A section of the disclosure of this patent document and its figures contain material subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document, but otherwise reserves all copyright rights whatsoever.
The present invention generally relates to testing of field programmable gate arrays. The present invention more particularly relates to testing field programmable gate arrays for delay faults.
A field programmable gate array (FPGA) is a general-purpose integrated circuit consisting of a two-dimensional array of programmable logic blocks interfaced with a programmable routing network and programmable input/output cells (I/O cells). By programming interconnections between the logic blocks, routing network, and I/O cells, a generic FPGA can be selectively configured to provide a wide variety of specific circuit functions.
It is desirable to thoroughly test FPGA's for defects. Two common defects are hard faults and delay faults. A hard fault is a defect that causes a functional failure within a circuit, while a delay fault is a defect that affects a circuit's delay. Though various conventional methods exist for efficiently testing FPGA's for hard faults, conventional methods of delay-fault testing are either non-comprehensive or require expensive test equipment and significant time to implement.
The conventional testing method that is regularly employed by FPGA manufacturers relies on iteratively configuring an FPGA with many designs and running each design at speed. This conventional method does not provide comprehensive delay-fault testing, because it is virtually impossible to test every circuit design that could be conceivably implemented on each FPGA. Given the difficulties and deficiencies of manufacturer conducted testing, other conventional methods of delay fault testing have considered only the testing of a user's specific FPGA configuration.
Configuration specific testing is problematic for a number of reasons. By their very nature, configuration specific tests are not feasible for wide-scale use by FPGA manufacturers. Consequently, significant overhead costs are imposed on individual users. Development time for configuration specific tests may be significant and test execution requires expensive machinery. Even after users have developed and executed these tests, it may be difficult to distinguish between problems caused by manufacturing defects and problems caused by user configuration errors. Furthermore, testing only static configurations is insufficient for users employing FPGA's in adaptive computing systems that dynamically reconfigure FPGA's while the system is on-line and running. However, conventional methods are not comprehensive for online testing, because delay faults are just as likely to occur in currently unused portions of the operational system. This problem is particularly significant for users employing FPGA's in high-reliability and high-availability applications, such as telecommunication network routers, in which the FPGA hardware cannot be taken offline for testing, maintenance, or repair without considerable cost or inconvenience. Conventional testing methods leave much to be desired.
An improved method of efficiently testing FPGA's for delay faults is needed.
Embodiments of the present invention provide systems and methods for delay-fault testing FPGA's, applicable both for off-line manufacturing and system-level testing, as well as for on-line testing within the framework of the roving self-testing areas (STAR's) approach. In one method according to the present invention, two or more paths under test receive a test pattern approximately simultaneously. The two paths are substantially identical and thus should propagate the signal in approximately the same amount of time. An output response analyzer receives the signal from each of the paths and determines the interval between them. The output response analyzer next determines whether a delay fault has occurred based at least in part on the interval.
The interval may be determined in any of several ways. For example, in one embodiment, the output of the first of the signals results in the activation of an oscillator. The oscillator continues to oscillate until all signals have been propagated. By counting the number of oscillation cycles occurring during the oscillation, the interval is determined. An embodiment of the present invention is able to test both low-to-high and high-to-low transitions.
One system according to the present invention includes an input, at least two paths under test in communication with the input, and an output response analyzer in communication with the paths that is operable to determine an interval between the time a data signal passes through the first path under test and the second path under test. In one embodiment, the output response analyzer includes an oscillator and a counter. Various configurations of the oscillator are possible. For example, in one embodiment, the oscillator includes an NAND gate and an OR gate. The inputs of the two gates are connected to the end of each path under test. The outputs of the two gates are connected to a second NAND gate. The output of the second NAND gate is connected to a counter and to the input of the second NAND gate. This results in an oscillation after a state transition until the state transition has propagated through all of the paths.
One system according to the present invention provides complete delay-fault testing of all paths through look-up tables (LUT's). For example, in one embodiment each of the paths under test comprises at least one lookup table (LUT), and each LUT is configured to produce a transition when the input of the LUT changes to a specified target address. For example, in one embodiment, the LUT content of the target address may be set to 1 and the LUT contents of all others set to 0. In another embodiment, the LUT content of the target address may be set to 0 and the LUT contents of all others set to 1. In one embodiment, the paths contain only LUT's and do not include flip-flops. In one such embodiment, each LUT comprises k inputs and each of the first path under test and second path under test comprises consecutive groups of 2k pairs of LUT's, wherein each of the groups comprises the same configuration and each pair comprises a different target address.
An embodiment of the present invention provides many advantages over conventional BIST-based techniques and systems for delay-fault testing in FPGA's. For example, an embodiment of the present invention is independent of the system applications implemented on the FPGA, and it is applicable for both on-line testing and for off-line manufacturing and system-level testing. An embodiment of the present invention is based on BIST, it is comprehensive, and it can work with any low-cost ATE.
Further details and advantages of embodiments of the present invention are set forth below.
These and other features, aspects, and advantages of the present invention are better understood when the following Detailed Description is read with reference to the accompanying drawings, wherein:
Embodiments of the present invention provide systems and methods for BIST-based delay-fault testing of a field-programmable gate array (FPGA). In one embodiment of the present invention, a test generator, a plurality of paths under test, and an output response analyzer are configured on an FPGA. The output response analyzer includes a combination of logic gates that create an oscillation during the interval between when the first of the plurality of paths under test propagate a signal from the test generator and when the last of the plurality of paths under test propagate the signal. If the interval is greater than a predetermined minimum threshold, a fault has occurred. The threshold may be zero.
Referring now to the drawings in which like numerals indicate like elements throughout the several figures,
An FPGA 102 comprises a plurality of programmable logic blocks (PLB's), such as PLB 104. The PLB's are installed on a chip 106 and are programmed to perform digital logic. The PLB 104 comprises flip-flops and/or look up tables to perform computational logic. For a complex digital circuit, the chip 106 may comprise an array of PLB's, such as a ten by ten or fifty-by-fifty array.
The FPGA 102 is a two-dimensional array of PLB's, interfacing to its Input/Output (I/O) pins via programmable I/O cells 108. Communication among PLB's and I/O cells is done through a programmable interconnect network 110, consisting of wire segments that can be connected via program able switches referred to as configurable interconnect points (CIP's) and also known as programmable interconnect points (PIP's) 112. The PLB logic functions and the CIP's are controlled by writing the configuration RAM. Wire segments in the programmable interconnect network are bounded by these CIP's and are considered to be either global or local routing resources. Global routing resources connect non-adjacent PLB's, while local routing resources connect a PLB to global routing resources or to adjacent PLB's. The routing resources are bus-oriented, with the number of wires per bus typically ranging between 4 and 8.
The outputs of the paths under test 504, 506 are connected to an output response analyzer (ORA) 508. When the ORA 508 receives a first signal from the paths under test 504, 506 to which it is connected, the ORA 508 starts measuring the timer interval until the last signal propagating along the paths under test 504, 506 arrives at the ORA 508. For example, in one embodiment, the ORA 508 begins an oscillation upon receiving the first signal and counts the number of oscillation cycles until the last of the paths under test 504, 506 propagates the signal from the test generator 502.
In the embodiment shown, the test is initiated using a flip flop with its input tied to a logical 1. The flip-flop is cleared out prior to initiating the test, so that it begins from a known state. The first cycle of the test causes a 0 to 1 (low to high) transition. In another embodiment, the input is tied to a logical 0, resulting hi a 1 to 0 (high to low) transition at the first clock cycle. Delay faults may be sensitive to the type of transition propagated through the paths under test; therefore, a tester may run both types of transitions through the same array of paths. After initiation, the test executes in a matter of nano-seconds 608.
Upon completion of the test 610, the tester reads the counter to determine the delay that occurred between the first of the paths under test to propagate the signal from the test generator and the last of the paths to propagate the signal 612. The counter may be read in various ways. In one embodiment, the tester reads the contents of the configuration memory corresponding to the counter value. In another embodiment, the counter functions as a shift register, and the tester shifts the contents of the register out through a boundary scan interface.
The tester utilizes the data in the counter to determine if a delay fault has occurred 614. An interval between when the first of the signals and the last of the signals propagated may or may not indicate a delay fault. For example, in one embodiment, the tester sets a minimum threshold. Unless the interval exceeds that threshold, no fault has occurred.
Assume that a rising transition is applied at the common input 1702. This transition propagates along every PUT 704, 706, and it will eventually appear at the inputs of the OR 708 and the NAND 710 gates. The signal FIRST responds to the fastest arriving transition, while LAST changes only after the slowest one has arrived. FIRST enables a local oscillator loop, and LAST stops the oscillations. Thus the count of oscillation pulses measures the difference D between the fastest and the slowest propagation delays along the PUT's 704, 706. In a circuit free of delay-faults, D should be smaller than a predetermined threshold; otherwise a delay fault is detected. The value of the threshold may be relative to the technology used to implement the FPGA since the FPGA will determine the rate of the oscillator. The threshold may be large (e.g., over 10 counts) or small (e.g., on the order of 3 or 4 counts) for indicating a delay fault. For example, in one embodiment, any value D smaller than 5% of the expected delay along the path is correct. Note that the same circuit can detect a delay-fault affecting the propagation of a 1/0 transition, the only difference being that the roles of FIRST and LAST are reversed. Since the first oscillation pulse may be generated (possibly as a partial pulse) even when the transitions of FIRST and LAST are very close, a count of one may not be interpreted as indicating a delay-fault.
It is interesting to observe that, unlike application-specific integrated circuit (ASIC) delay-fault testing, this technique does not involve clocking using the system clock. As a result, the clock distribution network in the FPGA is not tested for delay-faults using this technique. This is not a problem, since delays on the clock distribution paths are implicitly checked during speed-binning tests. Thus the delay-fault BIST described herein should be done in addition to, and not as a replacement of speed binning.
In any addressing/multiplexing mechanism with k address bits, there are 2k paths from every address input to the output, and each one of the 2k input combinations sensitizes a different set of k paths. As described so far, a method according to the present invention applies only the all-0 and all-1 input vectors. A complete delay-fault test of the LUT applies every possible address i to the LUT inputs, with the LUT programmed to produce a transition when the inputs change to the target address i; for every target address, the LUT should generate once a 0/1 and once a 1/0 transition.
In one embodiment of the present invention, a simple method of generating a 0/1 transition is utilized: programming a 1 at the address i and 0 at all other addresses. Then every input change from any other address to i will create a 0/1 spike-free output transition, occurring in response to the slowest input-output propagation through the LUT. Similarly, programming the LUT with a 0 at the target address and 1 elsewhere will generate a spike-free 1/0 transition.
Other modes of operation of a PLB, such as an adder, may involve dedicated logic and dedicated interconnect resources whose delays can be tested only when the PLB is configured for these operations.
In
The same configuration may be used to test the propagation of a falling transition (
In one embodiment of the present invention, the delay-fault BIST circuitry is simple: the TPG generates the two transitions, and the output response analyzer (ORA) consists of the three gates that produce the oscillation and the counter. The counter is reset before each experiment. Both the TPG and the counter can be initialized, and the ORA counter results can be read, via the FPGA boundary-scan access mechanism; this is the preferable method for on-line testing. Alternatively, for off-line testing, the ORA counter results can be read via configuration memory readback with the TPG and counter initialized via a global reset following download of the BIST configuration. The smallest difference between the delay of the fastest and slowest PUT's detectable with our scheme corresponds to one oscillation (OSC) cycle. When testing a path with ASIC-type delay-fault testing, the smallest detectable delay-fault is generally about 5% of the path delay. To achieve a similar feature, PUT's are constructed so that their total propagation delay corresponds to at least 20 OSC cycles. While making PUT's as long as possible would increase the number of FPGA resources concurrently tested, and possibly reduce the total number of BIST configurations required for a complete delay-fault test, it may also cause false negative results. For example, assume a path P1 where all of its components (PLB's, CIP's, and wire segments) are just 1% slower than their counterparts on path P2. If the PUT's involve a large number of components, the accumulated difference between the delays of P1 and P2 may be incorrectly reported as a delay-fault. Therefore, PUT's should be constructed so that their delay is not significantly larger than that of an average path that would be used in “normal” system circuits implemented in the FPGA while, at the same time being large enough to obtain the desired delay-fault detection resolution (for example, the 20 OSC cycles described above). In any comparison-based BIST approach, a passing result may be produced when the compared elements are all faulty, e.g., when all the compared PUT's are equally slow. Such a situation is unlikely when we compare several (4 to 8) paths. However, if desired, a validation test to protect against this case can be easily done by selecting one of the paths that passed the test and comparing it with a new path, which was not part of the compared group that passed the initial test.
No delay-faults will be detected in a slow device where all paths are equally slow. This is the correct result, and such a chip will be identified by speed binning and may be allowed to work as a lower speed-grade device. The approach described herein may fail if a PUT has compensating delay-faults, where the detection of a slow path segment is masked by the presence of a fast segment, so that the overall path delay remains about the same as the other PUT's. In general, however, most delay-faults slow down the circuit, and such a multiple fault is unlikely to occur in practice. Accordingly, if each resource is included that can contribute to a delay-fault in one of the PUT's, and each PUT is tested for both rising and falling transitions, the delay-fault test is complete, that is, it will detect any delay-fault that creates a meaningful difference between compared PUT's. Path selection in an embodiment of the present invention follows the scheme used in the interconnect testing approach detailed in C. Stroud, S. Wijesuriya, C. Hamilton, and M. Abramovici, “Built-In Self-Test of FPGA Interconnect,” Proc. Intn'l. Test Conf., pp. 404-411: 1998, and guarantees that every resource is included (at least once) in a PUT. Hence there is no need to need to compute the resulting delay-fault coverage. The use of the local oscillator created from the inverting feedback in the PLB logic could give rise to concerns of the quality of the clock feeding the ORA counter, specifically, the duty cycle and period needed for proper operation of the counter. One solution to this problem is to configure a single flip-flop as a toggle flip-flop with the output of the local oscillator driving the clock input to this flip-flop, and the output of the toggle flip-flop driving the clock input of the ORA counter. This effectively divides the local oscillator frequency by 2 and ensures a near 50% duty cycle to the ORA counter. The lower frequency clock will only reduce the resolution of delay-fault detection as opposed to preventing this delay-fault built-in self test (BIST) approach from working. However, the delay-fault BIST approach has been implemented in an ORCA 2C15A FPGA and found the oscillator clock to run at 243 MHz while producing a duty cycle and clock waveform of sufficient quality to obtain reproducible results from one execution of the delay-fault BIST sequence to the next. Therefore, dividing the clock may not be necessary.
Various techniques for testing the entire FPGA may be utilized. For example, in one embodiment, a roving self-test area (STAR) is utilized for on-line FPGA testing, diagnosis, and fault tolerance. Such a method is applicable to any FPGA supporting incremental run-time reconfiguration (RTR) via its boundary-scan interface. A STAR is a temporarily off-line section of the FPGA where self-testing occurs without disturbing the normal system activity in the rest of the chip. Roving the STARs periodically brings every section of the FPGA under test. This approach guarantees complete testing of the FPGA, including all its spare resources, and does not require any part of the chip to be fault-free.
Testing for delay-faults follows the pattern of interconnect testing in an on-line routing BIST, where horizontal and vertical routing resources are tested in H-STAR and V-STAR, respectively. Testing for delay-faults takes place after completing the test for logic and interconnect resources within the STAR.
One way of characterizing the difference between on-line and off-line (manufacturing or system-level) testing is that no system function exists during off-line testing. Hence for off-line testing, the entire FPGA can be populated with a “galaxy” of parallel STARs (either vertical or horizontal), all executing concurrently the same delay-fault BIST configurations (
Another embodiment of the present invention is implemented in a Xilinx Spartan series FPGA.
Our method is based on BIST, it is comprehensive, and does not require expensive ATE. We have successfully implemented this BIST approach on the ORCA 2C and Xilinx Spartan FPGA's and have verified that the approach is not only feasible but is also practical. We have emulated many delay-faults by creating a “faulty” PUT longer than the other “fault-free” PUT's, (the longer PUT is routed through additional wire segments, CIP's, and PLB's). In all cases, methods according to the present invention successfully detected all emulated delay-faults. The current diagnostic resolution for delay-faults detected using this approach is to a STAR.
The foregoing description of the preferred embodiments of the invention has been presented only for the purpose of illustration and description and is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Numerous modifications and adaptations thereof will be apparent to those skilled in the art without departing from the spirit and scope of the present invention.
This application claims priority to U.S. Provisional Patent application Ser. No. 60/392,543, filed Jul. 1, 2002, the entirety of which is hereby incorporated by reference.
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/US03/20705 | 7/1/2003 | WO | 00 | 3/24/2005 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2004/003582 | 1/8/2004 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
5675502 | Cox | Oct 1997 | A |
5761097 | Palermo | Jun 1998 | A |
5991907 | Stroud et al. | Nov 1999 | A |
6003150 | Stroud et al. | Dec 1999 | A |
6108806 | Abramovici et al. | Aug 2000 | A |
6134191 | Alfke | Oct 2000 | A |
6202182 | Abramovici et al. | Mar 2001 | B1 |
6256758 | Abramovici et al. | Jul 2001 | B1 |
6466520 | Speyer et al. | Oct 2002 | B1 |
6530049 | Abramovici et al. | Mar 2003 | B1 |
6550030 | Abramovici et al. | Apr 2003 | B1 |
6574761 | Abramovici et al. | Jun 2003 | B1 |
6631487 | Abramovici et al. | Oct 2003 | B1 |
6725442 | Cote et al. | Apr 2004 | B1 |
6996020 | Yoshida | Feb 2005 | B2 |
6874108 | Abramovici et al. | Mar 2005 | B1 |
6973608 | Abramovic et al. | Dec 2005 | B1 |
Number | Date | Country | |
---|---|---|---|
20050154552 A1 | Jul 2005 | US |
Number | Date | Country | |
---|---|---|---|
60392543 | Jul 2002 | US |