iMAGiNE Consortium Symposium 2023

Technical Talk - Meta
Dilin Wang
Research Scientist, Meta

“Detecting and generating 3D objects”

With the growing popularity of LiDAR and VR/AR devices, it has become increasingly important to design machine learning algorithms that enable us to use these devices to understand real-world 3D scenes, create 3D content, and develop more immersive human-computer interaction experiences. In this talk, I will introduce two recent works that we collaborated with UT Austin. In the first work, we leverage 2D camera features to improve 3D detection from Lidar point clouds. Our method works by introducing a path consistency regularization to address a semantic mis-alignment issue between 2D features and 3D features. In the second work, we accelerate 3D diffusion models and demonstrate real-time point cloud generation capability with straight flows.

Technical Talk - Samsung
Jun-Woo Jang
Samsung Electronics Principal Researcher, Samsung Advanced Institute of Technology

“Memory Coupled Computing: Processing-In/Near-Memory”

As deep learning models continue to scale up, improvements in processor and memory have become increasingly crucial. However, processor enhancement has outpaced memory, leading to memory bottlenecks in large-scale applications. Additionally, significant data movement between memory and processor required by these applications results in excessive energy consumption. To address these challenges, memory-coupled computing technologies such as processing-in-memory (PIM) and processing-near-memory (PNM) have been proposed. These technologies bring computing capability closer to memory, allowing for internal memory bandwidth utilization and reduced data movement between memory and processor. In this talk, we introduce our DRAM-based memory-coupled computing solutions. We also provide examples of how these technologies can improve the performance and energy efficiency of large-scale applications, including transformer-based neural networks and deep-learning-based recommendation systems.

Technical Talk - AMD
Michael Schulte
Senior Fellow Design Engineer, AMD

“Big Compute: Architecting our Future”

Big Compute comprises techniques and technologies from artificial intelligence, data analytics, and high-performance computing to solve previously intractable problems. In this talk, I’ll describe key technologies for Big Compute and discuss how the availability of massive amounts of data is changing the compute landscape. I’ll also give some examples of applications enabled by Big Compute and describe how we can use Big Compute to build a better future.

iMAGiNE Thrust Overviews

Cloud - Lizy John, Faculty Lead

Edge - Radu Marculescu, Faculty Lead

Applications - Christine Julien, Faculty Lead

Panel

“Enabling Intelligence in the Edge to Cloud Continuum: Challenges and Next Steps”

iMAGiNE Faculty Moderator:
Diana Marculescu

Faculty Panelists:
Mattan Erez
Andreas Gerstlauer
Lizy John
Radu Marculescu

PM Keynote - Amazon
Randy Huang
Principal Engineer at Annapurna Labs, an Amazon Company

“Optimizing ML workloads with AWS Inferentia & Trainium”

In this talk, we detail the AWS Inferentia and Trainium systems and show key value proposition for our customers.

Technical Talk - Amazon
Rana Ali Amjad
Senior Applied (Research) Scientist, Amazon

“Large model training”

In this talk we will focus on how Amazon search trained 100B parameter transformer-based encoder models efficiently. We will illustrate our journey of incrementally scaling to larger models and more data, introducing new techniques as needed to either scale up or improve efficiency of the training.

Technical Talk - Microsoft
Derek Chiou
Partner Hardware Architect, Microsoft

“Infrastructure Offload in Microsoft Azure”

Running a public cloud requires virtualizing tenants networking and storage activity. Virtualization can be done in software running on the host server, alongside the tenant VMs. Doing so, however, consumes many cores that could otherwise be rented to tenants. This talk will Azure’s approach to offloading such virtualization infrastructure to SmartNICs and some of the tradeoffs that were considered.