Best Paper Award at ISLPED 2026 for Farzad Razi

A paper co-authored by a cross-institutional team of researchers was recently recognized with the Best paper Award at the International Symposium on Low Power Electronics and Design (ISLPED) 2026 recently held in Evanston, Illinois. The first author of the paper is Farzad Razi, a postdoctoral research associate working with Professor Marc Riedel who is the corresponding author. The other authors included Mehran Moghadam and M. Hassan Najafi* of Case Western Reserve University and Sercan Aygun of University of Louisiana. The paper is titled “MITRA: Reconfigurable, Low-Latency, and Power-Efficient In-Memory Stochastic Architecture for Transcendental Functions.” 

MITRA is a reconfigurable magnetic tunnel junction (MTJ)-based in-memory architecture that combines stochastic bit-stream processing with compact finite state-machines (FSMs) that are embedded in the MTJ structures. The design offers low latency along with power efficient computation, unlike the binary options, thereby offering a robust processing-in-memory (PIM) option to address the data movement bottleneck that exists in contemporary data systems. 

Razi addresses the distinctive nature of MITRA: "What makes MITRA unique is its reconfigurable design, which combines stochastic computing with FSM embedded within magnetic memory. It also uses a fine-grained power-gating technique that limits write activity to the memory states involved in each transition, further improving energy efficiency. For AI applications, the work introduces a novel approximation-aware training method that enables neural networks to learn and compensate for approximation errors arising from stochastic computing and FSMs, thereby maintaining high inference accuracy."

In conventional computers, data must be repeatedly transferred between memory and the processor, creating a memory–compute bottleneck that increases delay and energy use. PIM addresses this problem by performing operations directly where the data is stored. Stochastic computing is particularly well suited to PIM because it represents numbers as bit streams, allowing complex computations to be implemented with significantly simpler circuits than conventional binary computing. Magnetic tunnel junctions (MTJs) are attractive for PIM because they retain data without continuous power, are compatible with CMOS technology, and can support both storage and computation, allowing intermediate results to remain within the memory array and reducing data movement, latency, and energy consumption.

Commenting on the significance of the team's work, Razi says, "Transcendental functions are essential to modern computing, but they are costly to execute using conventional hardware. MITRA uses finite-state machines (FSMs) to perform these computations directly within memory, reducing the need to move data repeatedly between memory and the processor. This approach can lower latency and energy consumption while supporting the growing demands of modern data-intensive systems and edge AI applications.

Read the complete paper at the ACM Digital Library

Farzad Razi earned his Ph.D. in Computer Engineering from the University of Tehran. He is currently a postdoctoral research associate in the Department of Electrical and Computer Engineering at the University of Minnesota Twin Cities. His research focuses on energy-efficient computer architectures, stochastic and in-memory computing, emerging memory technologies, and hardware acceleration for AI and edge computing. 

*M. Hassan Najafi, one of the authors of the paper, is an alumnus of the University of Minnesota Twin Cities. He was a recipient of a Doctoral Dissertation Fellowship for the 2017-2018 academic year and earned his doctoral degree 2018 under the guidance of Professor Emeritus David Lilja. Currently, Najafi is an associate professor with the Department of Electrical, Computer, and Systems Engineering at Case School of Engineering

The research was supported in part by National Science Foundation (NSF), National Aeronautics and Space Administration (NASA), Vernon and Ruby Langlinais Non-Endowed Research Fund, the Lockheed Martin Corporation Endowed Professorship Fund, the NASA award, and gifts from NVIDIA and Google.