Architectural Support for Approximate Computations

Overview:

Many programs compute on approximate and/or probabilistic data that are highly tolerant of error. Such computations can tolerate significant corruption and still produce meaningful and useful results. Multimedia computations exhibit such properties. In addition, artificial intelligence computations, an increasingly important application domain, also exhibit such properties. In this project, we characterize the error resilience properties of these computations, and investigate architectural techniques for exploiting them. Techniques of interest include mechanisms for increased resilience to soft errors, statistically correct architectures for low power, and runtime systems that tradeoff solution quality for improved performance and/or real-time guarantees.

People:

Faculty

  • Donald Yeung

    • Students

  • Hameed Badawy
  • Xuanhua Li
  • Wanli Liu
  • Meng-Ju Wu
  • Xu Yang

    • Publications:

  • Xuanhua Li and Donald Yeung. Application-Level Correctness and its Impact on Fault Tolerance. In Proceedings of The 13th International Symposium on High-Performance Computer Architecture (HPCA-XIII). Phoenix, AZ. February 2007. (pdf, postscript)
  • Xuanhua Li and Donald Yeung. Exploiting Soft Computing for Increased Fault Tolerance. Appears in Proceedings of the 2006 Workshop on Architectural Support for Gigascale Integration(ASGI'06). co-held with the ISCA-33. Boston, MA. June 2006. (pdf, postscript)

    • Funding:

    This project is funded by the Defense Advanced Research Projects Agency (DARPA) through the Department of the Interior National Business Center under grant #NBCH104009.

    Last updated: September 2006 by Donald Yeung (yeung@eng.umd.edu)