Student Spotlight on Varun Parekh: Divide and conquer: optimizing efficiency and workload distribution for artificial intelligence models

Varun's research focuses on designing energy-efficient hardware accelerators for deep learning, with a particular emphasis on chiplet-based architectures and optimizing resource utilization for next-generation AI workloads.

Varun profile picture professional

Group projects in school, call centers with multiple agents managing a constant stream of incoming calls, and companies using multiple warehouses across the country to fulfill orders are all examples of workload partitioning in our everyday lives. Balancing these workloads is crucial to ensure that no single node in any pipeline becomes overburdened with tasks.

Workload partitioning is also important for computer algorithms, like those that drive AI systems like ChatGPT. “As generative AI models continue to rise in popularity and use, the need for more memory and power rises.” Varun Parekh said, a PhD student in Computer Science and Engineering at Penn State's Microsystems Design Lab under Dr. Vijaykrishnan Narayanan. The large amounts of energy consumed to produce the rapid result for the user results in massive temperature changes to the computers that these models are run on. Therefore, it is imperative to distribute the workload of these AI models in such a way so that the temperature rise is balanced while still maintaining efficiency.

“Computer chips are made up of many microchips called chiplets. My research looks at different chiplet designs and how they can be physically arranged on a substrate to better distribute heat during the AI model workloads,” Varun explained. In other words, Varun Parekh is working on distributing these workloads through optimizing the architecture and thermal distribution of the devices running these AI models. "When designing chiplet architectures for AI models, it's not just about maximizing performance—it's about managing the thermal impact of heavy computational loads," explains Varun, "Our focus is on creating thermally aware chiplet designs that distribute workloads efficiently, balancing power and temperature to enhance both sustainability and the longevity of the hardware."

Varun’s current PhD research explores the changes and optimizations that contribute to greater thermal distribution and efficiency when computers run AI models. His work is part of a collaborative project that includes researchers from devices, AI and material disciplines, all working on different aspects of optimizing and creating energy-efficient AI models, from detailed computer architecture and computer science to device applications. In a world so rapidly changing technology-wise, Varun serves as an interface between the application and the people working on the devices. “In a way, I’m serving as a bridge between the fundamental physics and applied parts of the project,” Varun said about his role in the collaborative project.

Varun’s interest in thermal analysis began during his undergraduate studies in mechanical engineering major where he learned how different metals can withstand different temperatures and how these temperatures of the metal affect performance and reliability of electronic systems. His later studies in computer science led him to begin working on AI models, which quickly showed him how much energy these models consume. Combining these two experiences, Varun discovered the field of computer architecture, particularly its role in optimizing the energy consumption and thermal efficiency.

As AI rapidly is becoming part of everyday life, Varun’s work provides opportunities for more sustainable use of AI systems. “By changing the structure or materials used in chips, we can optimize the resources that a computer uses to run these AI models, resulting in AI models being able to run in resource limited scenarios and broader distribution of its use in energy-efficient ways,” says Varun.

While Varun is focused on enhancing the design and efficiency of hardware systems to handle larger AI workloads, he also takes a special interest in science policy, minoring in industry law and serving as the Secretary and Advocacy Co-Chair of Penn State’s Science Policy Society. His hopes after his PhD are to increase collaborative global research and development along with AI governance and hopes that he will be able to guide science policy through the lense of his research contributing to more equitable AI technology.

________

Varun Parekh is a PhD student in the Department of Computer Science and Engineering in the Microsystems Design Lab supervised by Dr. Vijaykrishnan Narayanan. This work is funded in part by Semiconductor Research Cooperation, National Science Foundation and Center for 3D Ferroelectric Microelectronics at Penn State.

This article was written by Avalon Miller, a PhD student mentored by Christina Rosa in the Department of Plant Pathology and Environmental Microbiology. It was prepared as part of the INSECT NET Science Communication workshop series coordinated by Drs. Christina Grozinger and Natalie Boyle in summer 2024.