4-8 July 2016
Kramer Law building
Africa/Johannesburg timezone
<a href="http://events.saip.org.za/internalPage.py?pageId=10&confId=86">The Proceedings of SAIP2016</a> published on 24 December 2017

Optimising GPU Integration into the ATLAS Trigger

6 Jul 2016, 10:20
20m
4B (Kramer Law building)

4B

Kramer Law building

UCT Middle Campus Cape Town
Oral Presentation Track F - Applied Physics Applied Physics (1)

Speaker

Mr Marc Sacks (University of the Witwatersrand)

Apply to be<br> considered for a student <br> &nbsp; award (Yes / No)?

Yes

Abstract content <br> &nbsp; (Max 300 words)<br><a href="http://events.saip.org.za/getFile.py/access?resId=0&materialId=0&confId=34" target="_blank">Formatting &<br>Special chars</a>

General Purpose Graphics Processing Units (GPGPU) are suited to rapidly
performing independent (non-sequential) computations on large datasets and it seems likely
that it will be the workhorse for applications involving massive parallelism in the near future.
The ATLAS detector in the Large Hadron Collider is currently the subject of investigation
with regards to the use of GPGPU. The planned increase in detector luminosity will lead to
increased pile-up (a time-energy resolution artifact). Preliminary tests indicate a reduction in
trigger latency with the introduction of GPUs, implying they can be used to run more complex
algorithms in a similar or smaller amount of time than CPUs, thereby reducing pile-up associated
errors. These tests have been conducted on high-end server-grade GPUs for demonstrative
purposes. However when selecting a GPU platform for large-scale integration into the ATLAS
Trigger, performance/watt and performance/dollar are the parameters of interest. It is not
trivial to categorise the relative performance of GPUs because of the factors involved in its
determination i.e. FLOPS, global and local memory bandwidth and size, core count, etc. This
investigation focuses on the categorisation of GPUs representative of those most likely to be
integrated into the trigger system. Criteria include power consumption, cost, ease-of-use, and
reliability. Preliminary results indicate that lower-end, more energy-efficient GPUs could, in
this context, be used in place of high-end GPUs with similar results.

Main supervisor (name and email)<br>and his / her institution

Bruce Mellado, Bruce.Mellado.Garcia@cern.ch, University of the Witwatersrand

Level for award<br>&nbsp;(Hons, MSc, <br> &nbsp; PhD, N/A)?

MSc

Please indicate whether<br>this abstract may be<br>published online<br>(Yes / No)

Yes

Would you like to <br> submit a short paper <br> for the Conference <br> Proceedings (Yes / No)?

Yes

Primary author

Mr Marc Sacks (University of the Witwatersrand)

Co-author

Prof. Bruce Mellado (University of Wisconsin - Madison)

Presentation Materials

There are no materials yet.