BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260730T152729Z
LOCATION:Mtg Room 101A
DTSTART;TZID=America/Los_Angeles:20260728T104200
DTEND;TZID=America/Los_Angeles:20260728T105400
UID:dac_DAC 2026_sess142_RESEARCH2094@linklings.com
SUMMARY:DAPA: Distribution Aware Piecewise Activation Functions for On-Dev
 ice Transformer Inference and Training
DESCRIPTION:Maoyang Xiang and Bo Wang (Singapore University of Technology 
 and Design)\n\nNon-linear activation functions play a pivotal role in on-d
 evice inference and training, as they not only consume substantial hardwar
 e resources but also impose a significant impact on system performance and
  energy efficiency.\nIn this work, we propose Distribution-Aware Piecewise
  Activation (DAPA), a differentiable and hardware-friendly activation func
 tion for Transformer architectures by exploiting the distribution of pre-a
 ctivation data. DAPA employs a non-uniform piecewise approximation that al
 locates finer segments to high-probability regions of the distribution, im
 proving generalizability over prior piecewise linear methods. The resultin
 g approximation is further quantized using Distribution-Weighted Mean Squa
 re Error to reduce latency and resource utilization for hardware deploymen
 t.\nOur HLS implementation demonstrates that DAPA speeds up GELU computati
 on by 16x and decreases DSP utilization by 16x while maintaining comparabl
 e or better performance across vision Transformers and GPT-2 models.\n\nTo
 pics: AI\n\nTrack: AI4-II. AI/ML Architecture Design\n\nSession Chairs: Xi
 ang Chen (Peking University) and Sangyeob Kim (Yonsei University)\n\n
END:VEVENT
END:VCALENDAR
