E&CE 720-t4: Asn-2: Pipelines and Performance
Overview
Due: Wed Mar 12 12:30pm.
You may turn in neatly written/drawn answers, there is no need to
type up your answers (although typed anwers are fine).
Purpose:
Review concepts on performance, pipelines, and superscalar processors
from chapters 1,2, 4, 5.
Tasks
Answer each of the following questions. The bold numbers are
questions from the textbook. If you have done the readings
carefully, most questions should take no more than about an hour to
answer. For the questions that ask you to design hardware, provide
a sketch of the hardware and an explanation of how the hardware
works.
- 1.2: SPEC performance vs frequency
- 1.6,7,8,9,10: Frequency vs performance for multiply and
shift/add.
- 2.9,10: Load-update instruction
- 2.13: Caches and pipelining
- 4.4: I-cache and superscalar fetch
- 4.13,14: IPC and superscalar
- For a typical 4-wide superscalar microprocessor, how much will
performance increase if branch-prediction accuracy increases from
90% to 95%? This is a very open-ended question. There are two
important aspects: making realistic assumptions/generalizations
and finding an approach that makes a good tradeoff between the
accuracy of your results and the effort taken to get the result.
I'm looking for a conceptual analysis, not tables and graphs.
- In register renaming, the conventional approach is for
instructions sitting in reservation stations to read each
operand as soon as it becomes valid. This question explores an
alternative approach, where instructions wait for both operands to
become valid and then read both operands at the same time, as the
instruction exits the reservation station.
- Describe the differences in the hardware between the
conventional approach and the alternative.
- For each of the three types of data hazards, describe how
you would guarantee data-hazard correctness in the alternative
approach.
- Analyze the tradeoffs (area, performance, etc) between the
conventional approach and the alternative approach. Will one
of the approaches always be better, or are there some
situations/processors where the conventional approach is
better and others where the alternative approach is
better?
Marking
The focus of the marking will be on your ability to apply the
concepts from the textbook and lecture to answer the questions.
Almost all of the questions have multiple correct answers.