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:20260202T201247Z
LOCATION:Second Floor Atrium
DTSTART;TZID=America/Chicago:20251120T080000
DTEND;TZID=America/Chicago:20251120T170000
UID:submissions.supercomputing.org_SC25_sess533_post286@linklings.com
SUMMARY:Energy-Efficient Multimodal LLM Inference: Stage-Level Characteriz
 ation and Input-Aware Controls​
DESCRIPTION:Mona Moghadampanah, Adib Rezaei Shahmirzadi, and Dimitrios S. 
 Nikolopoulos (Virginia Tech)\n\nMultimodal large language models (MLLMs) e
 xtend text-only LLMs with image and video encoders, enabling new capabilit
 ies but introducing high and poorly understood energy costs. This work cha
 racterizes the energy footprint of MLLM inference at the stage level, deco
 mposing serving into vision encoding, prefill, and decoding for image–text
  models. Using NVML-based measurements on an NVIDIA A100 with realistic wo
 rkloads, we demonstrate how encoder design and input complexity (resolutio
 n, image count) increase the number of visual tokens and shift energy towa
 rd prefill. Our novel contribution is linking token growth to serving inef
 ficiency and demonstrating two practical controls: complexity-aware batchi
 ng and stage-conditioned DVFS, which reduce energy while meeting latency S
 LOs. Current results highlight disproportionate energy growth from multimo
 dal inputs, and the study outlines stage-wise breakdowns, token-driven sca
 ling curves, and prototype controls that motivate future input-aware sched
 uling policies for energy-efficient multimodal inference.\n\nTag: Research
  & ACM SRC Posters\n\nRegistration Category: Technical Program Reg Pass\n\
 nSession Chairs: Kento Sato (RIKEN Center for Computational Science (R-CCS
 )); Chris Schlipalius (Pawsey Supercomputing Research Centre; Commonwealth
  Scientific and Industrial Research Organisation (CSIRO), Australia); and 
 Anja Gerbes (Georg-August-Universität Göttingen)\n\n
END:VEVENT
END:VCALENDAR