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:20260728T124000
DTEND;TZID=America/Los_Angeles:20260728T124100
UID:dac_DAC 2026_sess315_RESEARCH2432@linklings.com
SUMMARY:Hybrid PIM-oriented Architecture-Dataflow Co-Design for Heterogene
 ous LLM Inference
DESCRIPTION:Hai Huang and Feng Qiu (Beihang University), Xiang Chen (Pekin
 g University), and Weisheng Zhao and Chenchen Liu (Beihang University)\n\n
 Recent studies show that incorporating Processing-in-Memory (PIM) into neu
 ral processing units (NPUs) can significantly improve the efficiency of la
 rge language model (LLM) inference on edge devices.\nHowever, the limited 
 compute capability of DRAM-PIM and the use of fixed dataflow strategies st
 ill constrain overall performance gains.\nTo address these limitations, we
  propose a hybrid PIM architecture that integrates non-volatile spin-orbit
  torque magnetic random-access memory (SOT-MRAM)-based PIM into existing N
 PU-PIM systems, fully exploiting the high compute capability and bandwidth
  of SOT-MRAM together with the large storage capacity of DRAM.\nOn top of 
 this architecture, we further design an adaptive compute scheduling and da
 taflow optimization framework.\nUsing NSGA-II-based multi-objective datafl
 ow space exploration, our framework identifies Pareto-optimal hardware res
 ource allocations and dataflow configurations under different deployment o
 bjectives.\nExperimental results show that our approach reduces inference 
 latency by up to 8.72× and power consumption by 11.74× compared to NPUs, a
 nd further achieves 6.24× lower latency and 7.76× higher power efficiency 
 than NPU-PIM baselines.\n\nTopics: Quantum\n\nTrack: DES6. Quantum Computi
 ng, Systems\n\nSession Chairs: Hokeun Kim (Arizona State University); Baek
 Gyu Kim (Daegu Gyeongbuk Institute of Science and Technology); Xiang Chen 
 (Peking University); Sangyeob Kim (Yonsei University); Robert M. Radway (U
 niversity of Pennsylvania); Prashanth Mohan (University of Pittsburgh); Xi
 aoxuan Yang (University of Virginia, Stanford University); Ramtin Zand (Un
 iversity of South Carolina); Hadi Kamali (University of Central Florida); 
 Jiafeng Xie (Villanova University); Prabuddha Chakraborty (University of M
 aine); Junichiro Kadomoto (University of Tokyo); and Yosuke Ueno (RIKEN)\n
 \n
END:VEVENT
END:VCALENDAR