CMSC CMSC818P (Spring 2013)

Exascale Computation

Dr. Jeff Hollingsworth
4155 AV Williams
301 405-2708
Office Hours:
Tu/Th 10:45-11:30 (4155 AVW)

This page contains information relating to the Course CMSC818P on Exascale computing offered during the Spring 2013 semester.


  • Final project presentations Tuesday & Thursday May 7 & 9 in class
  • Final Exam: Monday May 13, 2013 (8-10AM) - CSI 3120
  • Final Project Reports are due Tuesday May 14 (5:00 PM) - email me the PDF
  • Academic Integrity:

    "Integrity without knowledge is weak and useless, and knowledge without integrity is dangerous and dreadful." Samuel Johnson (The History of Rasselas, ch. 41 - 1759). Please read the statement on academic integrity.

    Course Overview

    The next big challenge in high-end computing is exascale computation. Exascale computation requires that a single computer perform 10^18 operations per second (often an operation means a floating point arithmetic operation, but can also mean integer operations or data movement operations). This level of performance is nearly 100 times the largest computer in the world today. In addition to the goal of pure performance, the exascale effort in the United States requires that such a computer not use more than 20 Megawatt/hr of electricity (the world's largest computer currently uses about 8 Megawatts/hr).

    Meeting the goals of computing at this level may require major changes in how computers work. This course will look at those challenges and the potential changes to computers that might be required to meet them. We will also briefly review the path that led to peta-scale computing (10^15 operations/sec) and the claims that we would need radically different computers to get there (we didn't). What is or is not different about exascale?

    Topics to be covered include:

  • Power and Energy Optimization including using dark silicon
  • Reduced hardware reliability to improve power utilization
  • Novel Architectures including streaming processors
  • Advanced compiler techniques for heterogeneous architectures
  • Novel approaches to parallelism to reduce communication and synchronization
  • Use of mixed precision algotihms to enhance performance and power utilization
  • Tools to optmization performance and energy consumption
  • Course Structure

    This class will mostly be offered in two forms. A two credit seminar, will only involve reading papers and participating in discussion. There is also a three credit version that includes a relatively small term project and a final. The two credit class will not count for comp credit. However, the three credit version will count for both M.S. and Ph.D. comp credit.

    Additional Information

  • Syllabus
  • Reading List (To access some of the papers you must be on the UMD campus)
  • Parallel Computing 101 Slides