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:20260728T174300
DTEND;TZID=America/Los_Angeles:20260728T174300
UID:dac_DAC 2026_sess317_RESEARCH461@linklings.com
SUMMARY:MoE-Balance: Unleashing the Full Potential of CPU-GPU Heterogeneit
 y for Efficient MoE Inference
DESCRIPTION:Tiantian Lin and Xiaohang Wang (Zhejiang University), Yingtao 
 Jiang (University of Nevada), Amit Kumar Singh (University of Essex), and 
 Kui Ren (Zhejiang University)\n\nSparse Mixture-of-Experts (MoE) models of
 fer a powerful way to increase the capacity of large language models witho
 ut proportionally increasing computation. Yet, their extremely large exper
 t weight parameters create severe memory challenges for consumer-grade GPU
 s and edge devices. Existing offloading and hybrid CPU–GPU approaches fall
  short in fully utilizing CPU resources, and their coarse-grained scheduli
 ng and poorly overlapped CPU-GPU data transfers leave CPU cores frequently
  idle, thus causing pipeline bubbles and exacerbating overall system ineff
 iciency. We present MoE-Balance, a system that efficiently utilizes hetero
 geneous CPU-GPU resources for MoE inference. The key innovations include a
  fine-grained dynamic scheduler that distributes expert tasks based on rea
 l-time load, and a cross-layer prefetching and chunked transfer mechanism 
 that overlaps computation with weight movement via PCIe. Experiments on Mi
 xtral-8x7B show that MoE-Balance achieves 1.6x higher throughput and 37.5%
  lower end-to-end latency compared with state-of-the-art hybrid baselines,
  demonstrating strong performance gains under memory-constrained settings.
 \n\nTrack: EDA7-II. Physical Design and Verification, Design\n\nSession Ch
 airs: Christian Pilato (Politecnico di Milano); Mohsen Imani (University o
 f California, Irvine); Flavio Ponzina (San Diego State University); Amin F
 arshidi (Cadence Design Systems, Inc.); Yonggan Fu (Georgia Institute of T
 echnology, NVIDIA); Dimitrios Tsaras (Huawei, Hong Kong University of Scie
 nce and Technology (HKUST)); Anh Nguyen (Fujitsu Limited); Ruiyang Qin (Vi
 llanova University); Jiliang Zhang (Hunan University); Meng Li (Peking Uni
 versity); and Rassul Bairamkulov (Advanced Micro Devices (AMD))\n\n
END:VEVENT
END:VCALENDAR