UNIVERSITY OF HERTFORDSHIRE
COMPUTER SCIENCE RESEARCH COLLOQUIUM
presents
"Using Neural Networks for Branch Prediction"
speaker:
Steve Hunt
Computer Science, University of Hertfordshire
25 February 2004 (Wednesday)
Lecture Theatre LC108
Hatfield College Lane Campus
3:00 - 4:00 pm
Refreshments and biscuits will be available.
Everyone is Welcome to Attend
Abstract:
---------
Modern computer architectures employ techniques such as pipelining to
boost throughput. In principle, significant performance gains can be
achieved by increasing pipeline depth; however, in practice there are
penalties for adopting such a strategy.
Where a program contains a conditional branch instruction (e.g. at
the head of a 'for' loop), it is necessary to know whether or not the
condition is satisfied in order to determine which instruction should
be fetched next. The problem is that the necessary information is not
available early enough in the fetch-execute cycle to ensure that the
right instruction can be fetched in time to fill the next slot in the
pipeline. If the slot is left empty then the processor 'loses' a
clock cycle (because there is nothing for it to do), and every lost
cycle leads to a reduction in throughput. One of the most common
approaches to this problem is to attempt to _predict_ whether or not
the branch will be taken.
In his talk Steve will be examining the application of neural
networks to the problem of dynamic branch prediction in
high-performance processors. He will start off by explaining the
branch prediction problem and why it is so important to find a good
solution. He will then say a little about conventional solutions to
the problem, before he goes on to explain how neural networks might
be applied in this area. He will finish up by showing the results of
some experiments he has been running on applying simple threshold
logic units to the branch prediction problem, and by comparing such
systems with more complex neural networks and with conventional
predictors.
------------------------------------------------------------------
Computer Science Colloquium Abstracts On-line:
http://homepages.feis.herts.ac.uk/~nehaniv/colloq/