BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Computer Science and Engineering - ECPv6.13.0//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-ORIGINAL-URL:https://homecse.iitd.ac.in
X-WR-CALDESC:Events for Computer Science and Engineering
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
BEGIN:STANDARD
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
DTSTART:20260101T000000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=Asia/Kolkata:20260819T110000
DTEND;TZID=Asia/Kolkata:20260819T120000
DTSTAMP:20260922T154808
CREATED:20260813T065722Z
LAST-MODIFIED:20260813T065722Z
UID:2636-1787137200-1787140800@homecse.iitd.ac.in
SUMMARY:Foundations of Learning from Positive Samples by Dr. Anay Mehrotra
DESCRIPTION:Abstract: What can be learned from data? Traditional answers to this question assume an idealized data-generating process where test and training distributions are symmetric\, which is rarely the case in applications.\nIn this talk\, we will revisit this question for positive-only learning\, a setting where only positive examples are observed. This is a challenging problem arising in bioinformatics and causal inference. Classical results show that learning is impossible in general. However\, we will show that the hard instances are fragile: under a smoothed analysis\, which rules out pathological distributions\, efficient learning is possible. The same ideas also lead to faster and more general algorithms for estimation and regression from truncated data\, a foundational problem in statistics.\nThe talk is based on joint work with Shai Ben-David\, Yang Cai\, Constantine Caramanis\, Alkis Kalavasis\, Alex Kouridakis\, Jane H. Lee\, Katerina Mamali\, Farnam Mansouri\, and Manolis Zampetakis. \nSpeaker bio: Anay Mehrotra (https://anaymehrotra.com) is a Motwani Postdoctoral Fellow at Stanford working with Amin Saberi. He recently completed his PhD at Yale\, advised by Amin Karbasi and Manolis Zampetakis. His work has received the Best Paper Award at COLT and the Sri Binay Kumar Sinha Award from IIT Kanpur\, was selected for the TCS4All Rising Star session at STOC 2026\, and has been featured in WIRED.
URL:https://homecse.iitd.ac.in/event/foundations-of-learning-from-positive-samples-by-dr-anay-mehrotra/
LOCATION:Bharti 501\, IIT Campus\, Hauz Khas\, New Delhi
CATEGORIES:Seminars
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=Asia/Kolkata:20260825T120000
DTEND;TZID=Asia/Kolkata:20260825T130000
DTSTAMP:20260922T154808
CREATED:20260818T060107Z
LAST-MODIFIED:20260818T060107Z
UID:2651-1787659200-1787662800@homecse.iitd.ac.in
SUMMARY:Scaling Up GPU Memory Management by Dr. Pratheek B
DESCRIPTION:Abstract:  \nThe volume of data generated worldwide is growing at an unprecedented rate\, and GPUs have emerged as the primary compute engine for processing this data. While GPUs offer massive compute power — reaching thousands of TFLOPS — they are constrained by the relatively low memory bandwidth (around a few TB/s) and limited memory capacity (in tens of GBs). As a result\, memory is often the primary performance bottleneck in many GPU applications.\n\nIn this talk\, we explore two key aspects of GPU memory management: memory oversubscription and address translation. Memory oversubscription enables GPU programs to work on datasets larger than the on-board GPU memory\, but can lead to severe slowdowns. Efficient address translation is important in GPUs\, as it lies in the critical path of memory accesses\, and thus impacts overall GPU performance.\n\nThis talk will primarily focus on the challenges posed by GPU memory oversubscription. GPU memory oversubscription enables GPU applications to work with datasets larger than the GPU memory capacity\, using the CPU memory as swap space. Unfortunately\, applications under GPU memory oversubscription often experience significant slowdowns. Our work\, ObservUVM\, improves GPU's eviction and prefetching policies by enabling observability into GPU’s memory accesses to pages resident on GPU memory — something current GPUs lack. We show that  current eviction and prefetching policies\, handled by the driver running on the CPU\, are limited in their ability to make informed decisions due to the lack of observability. ObservUVM enables observability into GPU’s memory accesses by repurposing existing hardware access counters\, enabling better-informed eviction and prefetching policies. ObservUVM improves UVM performance by around 33% across 14\napplications.\n\nWe will also touch upon a few other relevant problems in GPU memory management. We will discuss SUV\, our compiler-driven technique to perform automated data placement and migration in memory-oversubscribed GPUs. Then we will briefly discuss the impact of non-uniformity of Multi-Chip-Module (MCM) design on address\ntranslation in GPUs\, and on the impact of multi-tenancy on address translation in GPUs.\n\n\nBio:  \nPratheek is a Senior Software Engineer at NVIDIA\, working on improving GPU memory management. He obtained his Ph.D. from the Indian Institute of Science\, Bengaluru in 2026. His research revolves around improving GPU memory management\, focusing on address translation and data placement for large-memory workloads. His work spans the domains of GPU micro-architecture\, compiler techniques\, and operating systems.\nPreviously\, he had worked on reverse-engineering Nvidia GPUs.
URL:https://homecse.iitd.ac.in/event/scaling-up-gpu-memory-management-by-dr-pratheek-b/
LOCATION:Bharti 501\, IIT Campus\, Hauz Khas\, New Delhi
CATEGORIES:Seminars
END:VEVENT
END:VCALENDAR