BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
X-LIC-LOCATION:Europe/Stockholm
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20210916T132453Z
LOCATION:Mère Royaume
DTSTART;TZID=Europe/Stockholm:20210708T173000
DTEND;TZID=Europe/Stockholm:20210708T180000
UID:submissions.pasc-conference.org_PASC21_sess200_msa351@linklings.com
SUMMARY:Performance Measurement Techniques with Mixed-Precision Solvers
DESCRIPTION:Minisymposium\n\nPerformance Measurement Techniques with Mixed
 -Precision Solvers\n\nLuszczek\n\nBy adjusting the accuracy of the operati
 ons in a parallel solver it is possible<br />to drastically increase the o
 verall performance without much reduction in the<br />quality of the solut
 ion.  Regulating accuracy may come in form of mixed<br />precision or
  approximations that take advantage of the numerical structure of<br />the
  data and possibly augmenting it with the information from the data's<br /
 >origin.  When used in the context of modern parallel hardware, these
  methods<br />allow to take full advantage of the available compute resour
 ces and result in a<br />substantial increase of both: the runtime perform
 ance as well as the optimality<br />of the communication.  We also pr
 esent techniques of generating data for<br />testing this new breed of mix
 ed precision solvers, allowing the users to test<br />those solvers in a s
 calable manner. Our techniques focus on mixed precision<br />hardware and 
 software where both the solver and the hardware can take advantage<br />of
  mixing multiple floating precision formats. This enables the use of recen
 tly<br />generations of hardware platforms that focus on ML and DNN worklo
 ads but can<br />also be utilized for HPC applications as long as the new 
 algorithms are<br />combined with the custom floating-point formats. Such 
 a mixture of algorithmic<br />and data storage improvements delivers perfo
 rmance levels far beyond the<br />standard IEEE data types while still ach
 ieving comparable accuracy of the<br />results.\n\nDomain: CS and Math, Em
 erging Applications, Physics, Engineering
END:VEVENT
END:VCALENDAR