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:20260730T152638Z
LOCATION:
DTSTART;TZID=America/Los_Angeles:20260728T125100
DTEND;TZID=America/Los_Angeles:20260728T125100
UID:dac_DAC 2026_sess315_RESEARCH368@linklings.com
SUMMARY:TriMoE: Augmenting GPU with AMX-Enabled CPU and DIMM-NDP for High-
 Throughput MoE Inference via Offloading
DESCRIPTION:Yudong Pan (CICS, Institute of Computing Technology, Chinese A
 cademy of Sciences;University of Chinese Academy of Sciences;); Yintao He 
 (State Key Laboratory of Computer Architecture, Institute of Computing Tec
 hnology, Chinese Academy of Sciences); Tianhua Han (Institute of Computing
  Technology, CAS); Lian Liu (Institute of Computing Technology); Shixin Zh
 ao (Chinese Academy of Sciences); Zhirong Chen (Research Center for Intell
 igent Computing System, Institute of Computing Technology, Chinese Academy
  of Sciences); Mengdi Wang and Cangyuan Li (Institute of Computing Technol
 ogy, Chinese Academy of Sciences); yinhe han (Institute of Computing Techn
 ology, CAS); and Ying Wang (State Key Laboratory of Computer Architecture,
  Institute of Computing Technology, Chinese Academy of Sciences)\n\nTo dep
 loy large Mixture-of-Experts (MoE) models cost-effectively, offloading-bas
 ed single-GPU heterogeneous inference is crucial. While GPU–CPU architectu
 res that offload cold experts are constrained by host memory bandwidth, em
 erging GPU-NDP architectures utilize DIMM-NDP to offload non-hot experts. 
 However, non-hot experts are not a homogeneous memory-bound group: a signi
 ficant subset of warm experts exists is severely penalized by high GPU I/O
  latency yet can saturate NDP compute throughput, exposing a critical comp
 ute gap. We present TriMoE, a novel GPU–CPU–NDP architecture that fills th
 is gap by synergistically leveraging AMX-enabled CPU to precisely map hot,
  warm, and cold experts onto their optimal compute units. We further intro
 duce a bottleneck-aware expert scheduling policy and a prediction-driven d
 ynamic relayout/rebalancing scheme. Experiments demonstrate that TriMoE ac
 hieves up to 2.83$\times$ speedup over state-of-the-art solutions.\n\nTopi
 cs: Quantum\n\nTrack: DES6. Quantum Computing, Systems\n\nSession Chairs: 
 Hokeun Kim (Arizona State University); BaekGyu Kim (Daegu Gyeongbuk Instit
 ute of Science and Technology); Xiang Chen (Peking University); Sangyeob K
 im (Yonsei University); Robert M. Radway (University of Pennsylvania); Pra
 shanth Mohan (University of Pittsburgh); Xiaoxuan Yang (University of Virg
 inia, Stanford University); Ramtin Zand (University of South Carolina); Ha
 di Kamali (University of Central Florida); Jiafeng Xie (Villanova Universi
 ty); Prabuddha Chakraborty (University of Maine); Junichiro Kadomoto (Univ
 ersity of Tokyo); and Yosuke Ueno (RIKEN)\n\n
END:VEVENT
END:VCALENDAR