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:20260728T174900
DTEND;TZID=America/Los_Angeles:20260728T175000
UID:dac_DAC 2026_sess317_RESEARCH684@linklings.com
SUMMARY:ZipFloat: An Ultra-fast and Lossless Floating-Point Compressor for
  LLM Inference Systems
DESCRIPTION:Shihao Wang, Xiangyu Zou, Wen Xia, and Hao Hu (Harbin Institut
 e of Technology, Shenzhen)\n\nLLM inference stresses GPU memory with large
  model weights and KV caches, which requires efficient compression techniq
 ues.\nBut existing approaches either sacrifice accuracy (e.g., quantizatio
 n) or apply serial entropy coding (e.g., DFloat11) that limits throughput.
  Both constrain the overall inference efficiency.\n\nIn this work, we prop
 ose ZipFloat, an LLM-oriented floating-point compressor that achieves mass
 ively parallel com-/decom-pression while preserving full precision.\nWe ob
 serve that weights and KV-cache values in LLMs exhibit strong statistical 
 redundancies, yet existing floating-point formats mask this redundancy and
  hinder effective compression.\nTo address this, ZipFloat employs \n(1) Ex
 ponent Sparsification, which redefines the binary representation of floati
 ng-points to restore compressibility,\nand (2) Bit-Matrix Packing, which l
 everages this restored structure with GPU-native parallelism to deliver ex
 treme throughput.\n\nEvaluations show that ZipFloat delivers up to 700 GB/
 s com-/decom-pression throughput, outperforming SOTA methods (e.g., DFloat
 11) by several orders of magnitude while maintaining comparable compressio
 n ratios, thereby reducing over 95\% TTFT and improving over 4x inference 
 throughput in LLM systems.\n\nTrack: EDA7-II. Physical Design and Verifica
 tion, Design\n\nSession Chairs: Christian Pilato (Politecnico di Milano); 
 Mohsen Imani (University of California, Irvine); Flavio Ponzina (San Diego
  State University); Amin Farshidi (Cadence Design Systems, Inc.); Yonggan 
 Fu (Georgia Institute of Technology, NVIDIA); Dimitrios Tsaras (Huawei, Ho
 ng Kong University of Science and Technology (HKUST)); Anh Nguyen (Fujitsu
  Limited); Ruiyang Qin (Villanova University); Jiliang Zhang (Hunan Univer
 sity); Meng Li (Peking University); and Rassul Bairamkulov (Advanced Micro
  Devices (AMD))\n\n
END:VEVENT
END:VCALENDAR