Computer Laboratory

Course material 2010–11

Comparative Architectures

Principal lecturer: Dr Robert Mullins
Taken by: Part II
Syllabus
Past exam questions
Information for supervisors (contact lecturer for access permission)

Lecture Notes

Complete set of lecture notes (PDF) (or each lecture 1up)
Supervision Worksheet (PDF)

Reading

Lectures 1-3: I suggest you read the first Chapter of Hennessy and Patterson. Appendix B. may also be useful (this is Chapter 2 in the 3rd Edition).

Lectures 4-7: Chapter 2 of Hennessy and Patterson. Chapter 3. will probably also be useful.

Lecture 7: Appendex G in Hennessy and Patterson (this is covered in Chapter 4 in the 3rd Edition)

Lecture 8: Finish reading Chapter 3.

Lecture 9 and 10: Appendix C and Chapter 5.

Lecture 11: If you need further explanations read Appendix F (on CD).

Lecture 12 and 13: Chapter 4.

Lecture 14: "Principles and Practices of Interconnection Networks", Dally/Towles is an excellent book if you want to go into more detail or clarify any particular areas.

Additional Material

I have provided additional material for each lecture on the course wiki to help you go beyond the basic material. These papers and reports may clarify elements of the course or simply help you explore aspects that you find particularly interesting. It is certainly not a requirement for you to read and understand the additional material.

Link to course wiki. Older additional material can be found here.

Branch Prediction Competition

See competition wiki page

Errata

Lecture 6, Slide 17. "If the count...." should read "If the count is non-zero when the STORE commits a LOAD may have received a stale value. In this case we flush and restart the pipeline to ensure the load receives the correct data."

Lecture 8, Slide 31. This figure needs some explanation and the current text beneath it is confusing. It should read "Each square describes the performance of the row benchmark when sharing the processor with the column benchmark. Performance is consider poor (red) when it is less than 0.5X the benchmark's performance when running alone."