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:20211207T054809Z
LOCATION:240-241-242
DTSTART;TZID=America/Chicago:20211118T103000
DTEND;TZID=America/Chicago:20211118T110000
UID:submissions.supercomputing.org_SC21_sess178_pap329@linklings.com
SUMMARY:Arithmetic-Intensity-Guided Fault Tolerance for Neural Network Inf
 erence on GPUs
DESCRIPTION:Paper\n\nArithmetic-Intensity-Guided Fault Tolerance for Neura
 l Network Inference on GPUs\n\nKosaian, Rashmi\n\nNeural networks (NNs) ar
 e increasingly employed in safety-critical domains and in environments pro
 ne to unreliability (e.g., soft errors), such as on spacecraft. Therefore,
  it is critical to impart fault tolerance to NN inference. Algorithm-based
  fault tolerance (ABFT) is emerging as an efficient approach for fault tol
 erance in NNs.\n\nWe propose an adaptive approach to ABFT for NN inference
  that exploits untapped opportunities in emerging deployment scenarios. GP
 Us have high compute-to-memory-bandwidth ratios, while NN layers have a wi
 de range of arithmetic intensities. This leaves some layers compute bound 
 and others memory-bandwidth bound, but current approaches to ABFT do not c
 onsider these differences. We first investigate ABFT schemes best suited f
 or each of these scenarios. We then propose intensity-guided ABFT, an adap
 tive, arithmetic-intensity-guided approach that selects the most efficient
  ABFT scheme for each NN layer. Intensity-guided ABFT reduces execution-ti
 me overhead by 1.09--5.3x across many NNs compared to traditional approach
 es to ABFT.\n\nTag: System Software and Runtime Systems\n\nRegistration Ca
 tegory: Tech Program Reg Pass
END:VEVENT
END:VCALENDAR