BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Chicago
X-LIC-LOCATION:America/Chicago
BEGIN:DAYLIGHT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
TZNAME:CDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
TZNAME:CST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20211207T055407Z
LOCATION:Online
DTSTART;TZID=America/Chicago:20211115T152500
DTEND;TZID=America/Chicago:20211115T155500
UID:submissions.supercomputing.org_SC21_sess423_ws_mlhpce102@linklings.com
SUMMARY:Semantic-Aware Lossless Data Compression for Deep Learning Recomme
 ndation Model (DLRM)
DESCRIPTION:Workshop\n\nSemantic-Aware Lossless Data Compression for Deep 
 Learning Recommendation Model (DLRM)\n\nPumma, Vishnu\n\nDeep Learning Rec
 ommendation Model (DLRM), a new neural network for recommendation systems,
  introduces challenging requirements for deep neural network training and 
 inference.  The size of the DLRM model is typically large and not able to 
 fit on a single GPU memory.  DLRM requires both model-parallel and data-pa
 rallel for the bottom part and top part of the model when running on multi
 ple GPUs.  Due to the hybrid-parallel model, the all-to-all communication 
 is used for welding the top and bottom parts together.  We have observed t
 hat the all-to-all communication is costly and is a bottleneck in the DLRM
  training/inference.\n\nIn this presentation, we reduce the communication 
 volume by using DLRM's properties to compress the transferred data without
  information loss.  We demonstrate benefits of our method by training DLRM
  TeraByte on AMD Instinct MI100 accelerators.  The experimental results sh
 ow 38%-59% improvement in the time-to-solution of the DLRM TeraByte traini
 ng for FP32 and mixed-precision.\n\nTag: Online Only, Machine Learning and
  Artificial Intelligence\n\nRegistration Category: Workshop Reg Pass
END:VEVENT
END:VCALENDAR