1. events
  2. all
  3. 2025

Jan 24 (Fri) @ 10:00am: "The Upcoming Revolution of General-Purpose Computing: Democratizing AI/ML Accelerators,” Hung-Wei Tseng, UC Riverside

Date and Time
photo of h-w Tseng

Location: Engineering Science Building (ESB), Room 2001

Abstract

The rise of AI/ML has spurred the development of specialized accelerators, transforming the computing landscape. While these accelerators excel in AI/ML workloads, their potential extends beyond this domain. Tensor processing, the core of these accelerators, can be harnessed for a broader spectrum of applications. By democratizing AI/ML accelerators, we aim to accelerate scientific discovery, enable new applications in various domains, and ultimately revolutionize general-purpose computing. However, current designs often limit their applicability due to their specialized nature.

This talk explores our research efforts to democratize AI/ML accelerators for general-purpose computing. We focus on three key areas:

  1. Rethinking Applications and Systems: We demonstrate how to expose the features of AI/ML accelerators to general-purpose programming frameworks and applications. Our "Simultaneous and Heterogenous Multithreading" project concurrently utilizes diverse accelerators (GPUs, TPUs), achieving a 1.95x speedup for linear algebra applications.
  2. Rethinking Algorithms: We showcase how to adapt existing algorithms to leverage the unique capabilities of AI/ML accelerators. Our TCUDB work exemplifies this, achieving a 288x speedup for database join operations by exploiting NVIDIA's tensor cores.
  3. Rethinking Architectures: We investigate architectural extensions to broaden the scope of applications supported by AI/ML accelerators. We demonstrate how extending support for more matrix operations can yield significant performance gains (over 10x) for dynamic programming algorithms.

We will discuss the challenges and opportunities in democratizing AI/ML accelerators, drawing parallels to the successful story of GPGPUs. Finally, we will explore the potential of democratizing other emerging hardware accelerators.

Bio

Hung-Wei is an associate professor in the Department of Electrical and Computer Engineering at the University of California, Riverside. He is now leading the Extreme Scale & Computer Architecture Laboratory and focusing on accelerating applications through generalized computing on tensor processors and other emerging accelerators. Hung-Wei's research has been recognized by IEEE Micro's "Top Picks from Computer Architecture" in 2024 and 2020 and the Facebook faculty research award in 2017. He got his PhD from the Department of Computer Science and Engineering at the University of California, San Diego.

Hosted by: The ECE Department

Submitted by: Megan Ashley <mmashley@ucsb.edu>