Georgia Tech School of Electrical and Computer Engineering (ECE) Ph.D. student Ritik Raj received the IBM Ph.D. Fellowship for his work on neural processing units (NPU) design optimization.

The IBM fellowship program is a highly selective, invite-only award that supports exceptional Ph.D. students in their areas of research.

“I am grateful and humbled to receive the IBM PhD fellowship award,” Raj said. “This award, along with IBM mentorship, will help me drive my research more into agentic AI efficiency and design optimization for NPUs.”

Raj’s work focuses on neurosymbolic artificial intelligence (AI) workloads run on NPUs —specialized hardware accelerators designed to speed up neural-network computations.

 This combines learning-based machine learning (ML) models with more structured, logic-like reasoning. Neurosymbolic AI aims to mimic the “thinking fast and slow” reasoning processes of biological entities. 

Though promising, neurosymbolic AI workloads can be inefficient and hard to evaluate on today’s NPUs, which are typically optimized for dense neural-network computation rather than memory-bound, control-heavy symbolic reasoning.

Raj’s research addresses these challenges in two ways:

First is with CogSys, a novel co-design framework dedicated specifically to neurosymbolic AI systems. The key optimization is a customized processing element and dataflow that makes NPU efficient and reconfigurable for running both conventional AI (neural networks) and symbolic AI. It has achieved promising results, with a 75 times faster speed up against traditional NPUs and enables real-time abstract reasoning. 

The other is with cycle-accurate simulation, called SCALE-Sim v3, which helps computer architects to assess various aspects of NPU performance, including latency and energy, before fabrication.

SCALE-Sim v3 adds five features to its predecessor: multi-core simulation, support for sparse matrix multiplications (SpMM), detailed DRAM analysis, precise data layout modeling, and energy estimation. These improvements enable deeper end-to-end system analysis for next-gen AI accelerators, accommodating a wide variety of systems and workloads and providing detailed full-system insights into latency, bandwidth, and power efficiency. 

Additionally, Raj works on improving the efficiency of agentic AI systems, which add a decision-making layer on top of monolithic AI models, allowing them to plan and take actions.

With the fellowship, he will receive support from IBM, which is the largest industrial research organization in the world, both financially and in the form of a mentor.

Raj is advised by ECE Associate Professor Tushar Krishna. He received his bachelor’s in technology in electronics and communication engineering from the Indian Institute of Technology (IIT), Roorkee, India, in 2023.