BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Chicago
X-LIC-LOCATION:America/Chicago
BEGIN:DAYLIGHT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
TZNAME:CDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
TZNAME:CST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260202T201805Z
LOCATION:266
DTSTART;TZID=America/Chicago:20251116T145000
DTEND;TZID=America/Chicago:20251116T150500
UID:submissions.supercomputing.org_SC25_sess221_ws_indis116@linklings.com
SUMMARY:Scaling LLM Training Using RDMA over Converged Ethernet
DESCRIPTION:Alex Batlle Casellas (Qualcomm Europe, Inc.); Adrián Pérez Dié
 guez (Qualcomm Technologies, Inc.); Aleix Torres-Camps (Qualcomm Europe, I
 nc.); Harris Teague (Qualcomm Technologies, Inc.); and Arnau Padres and Jo
 rdi Ros-Giralt (Qualcomm Europe, Inc.)\n\nWe present a comprehensive bench
 marking study that evaluates the scaling performance of RDMA over Converge
 d Ethernet (RoCE) and compares it with Infiniband in the context of large-
 scale LLM training workloads. While Infiniband is traditionally favored fo
 r its low-latency, high-bandwidth characteristics, it imposes significant 
 infrastructure and operational costs. RoCE, leveraging commodity Ethernet 
 and RDMA, offers a cost-effective alternative. Through extensive experimen
 ts on production clusters, we demonstrate that RoCE can achieve near-linea
 r scaling performance comparable to Infiniband when properly configured. O
 ur analysis spans data sharding strategies, quantization and activation re
 computation techniques, batch size tuning, and system-level optimizations,
  providing practical guidance for designing scalable and efficient AI infr
 astructure.\n\nRecording: Livestreamed, Recorded\n\nRegistration Category:
  Technical Program Reg Pass, Workshop Reg Pass\n\nSession Chairs: Ezra Kis
 sel (Energy Sciences Network (ESnet)); Rafael Silva-Guimaraes (Federal Ins
 titute of Espírito Santo, Department of Informatics); Nik Sultana (Illinoi
 s Institute of Technology, SCinet); Cees de Laat (University of Amsterdam,
  Lawrence Berkeley National Laboratory (LBNL)); Sonja Filiposka (Ss. Cyril
  and Methodius Univeristy in Skopje (UKIM) North Macedonia); and Akbar Kar
 a (Ciena Corporation, SCinet)\n\n
END:VEVENT
END:VCALENDAR