 A- 54 E. Data is delivered to the parity compare logic 54 A- 54 E from the ECC module 40 A- 40 E via bus 46 A- 46 E and the XOR engine via bus 50 A- 50 E. The parity compare logic 54 A- 54 E is used to protect against parity miscompares, i.e., a non-match between the data from the ECC module 40 A- 40 E and the data from the XOR engine 48 . When the system is operating in a X4 mode, SEC is enabled. Thus, if a SBE is detected by the ECC module 40 A- 40 E, it will be corrected by the ECC module 40 A- 40 E. If the data is properly corrected, the data from the bus 46 A- 46 E should match the data on respective bus 50 A- 50 E. If a parity mismatch occurs at the parity compare logic 54 A- 54 E, an NMI may be initiated via a respective line 56 A- 56 E. When operating in X8 mode, the parity compare logic 54 A- 54 E is used as a fail safe to ensure that errors are not missed by always comparing the data from the ECC modules 40 A to the corresponding data from the XOR engine 48 . If data miscompares are detected, a non-maskable interrupt (NMI) may be initiated, as further discussed below.

[0043] An introduction to an exemplary embodiment of the functionality of the data striping and correction techniques as implemented during a read operation are briefly described with reference to FIGS. 3 and 4 . In general, RAID memory provides 4+1 memory data redundancy by splitting or “striping” a single cacheline of data across four data segments, plus a fifth parity segment. This configuration provides redundancy such that any data segment can be recreated from the other four data segments if an error is detected. The data may be recreated using the XOR engine 48 . As previously described, the XOR engine 48 relies on the ECC modules 38 A- 38 E and 40 A- 40 E to detect errors, and the XOR engine 48 may be implemented to correct those errors.

[0044] Turning now to FIG. 3 , during a read operation, RAID memory stripes a cache line of data 65 such that each of the four 72-bit data words 66 , 67 , 68 , and 69 is transmitted through a separate memory controller (or “memory control device”) 20 A- 20 D. A fifth parity data word 70 is generated from the original data line. Each parity word 70 is also transmitted through a separate memory controller 20 E. The generation of the parity data word 70 from the original cache line 65 of data words 66 , 67 , 68 , and 69 can be illustrated by way of example. For simplicity, four-bit data words are illustrated. However, it should be understood that these principals are applicable to 72-bit data words, as in the present system, or any other useful word lengths. Consider the following four data words:

[0045] DATA WORD 1 : 1 0 1 1

[0046] DATA WORD 2 : 0 0 1 0

[0047] DATA WORD 3 : 1 0 0 1

[0048] DATA WORD 4 : 0 1 1 1

[0049] A parity word can be either even or odd. To create an even parity word, common bits are simply added together. If the sum of the common bits is odd, a “1” is placed in the common bit location of the parity word. Conversely, if the sum of the bits is even, a zero is placed in the common bit location of the parity word. In the present example, the bits may be summed as follows:

[0050] DATA WORD 1 : 1 0 1 1

[0051] DATA WORD 2 : 0 0 1 0

[0052] DATA WORD 3 : 1 0 0 1

[0053] DATA WORD 4 : 0 1 1 1

[0054] NUMBER OF 1's: 2 1 3 3

[0055] PARITY WORD: 0 1 1 1

[0056] When summed with the four exemplary data words, the parity word 0111 will provide an even number of active bits (or “1's”) in every common bit. This parity word can be used to re-create any of the data words (1-4) if an error is detected in one of the data words as further explained with reference to FIG. 4 .

[0057] FIG. 4 illustrates the re-creation of a data word in which an error has been detected in one of the ECC modules 38 A- 38 E or 40 A- 40 E implemented during a read operation. As in FIG. 3 , the original cache line 65 comprises four data words 66 , 67 , 68 , and 69 and a parity word 70 . Further, the memory control devices 20 A- 20 E corresponding to each data word and parity word are illustrated. In this example, a data error has been detected in the data word 68 . A new cache line 72 can be created using data words 66 , 67 , and 69 along with the parity word 70 using the XOR engine 48 . By combining each data word 66 , 67 , 69 and the parity word 70 in the XOR engine 48 , the data word 68 can be re-created. The new and correct cache line 72 thus comprises data words 66 , 67 , and 69 copied directly from the original cache line 65 and data word 68 a (which is the re-created data word 68 ) which is produced by the XOR engine 48 using the error-free data words ( 66 , 67 , 69 ) and the parity word 70 . It should also be clear that the same process may be used to re-create a parity word 70 if an error is detected therein.

[0058] Similarly, if the memory controller 20 A- 20 E, which is associated with the data word 68 , is removed during operation (i.e. hot-plugging) the data word 68 can similarly be re-created. Thus, any single memory controller can be removed while the system is running or any single memory controller can return a bad data word and the data can be re-created from the other four memory control devices using the XOR engine 48 .

[0059] Similarly, the XOR engine 48 is used to generate parity data during a write operation. As a cache line of data is received by the host/ data controller 16 , 18 , the cache line is striped such that the data may be stored across four of the segments 24 A- 24 D. In one embodiment, the cache line is divided to form four 72-bit data words. As described with reference to FIGS. 3 and 4 and the read operation, the XOR engine 48 may be implemented to generate a parity data word from each or the four data words. The parity data word may then be stored in the memory segment 24 E to provide parity data for use in the read operations. It should also be noted that the ECC modules 40 A- 40 E and 38 A- 38 E may also be implemented during the write operation to detect and correct single bit errors in the data words and/or parity word.

[0060] Returning to FIG. 2 , when operating in X4 memory mode, the XOR engine 48 may be used to correct multi-bit errors only in addition to the error correction in the ECC module 38 A- 38 E in each memory controller 20 A- 20 E, as described above. Conversely, when operating in X8 memory mode, the XOR engine 48 corrects both single bit errors and multi-bit errors. That is to say that the SEC algorithm is disabled (i.e., the data corrected by the SEC algorithm in the ECC module 40 A- 40 E is ignored) when operating in X8 mode. When operating in X4 memory mode, each memory segment 24 A- 24 E may exhibit a single bit error which may be corrected by the ECC module 40 A- 40 E without even triggering the use of the re-created data generated by the XOR engine 48 . However, if a multi-bit error is detected, only a single memory error (single or multi-bit) on one of the memory segments 24 A- 24 E can be corrected per each memory transaction using the XOR engine 48 . When operating in X8 memory mode, the host/data controller 16 , 18 can correct only one single bit error or multi-bit error in one of the memory segments 24 A- 24 E. Thus, if more than one of the memory segments 24 A- 24 E exhibits a single bit error or a multi-bit error in X8 memory mode, or if more than one of the memory segments 24 A- 24 E exhibits a multi-bit error in X4 memory mode, the XOR engine 48 will be unable to create good data to be transmitted out of the host/data controller 16 , 18 on the buses 52 A- 52 E. To track this condition (i.e. simultaneous errors on multiple memory segments 24 A- 24 E), the XOR engine 48 may be notified via an error detection signal each time an error is detected, as further described below.

[0061] As described above, one side-effect of changing to the presently described approach to error correction is that a class of error that previously was corrected by the ECC SEC logic, becomes an uncorrectable error condition. Recall that when the system is operating in X4 memory mode, the dual-mode ECC module 38 A- 38 E and/or 40 A- 40 E is able to detect and correct single-bit errors. Thus, if multiple ECC modules 38 A- 38 E or 40 A- 40 E detect single-bit errors on the same transaction (“SBE-SBE condition”), each of these errors will be detected and corrected by the SEC code in the corresponding ECC modules 38 A- 38 E or 40 A- 40 E and the transaction will complete without necessitating a system interrupt or other-wise degrading system performance.

[0062] Conversely, in X8 memory mode, single-bit error correction by ECC is turned off. By turning off single-bit error correction, the above-mentioned misaliasing problem with X8 devices is eliminated and the dual-mode ECC algorithm provides X8 chipkill detection. In this mode of operation, all errors, including single-bit errors, are corrected by the XOR engine 48 . The drawback to this solution is that the previously harmless SBE-SBE condition described above will now overwhelm the XOR engine 48 , forcing it to produce a NMI. That is, if two ECC modules 38 A- 38 E or 40 A- 40 E experience a SBE on the same transaction, where previously they would have both corrected the data, will both pass uncorrected data. Since the XOR engine 48 can only fix one error per transaction, an uncorrectable error has occurred and a NMI is produced. An NMI is generated whenever an event occurs of which the operating system should be notified. After an NMI event is handled by the operating system, the system is usually reset and/or rebooted.

[0063] When operating in X8 mode, one solution to the SBE-SBE condition is a trade-off. If one single-bit error occurs, the XOR engine 48 corrects the error, but if more than one single-bit error occurs on a single transaction, then correct one or all of the single-bit errors using the ECC module 40 A- 40 E rather than the XOR engine 48 . The assumption is that if an SBE-SBE condition occurs, it is far more likely that the error is, in fact, due to multiple independent single-bit errors and not an X8 chipkill error misaligning with a single-bit error. Furthermore, even if it is an X8 chipkill, the error will result in a parity miscompare at the comparator circuits 54 A- 54 E and the system will produce a NMI.

[0064] To implement this solution in X8 mode, an XOR mux controller 58 in the XOR engine 48 may be implemented. If an error is detected in an ECC module 40 A- 40 E, an error flag is sent to the XOR mux controller 58 via an error flag path 60 A- 60 E. The error flag indicates that an error has been detected for a particular data word and further indicates whether the error is a single-bit error or a multi-bit error. The XOR mux controller 58 performs a comparison of each data word and the parity word for each transaction. Thus, the XOR mux controller 58 determines if an error is detected in more than one ECC module 40 A- 40 E in a single transaction. If two errors are detected, the XOR mux controller 58 toggles one or both of the respective multiplexors 44 A- 44 E via the error flag path 62 A- 62 E such that the data corrected by the ECC module 40 A- 40 E can be transmitted. Recall that if an error is detected in the ECC module 40 A- 40 E, the respective multiplexor 44 A- 44 E is toggled such that the data word is transmitted from the XOR corrected data bus 50 A- 50 E. If two errors are detected on the same transaction, one or both of the respective multiplexors 40 A- 40 E is toggled back by the XOR mux controller 58 such that the data is transmitted from the data bus 46 A- 46 E carrying the data corrected by the ECC module 40 A- 40 E. If both of the errors are SBEs, the XOR mux controller 58 may be configured to re-toggle either one or both of the respective multiplexors 44 A- 44 E. If one of the errors is a SBE and the other error is a MBE, the XOR mux controller 58 toggles only the multiplexor 44 A- 44 E corresponding to the SBE data path. That is to say that the SBE is corrected by the corresponding ECC module 40 A- 40 E while the MBE is corrected by the XOR engine 48 . If both errors are MBEs, an NMI may be initiated as described above.

[0065] The present techniques can also be implemented for X16 and X 32 devices, as well. One implementation incorporating X16 devices is illustrated with reference to FIG. 5 . Generally speaking, for a system implementing X16 devices and having X16 chipkill, the memory array 26 ( FIG. 1 ) comprises two memory subsystems 26 A and 26 B, as illustrated in FIG. 5 . For simplicity, for each memory subsystem 26 A and 26 B, a single memory cartridge 25 AA and 25 BA, respectively, are illustrated. However, as can be appreciated, the memory subsystem 26 A may include five memory cartridges 25 AA- 25 AE, and the memory subsystem 26 B may include five memory cartridges 25 BA- 25 BE. Each of the memory cartridges 25 includes respective memory segments 24 and respective memory controllers 20 , as previously described. In the present exemplary embodiment, a memory cartridge 25 AA having a memory segment 24 AA and memory controller 20 AA is illustrated in the memory subsystem 26 A. Similarly, a memory cartridge 25 BA having a memory segment 24 BA and memory controller 20 BA is illustrated in the memory subsystem 26 B. In the present exemplary embodiment, each of the memory devices in the memory segments 24 AA- 24 AE and 24 BA- 24 BE comprises a X16 SDRAM device.

[0066] Referring briefly to FIG. 2 , when the system includes X4 or X8 memory devices, each ECC module 40 A- 40 E is configured to receive data from or pass data to a respective one of the memory segments 24 A- 24 E. As previously described, each ECC module 40 A- 40 E implementing the dual mode ECC algorithm is able to detect errors in up to 8 adjacent bits.

[0067] To facilitate X16 chipkill, the data controller 18 comprise two sets of ECC modules 40 AA- 40 AE and 40 BA- 40 BE and two XOR engines 48 A and 48 B, as illustrated in FIG. 5 . Generally speaking, during a read operation, the first 16 bits from the same SDRAM device may be split evenly between ECC modules. As previously described, each ECC module 40 AA- 40 AE and 40 BA- 40 BE is capable of detecting memory errors within the same byte (i.e. 8 adjacent bits). Accordingly, two ECC modules operating in tandem can detect any error within the same 16-bit memory device. Data correction will be handled by the respective XOR engines 48 A and 48 B.

[0068] To facilitate the division of bits from each X16 device in the memory segments 24 AA- 24 AE, a splitter 80 A is provided. The splitter 80 A is configured to split the data bits evenly between two corresponding ECC units. For instance, during a first read operation, a 16-bit data word is sent from the SDRAM device in the memory segment 24 AA, through the memory controller 20 AA and to the splitter 80 A. The 16-bit data word is divided into two 8-bit data words. The first 8-bit data word may be delivered to the ECC module 40 AA, and the second 8- bit data word may be delivered to the ECC module 40 BA. Each of the ECC modules 40 AA and 40 BA is configured to detect errors in up to 8 bits of data.

[0069] Similarly, to facilitate the division of bits from each X16 device in the memory segments 24 BA- 24 BE, a splitter 80 B is provided. The splitter 80 B is configured to split the data bits evenly between two corresponding ECC units. For instance, during a first read operation, a 16-bit data word is sent from the SDRAM device in the memory segment 24 BA, through the memory controller 20 BA and to the splitter 80 B. The 16-bit data word is divided into two 8-bit data words. The first 8-bit data word may be delivered to the ECC module 40 AA, and the second 8-bit data word may be delivered to the ECC module 40 BA. Each of the ECC modules 40 AA and 40 BA is configured to detect errors in up to 8 bits of data.

[0070] As can be appreciated, the dual sets of ECC modules 40 AA- 40 AE and 40 BA- 40 BE may be implemented in tandem to detect any errors in the X16 memory devices. Further, and as described above with reference to the X8 chipkill, each of the XOR engines 48 A and 48 B may be implemented to provide error correction in any of the 8-bit data words. By implementing two XOR engines 48 A and 48 B, along with the corresponding ECC modules 40 AA- 40 AE and 40 BA- 40 BE, any errors detected in the 8-bit data words may be corrected. Accordingly, the system of FIG. 5 illustrates a technique for providing X16 chipkill.

[0071] Similarly, X32 chipkill may be provided by implementing a third and fourth memory subsystem 26 C and 26 D (not illustrated), along with an additional splitter, an additional set of ECC modules in the data controller 18 and an additional XOR engine for each corresponding memory subsystem 26 C and 26 D. The splitter for a X32 chipkill system would be configured to split a 32-bit word into four 8-bit data words. As can be appreciated, each set of ECC modules and each XOR engine would be capable of detecting/correcting errors in up to 8 adjacent bits. Accordingly, operating in tandem, the presently described system provides for X32 chipkill, as can be appreciated by those skilled in the art.

[0072] While the invention may be susceptible to various modifications and alternative forms, specific embodiments have been shown by way of example in the drawings and will be described in detail herein. However, it should be understood that the invention is not intended to be limited to the particular forms disclosed. Rather, the invention is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the invention as defined by the following appended claims.