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:20260402T024533Z
LOCATION:3001\, Level 3
DTSTART;TZID=America/Los_Angeles:20250625T161500
DTEND;TZID=America/Los_Angeles:20250625T163000
UID:dac_DAC 2025_sess132_RESEARCH475@linklings.com
SUMMARY:PIMoE: Towards Efficient MoE Transformer Deployment on NPU-PIM Sys
 tem through Throttle-Aware Task Offloading
DESCRIPTION:Lizhou Wu, Haozhe Zhu, Siqi He, Xuanda Lin, and Xiaoyang Zeng 
 (State Key Lab of Integrated Chips and Systems, Fudan University) and Chix
 iao Chen (Fudan University)\n\nMixture-of-experts (MoE) technique holds si
 gnificant promise for scaling up Transformer models. \nHowever, the data t
 ransfer overhead and imbalanced workload hinder efficient deployment.\nThi
 s work presents PIMoE, a heterogeneous system combining processing-in-memo
 ry (PIM) and neural-processing-unit (NPU) to facilitate efficient MoE Tran
 sformer inference. \nWe propose a throttle-aware task offloading method th
 at addresses workload imbalance between NPU and PIM, achieving optimal tas
 k distribution.\nFurthermore, we design a near-memory-controller data cond
 enser to address the mismatch of sparse data layout between NPU and PIM, e
 nhancing data transfer efficiency. \nExperimental results demonstrate that
  PIMoE achieves 4.5× speedup and 13.7× greater energy efficiency compared 
 to the A100, and 1.4× speedup over a state-of-the-art MoE platform.\n\nTop
 ics: Design\n\nTracks: DES2B: In-memory and Near-memory Computing Architec
 tures, Applications and Systems\n\nSession Chairs: Steve Dai (NVIDIA) and 
 Haitong Li (Purdue University)\n\n
END:VEVENT
END:VCALENDAR