A defect-tolerant computer architecture: Opportunities for nanotechnology

Science

Volumn 280, Issue 5370, 1998, Pages 1716-1721

  Heath, James R ^a   Kuekes, Philip J ^b   Snider, Gregory S ^b   Williams, R Stanley ^b  

^a UNIVERSITY OF CALIFORNIA   (United States)

^b HEWLETT PACKARD LABORATORIES   (United States)

Author keywords

[No Author keywords available]

Indexed keywords

METAL; OXIDE; SILICON;

ALGORITHM; COMPUTER; COMPUTER PROGRAM; PRIORITY JOURNAL; REVIEW; SEMICONDUCTOR;

EID: 0032510985     PISSN: 00368075     EISSN: None     Source Type: Journal    
DOI: 10.1126/science.280.5370.1716     Document Type: Review

References (28)

1. The 1997 National Technology Roadmap for Semiconductors is a publication of the Semiconductor Industry Association (www.semichips.org). The Roadmap costs $250. Distribution is being coordinated by SEMATECH, based in Austin, Texas. Copies may be obtained on-line by contacting SEMATECH's home page (www.sematech.org).
2. R. R. Schaller, IEEE Spectrum 34, 53 (June 1997).
3. R. Landauer, IBM J. Res. Dev. 3, 183 (1961).
4. An especially good introduction to information science for physical scientists is R. P. Feynman, Lectures in Computation, A. J. G. Hey and R. W. Allen, Eds. (Addison-Wesley, Menlo Park, CA, 1996).
5. R. H. Chen, A. N. Korotov, K. K. Likharev, Appl. Phys. Lett. 68, 1954 (1996).
6. P. D. Tougaw and C. S. Lent, J. Appl. Phys. 75, 181 (1994).
7. W. A. Caldwell et al., Science 277, 93 (1997).
8. C. Mead, Proc. IEEE 78, 1629 (1990).
9. J. J. Hopfield and D. W. Tank, Science 233, 625 (1986).
10. A. Aviram and M. Ratner, Chem. Phys. Lett. 29, 277 (1974).
11. M. C. Petty, M. R. Bryce, D. Bloor, Eds., Introduction to Molecular Electronics, (Edward Arnold, London, 1995).
12. The assembly of the machine is chemical but the operation of the machine is electronic. This is to be distinguished from the DNA computers and others in which computation proceeds by way of chemical reactions, as discussed in L. M. Adelman [Science 266, 1021 (1994)].
13. W. B. Culbertson, R. Amerson, R. J. Carter, G. S. Snider, P. J. Kuekes, Proc. IEEE Symp. FPGA's for Custom Computing Machines (1997).
14. We distinguish between "defect tolerant," the capacity of a machine to operate perfectly in the presence of errors made in the hardware during manufacture, from "fault tolerant," the ability of a machine to recover from errors made during a calculation. Fault tolerance is an active research area and will also be required for computational nanotechnology. However, we do not discuss fault tolerance in this article.
15. The Teramac approach is actually more conservative than other massively parallel defect tolerant architectures that have been proposed. Once the defective components have been found, we propose to construct a fairly conventional computer with a standard instruction set out of the good components. Others propose to use cellular automata or neural nets to achieve robustness in the presence of defects. These approaches depend on significant progress in algorithm development through use of cellular automata or neural nets (6-9).
16. An excellent recent discussion of configurable computing can be found in J. Villasenor and W. Mangione-Smith, Sci. Am. 276, 68 (June 1997).
17. C. E. Leiserson, IEEE Trans. Comput. C34, 892 (1985).
18. L. R. Goke and G. L. Lipovski, in Proc. First Annual Symposium On Computer Architecture (IEEE, New York, 1973), pp. 21-28.
19. HP scientists chose a fat-tree architecture with larger than necessary Rent's rule exponents to enable fast compilation.
20. Rent's rule was discovered by Richard Rent at IBM in the 1960s. Although it was never published by Rent, subsequent work has confirmed its validity [B. S. Landman and R. L. Russo, IEEE Trans. Comp. C20, 1469 (1971)].
21. This small percentage of area devoted to logic holds true for all VLSI chips built today.
22. U. Kanus et al., 11th Eurographics Workshop on Computer Graphics Hardware, B.-O. Schneider and A. Schilling, Eds. (IBM, New York, 1996).
23. This version of Teramac was running a DNA string matching algorithm. For further detail on this algorithm, see R. J. Lipton and D. P. Lopresti, in 1985 Chapel Hill Conference on VLSI, H. Guchs, Ed. (Computer Science Press, Maryland, 1985), pp. 363-376.
24. N. J. Howard, A. M. Tyrrell, N. M. Allinson, in IEEE Transactions on VLSI Systems 2, 115 (1994).
25. J. Otterstedt, H.-J. Iden, M. Kuboshek, Proceedings of the 1994 International Conference on Wafer Scale Integration, R. M. Lea and S. Tewksbury, Eds. (IEEE, New York, 1994), pp. 315-323.
26. We estimate that Teramac could have contained up to 50% defective wires and still have operated successfully, although the question of how many defects can be tolerated in a Teramac-like system is a subject of investigation. One critical aspect of the cross-bar architecture is that if switches fail in the "open" state, then many more defects can be tolerated. Switches that fail "closed," however, will render an entire row or column of associated gates inoperable.
27. B. Krishnamurthy, IEEE Trans. Comput. C-33, 438 (1984).
28. We thank W. Robinett for helpful comments on the manuscript. J.R.H. acknowledges support from an NSF-GOALI grant and the Hewlett Packard Corporation during the writing of this manuscript.