BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:America/New_York
BEGIN:DAYLIGHT
DTSTART:20260308T030000
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3
TZNAME:EDT
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20261101T010000
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11
TZNAME:EST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20261017T101513Z
UID:D4772576-B18A-4040-8BF5-C51CD42446B2
DTSTART;TZID=America/New_York:20261016T120000
DTEND;TZID=America/New_York:20261016T130000
DESCRIPTION:Title: Memory and Energy Efficient Inference of Large Mixture-o
 f-Experts Foundation Models\n\nAbstract:\n\nLarge Mixture-of-Experts (MoE)
  foundation models improve scalability by activating only a sparse subset 
 of parameters for each input\, but inference remains bottlenecked by memor
 y movement\, hardware inefficiency\, and energy consumption. This talk pre
 sents two complementary\, theory-guided approaches for efficient MoE infer
 ence across digital and analog hardware platforms. First\, I will introduc
 e an expert-wise mixed-precision quantization framework that identifies cr
 itical experts through router training dynamics and allocates quantization
  bit-widths accordingly\, achieving improved low-bit inference accuracy an
 d lower memory cost on large models such as Switch Transformer and Mixtral
 . Second\, I will discuss a theory-guided analog–digital mapping strateg
 y for analog in-memory computing (AIMC)\, where a proposed metric identifi
 es noise-sensitive experts and dense modules for selective deployment on d
 igital hardware\, while the remaining operations are executed on energy-ef
 ficient analog units. Experiments on models including DeepSeek MoE and OLM
 oE demonstrate near full-precision accuracy with substantially reduced mem
 ory movement and energy consumption.\n\nSpeaker(s): Meng Wang\, \n\nVirtua
 l: https://events.vtools.ieee.org/m/582662
LOCATION:Virtual: https://events.vtools.ieee.org/m/582662
ORGANIZER:guoxian2012@gmail.com
SEQUENCE:52
SUMMARY:IEEE PES &amp; WIE AG Schenectady Chapter &amp; IEEE RPI Student Branch (Vi
 rtual) Lecture on “Memory and Energy Efficient Inference of Large Mixtur
 e-of-Experts Foundation Models”
URL;VALUE=URI:https://events.vtools.ieee.org/m/582662
X-ALT-DESC:Description: &lt;br /&gt;&lt;p class=&quot;MsoNormal&quot;&gt;&lt;strong&gt;&lt;span style=&quot;fon
 t-family: &#39;Arial&#39;\,sans-serif\;&quot;&gt;Title:&lt;/span&gt;&lt;/strong&gt;&lt;span style=&quot;font-f
 amily: &#39;Arial&#39;\,sans-serif\;&quot;&gt; Memory and Energy Efficient Inference of La
 rge Mixture-of-Experts Foundation &lt;/span&gt;&lt;span style=&quot;font-family: &#39;Arial&#39;
 \,sans-serif\;&quot;&gt;Models&lt;/span&gt;&lt;/p&gt;\n&lt;p class=&quot;MsoNormal&quot;&gt;&lt;strong&gt;&lt;span styl
 e=&quot;font-family: &#39;Arial&#39;\,sans-serif\;&quot;&gt;Abstract:&amp;nbsp\;&lt;/span&gt;&lt;/strong&gt;&lt;/p
 &gt;\n&lt;p class=&quot;MsoNormal&quot; style=&quot;text-align: justify\;&quot;&gt;&lt;span style=&quot;font-si
 ze: 12pt\;&quot;&gt;&lt;span style=&quot;font-family: Verdana\, sans-serif\; color: rgb(19
 \, 19\, 19)\;&quot;&gt;Large Mixture-of-Experts (MoE) foundation models improve sc
 alability by activating only a sparse subset of parameters for each input\
 , but inference remains bottlenecked by memory movement\, hardware ineffic
 iency\, and energy consumption. This talk presents two complementary\, the
 ory-guided approaches for efficient MoE inference across digital and analo
 g hardware platforms. First\, I will introduce an expert-wise mixed-precis
 ion quantization framework that identifies critical experts through router
  training dynamics and allocates quantization bit-widths accordingly\, ach
 ieving improved low-bit inference accuracy and lower memory cost on large 
 models such as Switch Transformer and Mixtral. Second\, I will discuss a t
 heory-guided analog&amp;ndash\;digital mapping strategy for analog in-memory c
 omputing (AIMC)\, where a proposed metric identifies noise-sensitive exper
 ts and dense modules for selective deployment on digital hardware\, while 
 the remaining operations are executed on energy-efficient analog units. Ex
 periments on models including DeepSeek MoE and OLMoE demonstrate near full
 -&lt;/span&gt;&lt;span style=&quot;font-family: Verdana\, sans-serif\; color: rgb(19\, 1
 9\, 19)\;&quot;&gt;precision accuracy with substantially reduced memory movement a
 nd energy consumption.&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;\n&lt;p class=&quot;MsoNormal&quot; style=&quot;marg
 in-bottom: 0in\; line-height: 150%\; text-align: justify\;&quot;&gt;&amp;nbsp\;&lt;/p&gt;
END:VEVENT
END:VCALENDAR

