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:20260522T150118Z
LOCATION:C140/142
DTSTART;TZID=America/Chicago:20181115T140000
DTEND;TZID=America/Chicago:20181115T143000
UID:submissions.supercomputing.org_SC18_sess190_pap429@linklings.com
SUMMARY:CosmoFlow: Using Deep Learning to Learn the Universe at Scale
DESCRIPTION:Amrita Mathuriya (Intel Corporation); Deborah Bard (National E
 nergy Research Scientific Computing Center (NERSC), Lawrence Berkeley Nati
 onal Laboratory); Pete Mendygral (Cray Inc); Lawrence Meadows (Intel Corpo
 ration); James Arnemann (University of California, Berkeley); Lei Shao (In
 tel Corporation); Siyu He (Carnegie Mellon University); Tuomas Karna (Inte
 l Corporation); Diana Moise (Cray Inc); Simon J. Pennycook (Intel Corporat
 ion); Kristyn Maschhoff (Cray Inc); Jason Sewall and Nalini Kumar (Intel C
 orporation); Shirley Ho (Lawrence Berkeley National Laboratory, Carnegie M
 ellon University); Michael F. Ringenburg (Cray Inc); Mr Prabhat (Lawrence 
 Berkeley National Laboratory, National Energy Research Scientific Computin
 g Center (NERSC)); and Victor Lee (Intel Corporation)\n\nDeep learning is 
 a promising tool to determine the physical model that describes our univer
 se.   To handle the considerable computational cost of this problem, we pr
 esent CosmoFlow: a highly scalable deep learning application built on top 
 of the TensorFlow framework.\n\nCosmoFlow uses efficient implementations o
 f 3D convolution and pooling primitives, together with improvements in thr
 eading for many element-wise operations, to improve training performance o
 n Intel Xeon Phi processors.  We also utilize the Cray PE Machine Learning
  Plugin for efficient scaling to multiple nodes. We demonstrate fully sync
 hronous data-parallel training on 8192 nodes of Cori with 77% parallel eff
 iciency, achieving 3.5 Pflop/s sustained performance. \n\nTo our knowledge
 , this is the first large-scale science application of the TensorFlow fram
 ework at supercomputer scale with fully-synchronous training. These enhanc
 ements enable us to process large 3D dark matter distribution and predict 
 the cosmological parameters Omega_M, sigma_8 and N_s with unprecedented ac
 curacy.\n\nTag: Applications, Cosmology, Data Analytics, Deep Learning, Ma
 chine Learning, Programming Systems, Storage, Visualization\n\nRegistratio
 n Category: Tech Program Reg Pass\n\nSession Chair: Tal Ben-Nun (Lawrence 
 Livermore National Laboratory (LLNL))\n\n
END:VEVENT
END:VCALENDAR