BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Denver
X-LIC-LOCATION:America/Denver
BEGIN:DAYLIGHT
TZOFFSETFROM:-0700
TZOFFSETTO:-0600
TZNAME:MDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0600
TZOFFSETTO:-0700
TZNAME:MST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260422T000713Z
LOCATION:403-404
DTSTART;TZID=America/Denver:20231116T113000
DTEND;TZID=America/Denver:20231116T120000
UID:submissions.supercomputing.org_SC23_sess178_pap438@linklings.com
SUMMARY:VENOM: A Vectorized N:M Format for Unleashing the Power of Sparse 
 Tensor Cores
DESCRIPTION:Roberto L. Castro (Universidade da Coruña), Andrei Ivanov (ETH
  Zürich), Diego Andrade (Universidade da Coruña), Tal Ben-Nun (ETH Zürich)
 , Basilio B. Fraguela (Universidade da Coruña), and Torsten Hoefler (ETH Z
 ürich)\n\nThe increasing success and scaling of Deep Learning models deman
 ds higher computational efficiency and power. Sparsification can lead to b
 oth smaller models as well as higher compute efficiency, and accelerated h
 ardware is becoming available. However, exploiting it efficiently requires
  kernel implementations, pruning algorithms, and storage formats, to utili
 ze hardware support of specialized sparse vector units. An example of thos
 e are the NVIDIA’s Sparse Tensor Cores(SPTCs), which promise a 2x speedup.
  However, SPTCs only support the 2:4 format, limiting achievable sparsity 
 ratios to 50%. We present the V:N:M format, which enables the execution of
  arbitrary N:M ratios on SPTCs. To efficiently exploit the resulting forma
 t, we propose Spatha, a high-performance sparse-library for DL routines. W
 e show that Spatha achieves up to 37x speedup over cuBLAS. We also demonst
 rate a second-order pruning technique that enables sparsification to high 
 sparsity ratios with V:N:M and little to no loss in accuracy in modern tra
 nsformers.\n\nTag: Artificial Intelligence/Machine Learning, Codesign, Per
 formance Optimization, Programming Frameworks and System Software\n\nRegis
 tration Category: Tech Program Reg Pass\n\nReproducibility Badges: Artifac
 t Available, Artifact Functional, Results Reproduced\n\nSession Chair: Apa
 rna Chandramowlishwaran (University of California, Irvine)\n\n
END:VEVENT
END:VCALENDAR