Course pages 2012–13
Lecturer: Dr R.D. Mullins
No. of lectures: 16
Suggested hours of supervisions: 4
Prerequisite course: Computer Design
This course examines the techniques and underlying principles that are used to design high-performance computers and processors. Particular emphasis is placed on understanding the trade-offs involved when making design decisions at the architectural level. A range of processor architectures are explored and contrasted. In each case we examine their merits and limitations and how ultimately the ability to scale performance is restricted.
- Introduction. The impact of technology scaling and market trends.
- Fundamentals of Computer Design. Amdahl’s law, energy/performance trade-offs, ISA design.
- Advanced pipelining. Pipeline hazards; exceptions; optimal pipeline depth; branch prediction; the branch target buffer [2 lectures]
- Superscalar techniques. Instruction-Level Parallelism (ILP); superscalar processor architecture [2 lectures]
- Software approaches to exploiting ILP. VLIW architectures; local and global instruction scheduling techniques; predicated instructions and support for speculative compiler optimisations.
- Multithreaded processors. Coarse-grained, fine-grained, simultaneous multithreading
- The memory hierarchy. Caches; programming for caches; prefetching [2 lectures]
- Vector processors. Vector machines; short vector/SIMD instruction set extensions; stream processing
- Chip multiprocessors. The communication model; memory consistency models; false sharing; multiprocessor memory hierarchies; cache coherence protocols; synchronization [2 lectures]
- On-chip interconnection networks. Bus-based interconnects; on-chip packet switched networks
- Special-purpose architectures. Converging approaches to computer design
At the end of the course students should
- understand what determines processor design goals;
- appreciate what constrains the design process and how architectural trade-offs are made within these constraints;
- be able to describe the architecture and operation of pipelined and superscalar processors, including techniques such as branch prediction, register renaming and out-of-order execution;
- have an understanding of vector, multithreaded and multi-core processor architectures;
- for the architectures discussed, understand what ultimately limits their performance and application domain.
* Hennessy, J. & Patterson, D. (2006). Computer architecture: a quantitative approach. Elsevier (4th ed.) ISBN 978-0-12-370490-0. (3rd edition is also good)