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:20260730T152637Z
LOCATION:
DTSTART;TZID=America/Los_Angeles:20260728T123000
DTEND;TZID=America/Los_Angeles:20260728T123000
UID:dac_DAC 2026_sess315_RESEARCH1582@linklings.com
SUMMARY:SliceMoE: Bit-Sliced Expert Caching under Miss-Rate Constraints fo
 r Efficient MoE Inference
DESCRIPTION:Yuseon Choi, Sangjin Kim, Jungjun Oh, Gwangtae Park, Byeongche
 ol Kim, and Hoi-Jun Yoo (Korea Advanced Institute of Science and Technolog
 y (KAIST))\n\nMoE models offer efficient scaling through conditional compu
 tation, but their large parameter size and expensive expert offloading mak
 e on-device deployment challenging. Existing acceleration techniques such 
 as prefetching or expert clustering often increase energy usage or reduce 
 expert diversity. We present SliceMoE, an energy-efficient MoE inference f
 ramework for miss-rate–constrained deployment. SliceMoE introduces Dynamic
  Bit-Sliced Caching (DBSC), which caches experts at slice-level granularit
 y and assigns precision on demand to expand effective expert capacity. To 
 support mixed-precision experts without memory duplication, we propose Cal
 ibration-Free Asymmetric Matryoshka Quantization (AMAT), a truncation-base
 d scheme that maintains compatibility between low-bit and high-bit slices.
  We further introduce Predictive Cache Warmup (PCW) to reduce early-decode
  cold misses by reshaping cache contents during prefill. Evaluated on Deep
 Seek-V2-Lite and Qwen1.5-MoE-A2.7B, SliceMoE reduces decode-stage energy c
 onsumption by up to 2.37× and 2.85×, respectively, and improves decode lat
 ency by up to 1.81× and 1.64×, while preserving near–high-bit accuracy. Th
 ese results demonstrate that slice-level caching enables an efficient on-d
 evice MoE deployment.\n\nTopics: Quantum\n\nTrack: DES6. Quantum Computing
 , Systems\n\nSession Chairs: Hokeun Kim (Arizona State University); BaekGy
 u Kim (Daegu Gyeongbuk Institute of Science and Technology); Xiang Chen (P
 eking University); Sangyeob Kim (Yonsei University); Robert M. Radway (Uni
 versity of Pennsylvania); Prashanth Mohan (University of Pittsburgh); Xiao
 xuan Yang (University of Virginia, Stanford University); Ramtin Zand (Univ
 ersity of South Carolina); Hadi Kamali (University of Central Florida); Ji
 afeng Xie (Villanova University); Prabuddha Chakraborty (University of Mai
 ne); Junichiro Kadomoto (University of Tokyo); and Yosuke Ueno (RIKEN)\n\n
END:VEVENT
END:VCALENDAR