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:20260730T152640Z
LOCATION:Exhibit Hall
DTSTART;TZID=America/Los_Angeles:20260728T174600
DTEND;TZID=America/Los_Angeles:20260728T174600
UID:dac_DAC 2026_sess306_LBR160@linklings.com
SUMMARY:Late Breaking Results: Hybrid-Bonding Architecture with Stage-Awar
 e Mapping for Masked Autoregressive Model Inference
DESCRIPTION:Wenxi Gao (Imperial College London), Zhenhua Zhu (Tsinghua Uni
 versity), Hongyi Wang (HKUST), Yuan Xie (Hong Kong University of Science a
 nd Technology), and Yu Wang (Tsinghua University)\n\nMasked autoregressive
  (MAR) models have recently shown strong potential in visual generation ta
 sks. Unlike traditional autoregressive model that generates tokens sequent
 ially, MAR predicts multiple tokens in parallel at each generation step (s
 tage), resulting in non-linear growth in computation and memory across sta
 ges. To meet the demands, we introduce Hybrid-bonding (HB) architectures t
 o provide high memory bandwidth and scalable compute capability through 3D
  integration of DRAM and logic banks. However, existing static mapping str
 ategies cannot effectively adapt to the stage-wise workload shifts due to 
 limited remote-bank bandwidth and constrained per-bank compute capacity. I
 n this work, we propose a \textbf{stage-aware mapping strategy} for MAR in
 ference on HB-based architectures that dynamically balances communication 
 and computation costs across stages. Experimental results show up to \text
 bf{1.75$\times$ speedup} and \textbf{1.56$\times$ energy efficiency} compa
 red with baselines.\n\nTrack: Student\n\n
END:VEVENT
END:VCALENDAR