BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20240310T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20231105T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20231201T014613Z UID:40EDE2CE-62BB-435D-888F-0F7C0101CE4B DTSTART;TZID=America/New_York:20231130T190000 DTEND;TZID=America/New_York:20231130T200000 DESCRIPTION:Foundation models such as GPT-4 have garnered significant inter est from both academia and industry. An outstanding feature of such models is so-called emergent capabilities\, including multi-step reasoning\, ins truction following\, and model calibration\, in a wide range of applicatio n domains. Such capabilities were previously only attainable with speciall y designed ML models\, such as those using carefully constructed knowledge graphs\, in specific domains. As the capabilities of foundation models ha ve increased\, so too have their sizes at a rate much faster than Moore's law. The training of foundation models requires massive computing power. F or instance\, training a BERT model on a single state-of-the-art GPU machi ne with multi-A100 chips can take several days\, while training GPT-3 mode ls on a large multi-instance GPU cluster can take several months to comple te the estimated 3*10^23 flops.\n\nThis talk provides an overview of the l atest progress in supporting foundation model training and inference with new AI accelerators. It reviews progress on the modeling side\, with an em phasis on the transformer architecture\, and presents the system architect ure supporting training and serving foundation models.\n\nExplore the fron tier of AI with us as we delve into the power and potential of foundation models like GPT-4. Discover how emergent capabilities are pushing the boun daries of what's possible and the groundbreaking AI accelerators making it all happen. Join our talk to uncover the future of AI training and applic ation!\n\nSpeaker(s): Jun (Luke) Huan\, \n\nVirtual: https://events.vtools .ieee.org/m/382797 LOCATION:Virtual: https://events.vtools.ieee.org/m/382797 ORGANIZER:lifanghescut@gmail.com SEQUENCE:27 SUMMARY:Training Large-scale Foundation Models on Emerging AI Accelerators URL;VALUE=URI:https://events.vtools.ieee.org/m/382797 X-ALT-DESC:Description: <br /><p>Foundation models such as GPT-4 have garne red significant interest from both academia and industry. An outstanding f eature of such models is so-called emergent capabilities\, including multi -step reasoning\, instruction following\, and model calibration\, in a wid e range of application domains. Such capabilities were previously only att ainable with specially designed ML models\, such as those using carefully constructed knowledge graphs\, in specific domains. As the capabilities of foundation models have increased\, so too have their sizes at a rate much faster than Moore's law. The training of foundation models requires massi ve computing power. For instance\, training a BERT model on a single state -of-the-art GPU machine with multi-A100 chips can take several days\, whil e training GPT-3 models on a large multi-instance GPU cluster can take sev eral months to complete the estimated 3*10^23 flops.</p>\n<p>This talk pro vides an overview of the latest progress in supporting foundation model tr aining and inference with new AI accelerators. It reviews progress on the modeling side\, with an emphasis on the transformer architecture\, and pre sents the system architecture supporting training and serving foundation m odels.<br /><br />Explore the frontier of AI with us as we delve into the power and potential of foundation models like GPT-4. Discover how emergent capabilities are pushing the boundaries of what's possible and the ground breaking AI accelerators making it all happen. Join our talk to uncover th e future of AI training and application!</p> END:VEVENT END:VCALENDAR