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:20260402T024532Z
LOCATION:3001\, Level 3
DTSTART;TZID=America/Los_Angeles:20250625T154500
DTEND;TZID=America/Los_Angeles:20250625T160000
UID:dac_DAC 2025_sess132_RESEARCH671@linklings.com
SUMMARY:AttenPIM: Accelerating LLM Attention with Dual-mode GEMV in Proces
 sing-in-Memory
DESCRIPTION:Liyan Chen, Dongxu Lyu, Zhenyu Li, Jianfei Jiang, Qin Wang, Zh
 igang Mao, and Naifeng Jing (Shanghai Jiao Tong University)\n\nLarge Langu
 age Models (LLMs) have demonstrated unprecedented generative performance a
 cross a wide range of applications. While recent heterogeneous architectur
 es attempt to address the memory-bound bottleneck from attention computati
 ons by processing-in-memory (PIM) offloading, they overlook two critical c
 haracteristics of attention GEMVs that distinguish them from traditional P
 IM scenarios: (1) dynamic matrix dimensions that scale with token length, 
 and (2) distinct GEMV patterns between score computation (Q×Kt) and contex
 t computation (S×V). Existing PIM designs, employing either uniform or tra
 nsposed computing modes, suffer from inefficiencies in newly generated ele
 ment preparation or distinct GEMV execution. To address these limitations,
  we propose AttenPIM, a software-hardware co-design for efficient PIM-base
 d attention acceleration. For bank-level execution, we propose dual-mode c
 omputing modes tailored for score and context computations with PIM-orient
 ed data layouts and execution flows for KV storage, supported by a low-cos
 t configurable per-bank PIM unit (PU). For system-level execution, we leve
 rage token-level and head-level concurrency to ensure workload balance and
  maximize bank PU parallelism. Furthermore, dynamic allocation and kernel 
 fusion method are proposed to further minimize memory overhead. Experiment
 al results demonstrate that AttenPIM achieves 1.13×-5.26× speedup and redu
 ces energy consumption by 17%-49% compared to two state-of-the-art PIM bas
 elines.\n\nTopics: Design\n\nTracks: DES2B: In-memory and Near-memory Comp
 uting Architectures, Applications and Systems\n\nSession Chairs: Steve Dai
  (NVIDIA) and Haitong Li (Purdue University)\n\n
END:VEVENT
END:VCALENDAR