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:20260402T024532Z
LOCATION:3001\, Level 3
DTSTART;TZID=America/Los_Angeles:20250625T140000
DTEND;TZID=America/Los_Angeles:20250625T141500
UID:dac_DAC 2025_sess109_RESEARCH243@linklings.com
SUMMARY:Discovering and Exploiting Untapped Buffer Resources in Many-Core 
 DNN Accelerators
DESCRIPTION:Yuchen Wei, Jingwei Cai, and Mingyu Gao (Tsinghua University);
  Sen Peng and Zuotong Wu (Xi'an Jiaotong University); and Guiming Shi and 
 Kaisheng Ma (Tsinghua University)\n\nIn large-scale DNN inference accelera
 tors, the many-core architecture has emerged as a predominant design, with
  layer-pipeline (LP) mapping being a mainstream mapping approach. However,
  our experimental findings and theoretical justifications uncover a hardwa
 re-independent and prevalent flaw in employing layer-pipeline mapping on m
 any-core accelerators: a significant underutilization of buffer space acro
 ss numerous cores, indicating substantial potential for optimization. Buil
 ding on this discovery, we develop a universal and efficient buffer alloca
 tion strategy, BufferProspector, which includes a Buffer Requirement Calcu
 lator and Buffer Allocator, to capitalize on these unused buffers, address
 ing the timing mismatch challenge inherent in LP mapping. Compared to the 
 state-of-the-art (SOTA) open-source LP mapping framework Tangram, BufferPr
 ospector averages a simultaneous increase in energy efficiency and perform
 ance by 1.44x and 2.26x, respectively. BufferProspector will be open-sourc
 ed.\n\nTopics: AI\n\nTracks: AI3: AI/ML Architecture Design\n\nSession Cha
 irs: Andreas Herkersdorf (Technische Universität München) and Johannes Max
 imilian Kuehn (Nvidia)\n\n
END:VEVENT
END:VCALENDAR