The present invention relates to a method for at-speed storage of faults for Built-In Self-Repair (BISR) of embedded-RAMs according to the pre-characterizing portion of claim 1. Furthermore the invention relates to a fault storage table circuit device for at-speed storage of faults for Built-In Self-Repair (BISR) of embedded-RAMs according to the pre-characterizing portion of claim 6.

A generic method for at-speed storage of faults for Built-In Self-Repair (BISR) of embedded-RAMs according to the pre-characterizing portion of claim 1 as well as a generic fault storage table circuit device for at-speed storage of faults for Built-In Self-Repair (BISR) of embedded-RAMs according to the pre-characterizing portion of claim 6 are known from Borri, S.: "A recursive 'at-speed' Built-In Self-Repair (BISR) method for the allocation of redundant rows and columns in an embedded-RAM", Siemens Technology Report, Dec. 2002, pp. 37 to 41.

Large embedded-RAMs usually employ wordline and/or bitline redundancy for yield improvement. When both spare rows and columns are present, the allocation of the redundant resources can be efficiently determined only when the full memory failure bitmap is known in advance. Since the complete bitmap can not be stored on-chip, there is a need for a mechanism to reduce the amount of information necessary to perform the diagnosis, while keeping full repair efficiency. The repair efficiency is defined as the following quotient: number of repaired memories / number of repairable memories.

During "at-speed" memory test, an additional constraint is that the test algorithm can not be stopped; otherwise the delay fault coverage would be affected. Therefore the storage of the faulty bitmaps has to be performed on-line during the BIST (Built-In Self-Test) algorithm and must complete within one clock cycle. For word-based memories (like the embedded-SRAMs, i.e. Static Random Access Memories, commonly used at Infineon) this implies the need to handle several bit faults in the same logical word, which can be detected in the same read cycle. So multiple-bit faults can appear at each cycle, and all of them must be handled within one clock cycle.

Most Built-In Self-Analysis / Repair (BISA / BISR) solutions available today focus on compressing the faulty bitmaps to reduce the area of the storage table. Being this compression performed during BIST, these methods need significant on-line processing capability during the test, so that they can not complete within only one clock cycle (i.e. they are not suitable for at-speed BIST) or their implementation leads to a large logic area overhead.

The architecture described in Bhavsar, D. K.: "An Algorithm for Row-Column Self-Repair of RAMs and Its Implementation in the Alpha 21264", Proc. of the International Test Conference (ITC) 1999, pp. 311 to 318 is possibly implemented with no cycle penalty but it is intended only for a one-row one-column redundancy with no extension possible.

In Kim, H.-C. et al.: "A BISR (Built-In Self-Repair) circuit for embedded memory with multiple redundancies", Proc. of IEEE 6th International Conference on VLSI and CAD (ICVC), 1999, pp. 602 to 605 the BIST has to be stopped for several cycles to allow the BISR engine to perform the needed on-line calculations. This pre-processing is necessary to analyze the failure bitmap and update the so-called Row / Column Fill Entries.

The architecture presented in Kawagoe, T. et al.: "A Built-In Self-Repair Analyzer (CRESTA) for embedded DRAMs", Proc. of the International Test Conference (ITC) 2000, pp. 567 to 574 is quite general but it makes use of dedicated CAMs whose area overhead increases exponentially with the number of spare rows and columns. Moreover single-cycle operation is achieved for bit-oriented memories only and it is not directly extendable to word-based memories without losing clock cycles.

The redundancy analyzer described in Nakahara, S. et al.: "Built-In Self-Test for GHz Embedded SRAMs Using Flexible Pattern Generator and new Repair Algorithm", Proc. of the International Test Conference (ITC) 1999, pp. 301 to 310 handles word-based memories with a single-cycle architecture, but at the expense of a reduced repair efficiency. This is caused by the use of a lossy compression scheme in which the information about bit error locations gets lost in case that multiple bit errors are detected in the same memory word.

The Repair Analysis (RA) unit described in Nagura, Y.: "Test cost reduction by at-speed BISR for embedded DRAMs", Proc. of the International Test Conference (ITC) 2001, pp. 182 to 187 solves all the previous issues reaching high repair efficiency. However parallel processing of the I/O data lines and pipelined structures cause a significant area increase (20 Kgates + 9 small SRAMs => 1.7 sqmm in 0.18 µm technology for the complete BISR). Thus this solution can be efficiently employed only for very large embedded-RAMs.

The main issue of the above described solutions is the need for on-line processing during BIST. Eventually the advantage coming from the size reduction of the fault storage table is weighed down by the increase of the preprocessing logic needed for bitmap compression. The situation gets worse for word-oriented memories in which several bit failures can be detected in the same clock-cycle, so that extensive parallel processing is needed to ensure single-cycle operation of the pre-processing logic.

It is an object of the present invention to provide a method and a fault storage table circuit device for at-speed storage of faults for Built-In Self-Repair (BISR) of embedded-RAMs, which method and device require reduced on-line computational overhead and provide a better speed performance in comparison with the state of the art.

According to the invention this object is achieved by a method according to claim 1 and by a device according to claim 4.

According to the present invention it is proposed to store the full failure bitmap for each failing word in a fault storage table. No bitmap compression is performed, significantly reducing the on-line computational overhead in comparison with most of the methods and devices from the state of the art. Moreover, the storage of the faults in the same cycle in which they are detected, is a feature which ensures a better speed performance of the method and device according to the invention even in comparison with the above mentioned generic method and device using full failure bitmap storage for each failing word, too.

The main advantages of the solution according to the invention are as follows:

• Reduced online processing during test. No bitmap compression necessary. The redundancy analysis is delayed to the end of test, so that it can be efficiently performed by low-speed logic (possibly by an on-chip microprocessor) with no need for dedicated parallel structures.
• One-cycle architecture. Suitable for at-speed BIST of word-based memories.
• Independent of the BIST algorithm.
• Can be used for ANY configuration of redundant resources.
• Small area overhead when compared to solutions with the same characteristics (e.g. compared with the solution described in Nagura, Y.: "Test cost reduction by at-speed BISR for embedded DRAMs", Proc. of the International Test Conference (ITC) 2001, pp. 182 to 187).
• The dimension N of the fault storage table can be traded off for the repair efficiency.

Preferred and advantageous embodiments of the method according to the invention are subject matter of claims 2 to 5. Preferred and advantageous embodiments of the device according to the invention are subject matter of claims 6 to 13.

Examples of preferred and advantageous embodiments of the invention will now be described with respect to the accompanying drawings in which

Fig. 1

is a top level block diagram of an embodiment of a fault storage table circuit device according to the invention,

Fig. 2

is a scheme of a circuit architecture of the i-th line in the fault storage table of the device of Fig. 1, and

Fig. 3

is a scheme of a circuit architecture of the common control logic block of the device of Fig. 1.

An embodiment of a fault storage table circuit device according to the invention, as it will be described in more detail hereinbelow, comprises a fault storage table and a common table control logic block (cf. Fig. 1). The fault storage table stores the wordline address, the faulty bitmap and a flag indicating that the location has already been used. The number (N) of lines of the fault storage table is determined in such a way that, if all the lines have been used, there is at least one forced choice for a spare wordline or a spare bitline. Therefore the redundancy allocation can start with no impact on the repair efficiency, for instance - but not necessarily - applying the recursive algorithm described in Borri, S.: "A recursive 'at-speed' Built-In Self-Repair (BISR) method for the allocation of redundant rows and columns in an embedded-RAM", Siemens Technology Report, Dec. 2002, pp. 37 to 41.

Otherwise, if not all the lines are used, then all the memory faults have been stored and a full redundancy allocation algorithm, e.g. the exhaustive search procedure described in Kawagoe, T. et al.: "A Built-In Self-Repair Analyzer (CRESTA) for embedded DRAMs", Proc. of the International Test Conference (ITC) 2000, pp. 567 to 574, can now be run off-line.

Given a memory array with m spare wordlines and p spare I/Os, it can be shown (see below at the end of the current specification) that the minimum value of N which is necessary to have at least one forced choice is:$${\mathrm{N}}_{\min }\mathrm{=}\mathrm{m}\mathrm{\cdot }\left(\mathrm{p}\mathrm{+}\mathrm{1}\right)\mathrm{+}\mathrm{1}\mathrm{=}\mathrm{m}\mathrm{\cdot }\mathrm{p}\mathrm{+}\mathrm{m}\mathrm{+}\mathrm{1}$$

If this condition is met, it is ensured that a proper redundancy allocation algorithm (see e.g. Borri, S.: "A recursive 'at-speed' Built-In Self-Repair (BISR) method for the allocation of redundant rows and columns in an embedded-RAM", Siemens Technology Report, Dec. 2002, pp. 37 to 41.) can lead to the optimal repair solution (maximum repair efficiency). However the size of the fault storage table N can be reduced if one accepts a slight degradation of the repair efficiency. In this case the final repair rate will depend on the N value and on the particular redundancy allocation algorithm chosen.

The architecture of an embodiment of a proposed fault storage table according to the invention is shown in Fig. 1.

The fault storage table is composed of N lines, each one including a wordline address register WL(i), a faulty bitmap register BMP(i) and an enable flag register EN(i). The detailed diagram of the i-th line is shown in Fig. 2.

When an error is detected (bist_error becomes set for one cycle), if the current faulty wordline address (cWL) is equal to the already stored WL(i) and the corresponding enable flag is set, then an update signal update(i) is activated for the i-th line of the table. Then the failure bitmap of the i-th line is updated by a bit-by-bit OR between the current bitmap value cBMP and the old one BMP(i).

The architecture of the control logic block is depicted in Fig. 3. A common counter cur_pos is used to point to the first available (blank) line of the fault storage table. The value of the common counter cur_pos can vary between 0 and N, wherein N, as above, is the number of lines in the fault storage table, so its output has B bits, where B=ceiling(log₂(N+1)). In case that the wordline address is not already stored in the fault storage table, i.e. all the update(i) signals are zero, the common logic block activates a load signal load(i) (with i=cur_pos) which commands the cur_pos-th line to store the current wordline address and faulty bitmap coming from the BIST, also setting the corresponding EN(i) flag. At the same time an incrementing signal inc_pos is activated, which increments the common counter cur_pos in the following cycle. The load(i) and inc_pos signals are activated only if the cur_pos value is less than N.

Therefore, once that the fault storage table is full (cur_pos=N) no more wordline addresses may be stored, and only the bitmap of the already stored wordlines can be updated. This is necessary because other bit faults belonging to the same wordline can be detected in subsequent phases of the test algorithm.

In view of the advantages of the present invention as outlined above, nevertheless it should be noted that at some embodiments of the present invention the area overhead may become important with large amount of redundancy because the table size N has to be proportional to m·p(m=number of spare wordlines, p=number of spare I/Os) to have the maximum repair efficiency. The table size also depends on the memory word width. Thus the area efficiency is better for memories with I/O based redundancy (as is the case for Infineon embedded SRAMs) and small I/O count. Furthermore it may happen that not all the faults get stored in the fault storage table during the first test pass. This implies the need to re-run the BIST several times for memories with complex failure bitmaps (see Borri, S.: "A recursive 'at-speed' Built-In Self-Repair (BISR) method for the allocation of redundant rows and columns in an embedded-RAM", Siemens Technology Report, Dec. 2002, pp. 37 to 41.).

VHDL (Very High-Speed Integrated Circuit Hardware Description Language) RTL (Register Level Transfer) descriptions, simulations and synthesis have been performed for the same case as it was simulated in Borri, S.: "A recursive 'at-speed' Built-In Self-Repair (BISR) method for the allocation of redundant rows and columns in an embedded-RAM", Siemens Technology Report, Dec. 2002, pp. 37 to 41 (i.e.: number of logical words = 8192, data width = 32, number of bitlines = 512, and number of wordlines = 512). The total area overhead is quite small: synthesis results in 0.18 µm technology for the case of the example (one memory of 8Kx32 with 1 redundant WL + 2 redundant I/Os, one for each side of the memory array, and fault storage table with N=3 locations) give a total area overhead (BIST + redundancy allocation + fault storage table + redundancy activation registers) of ~ 0.06 mm² (~5.5 Kgates). This figure is considerably smaller than the 1.7 mm² reported in Nagura, Y.: "Test cost reduction by at-speed BISR for embedded DRAMs", Proc. of the International Test Conference (ITC) 2001, pp. 182 to 187, although part of this difference is due to significant specifications discrepancies.

Summarizing it is stated that the present invention provides a mechanism to store detected memory faults, which mechanism minimizes on-line processing during the execution of the BIST algorithm and is thus suitable for at-speed BIST/BISR. No bitmap compression is performed and the full failure bitmap of a certain wordline is stored (and/or updated). This can easily be accomplished within one clock cycle without stopping the BIST algorithm. For word-oriented memories, in which multiple bit faults may be detected in the same memory access, direct bitmap storage avoids the need for massive parallel pre-processing and has no impact on the repair efficiency. This architecture is extensible to any configuration of spare wordlines/IOs, with the only drawback of requiring an increasing number of locations. However the fault table size can be traded off for the overall repair efficiency. A significant area advantage is demonstrated over equivalent solutions as they were described e.g. in Nagura, Y.: "Test cost reduction by at-speed BISR for embedded DRAMs", Proc. of the International Test Conference (ITC) 2001, pp. 182 to 187.

Last but not least as already announced above it should now be given an explication of the formula for N_min.

Given a memory array with r spare rows and s spare columns, it was shown in Bhavsar, D. K.: "An Algorithm for Row-Column Self-Repair of RAMs and Its Implementation in the Alpha 21264", Proc. of the International Test Conference (ITC) 1999, pp. 311 to 318 that the condensed maximally repairable sparse failure array (MSFA) has r*(s+1) rows and s*(r+1) columns. See Bhavsar, D. K.: "An Algorithm for Row-Column Self-Repair of RAMs and Its Implementation in the Alpha 21264", Proc. of the International Test Conference (ITC) 1999, pp. 311 to 318, and in particular Fig. 1 on page 313 there, for details of the MSFA definition.

According to an embodiment of the current invention the fault storage table stores partially condensed MSFA. The condensed values are in the vertical direction only (row addresses), and in the horizontal direction the full failure bitmaps are stored (there is no column "condensation").

Therefore, allowing for the storage of one more row than the condensed MSFA will lead to one of the following situations at the end of the test:

1. 1. If there is one row (column) with more faults than the number of available spare columns (rows), then at least one "row-must"/"column-must" forced choice exists. This forces the allocation of at least one spare row/column, and the test can now be run again.
2. 2. There is no forced repair and the table is full (i.e. all the lines have been used). Since the table contains more rows than the MSFA, then the memory is not repairable.
3. 3. There is no forced repair and the table is not full. In this case it is ensured that all the faults are stored in the table, so that the MSFA can now be analyzed by a proper allocation algorithm.

Hence, the minimum value of N needed to perform at least one sensible row/column choice is:$${\mathrm{N}}_{\min }\mathrm{=}\mathrm{Number\_of\_rows\_of\_MSFA}\mathrm{+}\mathrm{1}\mathrm{=}\mathrm{r}\mathrm{\cdot }\left(\mathrm{s}\mathrm{+}\mathrm{1}\right)\mathrm{+}\mathrm{1}\mathrm{=}\mathrm{r}\mathrm{\cdot }\mathrm{s}\mathrm{+}\mathrm{r}\mathrm{+}\mathrm{1}$$

If this condition is met, then there is no impact on the final repair efficiency, provided that a proper allocation algorithm is used. Such "proper" allocation algorithms are already known to persons skilled in the art. Several algorithms have been presented in the literature. One of these is the exhaustive procedure described in Kawagoe, T. et al.: "A Built-In Self-Repair Analyzer (CRESTA) for embedded DRAMs", Proc. of the International Test Conference (ITC) 2000, pp. 567 to 574.

It has to be noted that the formula given above is indeed applicable to the general case of r rows and s columns. However, it is not directly applicable to the case described in Borri, S.: "A recursive 'at-speed' Built-In Self-Repair (BISR) method for the allocation of redundant rows and columns in an embedded-RAM", Siemens Technology Report, Dec. 2002, pp. 37 to 41 since in that case the spare I/Os have some additional constraints in terms of allocation (each I/O can only be used for each half of the memory). In this case a minimum value for N still exists but it is fairly more complex to calculate.