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:20260728T174200
DTEND;TZID=America/Los_Angeles:20260728T174300
UID:dac_DAC 2026_sess317_RESEARCH2345@linklings.com
SUMMARY:ScoutAttention: Efficient KV Cache Offloading via Layer-Ahead CPU 
 Pre-computation for LLM Inference
DESCRIPTION:Qiuyang Zhang, Kai Zhou, Ding Tang, and Kai Lu (Huazhong Unive
 rsity of Science and Technology); Cheng Li and Zhenyu Yang (Huawei); Peng 
 Xu (Research Center for High Efficiency Computing Infrastructure, Zhejiang
  Lab); and Jiguang Wan (Huazhong University of Science and Technology)\n\n
 Large language models encounter critical GPU memory capacity constraints d
 uring long-context inference, where KV cache memory consumption severely l
 imits decode batch sizes. While existing research has explored offloading 
 KV cache to DRAM, these approaches either demand frequent GPU-CPU data tra
 nsfers or impose extensive CPU computation requirements, resulting in poor
  GPU utilization as the system waits for I/O operations or CPU processing 
 to complete.\n\nWe propose ScoutAttention, a novel KV cache offloading fra
 mework that accelerates LLM inference through collaborative GPU-CPU attent
 ion computation. To prevent CPU computation from bottlenecking the system,
  ScoutAttention introduces GPU-CPU collaborative block-wise sparse attenti
 on that significantly reduces CPU load. Unlike conventional parallel compu
 ting approaches, our framework features a novel layer-ahead CPU pre-comput
 ation algorithm, enabling the CPU to initiate attention computation one la
 yer in advance, complemented by asynchronous periodic recall mechanisms to
  maintain minimal CPU compute load. Experimental results demonstrate that 
 ScoutAttention maintains accuracy within 2.4% of baseline while achieving 
 2.1x speedup compared to existing offloading methods.\n\nTrack: EDA7-II. P
 hysical Design and Verification, Design\n\nSession Chairs: Christian Pilat
 o (Politecnico di Milano); Mohsen Imani (University of California, Irvine)
 ; Flavio Ponzina (San Diego State University); Amin Farshidi (Cadence Desi
 gn Systems, Inc.); Yonggan Fu (Georgia Institute of Technology, NVIDIA); D
 imitrios Tsaras (Huawei, Hong Kong University of Science and Technology (H
 KUST)); Anh Nguyen (Fujitsu Limited); Ruiyang Qin (Villanova University); 
 Jiliang Zhang (Hunan University); Meng Li (Peking University); and Rassul 
 Bairamkulov (Advanced Micro Devices (AMD))\n\n
END:VEVENT
END:VCALENDAR