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:20260202T201257Z
LOCATION:Second Floor Atrium
DTSTART;TZID=America/Chicago:20251121T080000
DTEND;TZID=America/Chicago:20251121T120000
UID:submissions.supercomputing.org_SC25_sess620_post166@linklings.com
SUMMARY:Forward Error Bounds and Efficient Algorithms for Computing a Tens
 or Times Matrix Chain in Low Precision on GPUs
DESCRIPTION:Julian Bellavita (Cornell University, Oak Ridge National Labor
 atory (ORNL)) and Piyush Sao and Ramakrishnan Kannan (Oak Ridge National L
 aboratory (ORNL))\n\nMany tensor processing algorithms require computing a
  tensor times matrix chain (TTMc) operation, and this operation is frequen
 tly the bottleneck in such algorithms. This work develops strategies for a
 ccelerating a TTMc using low-precision hardware. \n\nWe present a novel sc
 heme for scaling the TTMc operands to prevent overflow. Our scheme exploit
 s the Kronecker Product structure of a TTMc to allow for efficient applica
 tion. Additionally, we present the first forward error bound for TTMc, and
  we develop a heuristic for ordering the individual TTM operations within 
 a TTMc to reduce the forward error. \n\nOur scaling scheme allows for a TT
 Mc on the Miranda Tensor to be computed without overflow on an NVIDIA A100
  GPU using FP16 arithmetic, exhibiting a speedup of up to 2× over FP64 ari
 thmetic, even when accounting for the overhead of applying scaling. We sho
 w that our TTM ordering heuristic is effective for some tensors in certain
  cases.\n\nTag: Research & ACM SRC Posters\n\nRegistration Category: Techn
 ical Program Reg Pass\n\nSession Chairs: Kento Sato (RIKEN Center for Comp
 utational Science (R-CCS)); Anja Gerbes (Georg-August-Universität Göttinge
 n); and Chris Schlipalius (Pawsey Supercomputing Research Centre; Commonwe
 alth Scientific and Industrial Research Organisation (CSIRO), Australia)\n
 \n
END:VEVENT
END:VCALENDAR