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:
DTSTART;TZID=America/Los_Angeles:20260728T173100
DTEND;TZID=America/Los_Angeles:20260728T173100
UID:dac_DAC 2026_sess317_RESEARCH1441@linklings.com
SUMMARY:An Adaptive Prediction of Tokens Fluctuation for Decoupled Prefetc
 hing to Accelerate Long-Context LLM Decoding
DESCRIPTION:Yuanhua Xiao and Jihe Wang (Chinese Academy of Sciences), Zhiy
 u Sun and Linying Wu (Northwestern Polytechnical University), and Danghui 
 Wang (NPU)\n\nThe prohibitively slow speed of long-context LLM decoding is
  a critical bottleneck, caused by massive Key-Value (KV) access that satur
 ates memory bandwidth. Existing solutions fail due to a fundamental trade-
 off: static predictors cannot handle the Tokens Fluctuation, leading to pe
 rmanent accuracy loss from irreversible eviction, while an on-the-fly scor
 ing method creates a synchronous bottleneck that stalls the entire pipelin
 e. This paper presents a co-designed system that breaks this trade-off by 
 leveraging an Adaptive Prediction algorithm to enable a fully Decoupled Pr
 efetching architecture. The proposed Adaptive Prediction model first mirro
 rs the nature of Tokens Fluctuation, enabling a lightweight, reuse-based r
 ecovery mechanism that solves irreversible eviction. This high-fidelity pr
 ediction then unlocks the Decoupled Prefetching system, which completely h
 ides data transfer latency and eliminates the synchronous bottleneck. On l
 ong-context LLM decoding tasks, the proposed system matches the accuracy o
 f a full KV cache with just 256 tokens, while achieving an average 2.50x e
 nd-to-end speedup.\n\nTrack: EDA7-II. Physical Design and Verification, De
 sign\n\nSession Chairs: Christian Pilato (Politecnico di Milano); Mohsen I
 mani (University of California, Irvine); Flavio Ponzina (San Diego State U
 niversity); Amin Farshidi (Cadence Design Systems, Inc.); Yonggan Fu (Geor
 gia Institute of Technology, NVIDIA); Dimitrios Tsaras (Huawei, Hong Kong 
 University of Science and Technology (HKUST)); Anh Nguyen (Fujitsu Limited
 ); Ruiyang Qin (Villanova University); Jiliang Zhang (Hunan University); M
 eng Li (Peking University); and Rassul Bairamkulov (Advanced Micro Devices
  (AMD))\n\n
END:VEVENT
END:VCALENDAR