BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT DTSTART:20260308T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260321T215331Z UID:A0AC4316-E22A-42A4-9051-05542B02882B DTSTART;TZID=America/Los_Angeles:20260318T164500 DTEND;TZID=America/Los_Angeles:20260318T180000 DESCRIPTION:Join us for a talk on how SmartNICs and RDMA Power AI in the Cl oud\, check out an Embodied AI demo and get insights into the state of the Chapter.\n\nTraining modern Large Language Models (LLMs) requires tens of thousands of GPUs acting as a single "AI Supercomputer." To build this "A I Hypercomputer\," we must first address the CPU bottlenecks of traditiona l general-purpose networking. This talk begins by analyzing why standard T CP/IP processing limits Model Training performance and introduces the conc ept of "Kernel Bypass" and the role of SmartNICs in offloading network pro cessing from the host CPU. We will explore why modern AI clusters have mov ed toward hardware offloads (like RDMA) to achieve the high throughput and low latency required for GPU-to-GPU communication. We will also discuss t he specific challenges of running lossless transport protocols over lossy Ethernet\, where congestion and packet drops can cause severe performance degradation ("tail latency") in large-scale training jobs. The session con cludes by analyzing the architectural design patterns required to optimize flow control and ensure reliable delivery in massive AI infrastructure en vironments.\nDemo: Comparing Reinforcement Learning with Imitation Learnin g for Autonomous Warehouse Pick-and-Place using a Robotic Arm\n\nThis demo simulates a last-meter warehouse picking task\, inspired by Amazon/Kiva-s tyle systems but using general-purpose robotics. The experiments explicitl y contrast policy-gradient reinforcement learning methods such as PPO with imitation learning inside a physically realistic embodied-AI task built w ith Isaac Sim. The demo has been designed to expose where each algorithm s truggles or excels due to action spaces\, partial observability\, contact dynamics\, and reward structure. These are core issues in embodied AI.\n\n This event features a leading industry expert from Google addressing this important topic\, followed by a demo on Embodied AI using Isaac Sim / Lab updates on the state of our chapter from the IEEE CIS SCV Chair.\n\nšŸŽ¤ T alk 1\nThe Infrastructure of AI: How SmartNICs and RDMA Power the Cloud\nS peaker: Sujithra Periasamy\, Google\n\nšŸŽ¤ Demo and Talk\nComparing Model -Free RL Algorithms for Autonomous Warehouse Pick-and-Place with Mobile Ma nipulation\n\nSpeakers: Mayank Kapadia and Dr. Vishnu S. Pendyala\, Depart ment of Applied Data Science\, College of Information\, Data\, and Society \, San Jose State University\n\nšŸŽ¤ Talk 2\nState of the Chapter\nSpeaker : Dr. Vishnu S. Pendyala\, Chair\, IEEE CIS Santa Clara Valley Chapter\n\n Co-sponsored by: Vishnu S. Pendyala\, San Jose State University\n\nSpeaker (s): Sujithra Periasamy\, Dr. Vishnu S Pendyala\n\nRoom: MLK Room 225\, Dr . Martin Luther King\, Jr. Library (SJSU)\, 150 E San Fernando St San Jose \, California 95112\, San Jose\, California\, United States\, Virtual: htt ps://events.vtools.ieee.org/m/537154 LOCATION:Room: MLK Room 225\, Dr. Martin Luther King\, Jr. Library (SJSU)\, 150 E San Fernando St San Jose\, California 95112\, San Jose\, California \, United States\, Virtual: https://events.vtools.ieee.org/m/537154 ORGANIZER:pendyala@ieee.org SEQUENCE:61 SUMMARY:Chapter Open House\, talk on AI Infrastructure\, and Embodied AI de mo URL;VALUE=URI:https://events.vtools.ieee.org/m/537154 X-ALT-DESC:Description: <br /><p data-start="181" data-end="257">Join us fo r a talk on<strong data-start="206" data-end="226"> how SmartNICs and RDMA Power AI in the Cloud\, </strong>check out an <strong>Embodied AI demo </ strong>and get insights into the <strong>state of the Chapter.</strong></p >\n<p><span style="color: #0000ff\;" data-ogsc="" data-ogac="#0000ff">Trai ning modern Large Language Models (LLMs) requires tens of thousands of GPU s acting as a single "AI Supercomputer." To build this "AI Hypercomputer\, " we must first address the CPU bottlenecks of traditional general-purpose networking. This talk begins by analyzing why standard TCP/IP processing limits Model Training performance and introduces the concept of "Kernel By pass" and the role of SmartNICs in offloading network processing from the host CPU. We will explore why modern AI clusters have moved toward hardwar e offloads (like RDMA) to achieve the high throughput and low latency requ ired for GPU-to-GPU communication. We will also discuss the specific chall enges of running lossless transport protocols over lossy Ethernet\, where congestion and packet drops can cause severe performance degradation ("tai l latency") in large-scale training jobs. The session concludes by analyzi ng the architectural design patterns required to optimize flow control and ensure reliable delivery in massive AI infrastructure environments.</span ></p>\n<div data-ogsc="rgb(0\, 0\, 0)"><span style="color: rgb(22\, 145\, 121)\;">Demo: Comparing Reinforcement Learning with Imitation Learning for Autonomous Warehouse Pick-and-Place using a Robotic Arm</span></div>\n<di v data-ogsc="rgb(0\, 0\, 0)">&nbsp\;</div>\n<div data-ogsc="rgb(0\, 0\, 0) "><span style="color: rgb(22\, 145\, 121)\;">This demo simulates a last-me ter warehouse picking task\, inspired by Amazon/Kiva-style systems but usi ng general-purpose robotics. The experiments explicitly contrast policy-gr adient reinforcement learning methods such as PPO with imitation learning inside a physically realistic embodied-AI task built with Isaac Sim. The d emo has been designed to expose where each algorithm struggles or excels d ue to action spaces\, partial observability\, contact dynamics\, and rewar d structure. These are core issues in embodied AI.</span></div>\n<div data -ogsc="rgb(0\, 0\, 0)">&nbsp\;</div>\n<div data-ogsc="rgb(0\, 0\, 0)"><spa n style="color: rgb(0\, 0\, 0)\;">This event features a leading industry e xpert from Google addressing this important topic\, followed by a demo on Embodied AI using Isaac Sim / Lab updates on the state of our chapter from the IEEE CIS SCV Chair.</span></div>\n<p data-start="663" data-end="824"> šŸŽ¤ <strong data-start="666" data-end="676">Talk 1</strong><br data-start ="676" data-end="679"><em data-start="771" data-end="781">The Infrastructu re of AI: How SmartNICs and RDMA Power the Cloud<br>Speaker:</em> Sujithra Periasamy\, Google</p>\n<p data-start="663" data-end="824">šŸŽ¤ <strong d ata-start="829" data-end="839">Demo and Talk</strong><br data-start="839" data-end="842"><em>Comparing Model-Free RL Algorithms for Autonomous Wareh ouse Pick-and-Place with Mobile Manipulation</em></p>\n<p data-start="663" data-end="824"><em data-start="869" data-end="879">Speakers: Mayank Kapad ia and</em> Dr. Vishnu S. Pendyala\, Department of Applied Data Science\, College of Information\, Data\, and Society\, San Jose State University</p >\n<p data-start="826" data-end="952">šŸŽ¤ <strong data-start="829" data-e nd="839">Talk 2</strong><br data-start="839" data-end="842"><strong data-s tart="842" data-end="866">State of the Chapter</strong><br data-start="866 " data-end="869"><em data-start="869" data-end="879">Speaker:</em> <strong data-start="880" data-end="906">Dr. Vishnu S. Pendyala</strong>\, Chair\, IEEE CIS Santa Clara Valley Chapter</p> END:VEVENT END:VCALENDAR