Marzieh Lenjani

Rethinking Control, Access, Communication, and Partitioning Techniques for PIM-based Accelerators

Photo of Marzieh Lenjani

Marzieh Lenjani
Apple / University of Virginia


Recent studies show that the number of instructions required per each arithmetic operation and communication overhead can limit the benefits of PIM. In my talk, I argue that rethinking and simplifying the control, access, and communication mechanisms of PIM units can alleviate these overheads. To this end, first, I discuss the inefficacy of traditional SISD and SIMD/SIMT architectures for data-intensive applications, with few computations per data element and challenges due to control and access divergence. Then, I introduce our proposed architectures, Fulcrum and FulcrumV2 (Gearbox), that offer a trade-off between (i) full control and access divergence support in SIMD and (ii) no/costly control or divergence support in SIMD/SIMT approaches. We show the flexibility of our design by mapping important kernels from different domains, such as machine learning and graph processing. Finally, I conclude with our vision for identifying common requirements of data-intensive applications and designing a semi-general-purpose PIM-based accelerator that supports a wide range of applications.


Dr. Lenjani’s presentation is based on her Ph.D. work at the University of Virginia and does not represent her work at Apple.


Marzieh Lenjani received her Ph.D. from the University of Virginia, where she was advised by Professor Kevin Skadron. Marzieh’s research explores accelerator design for data-intensive applications, as well as the characterization and optimization of applications on emerging devices. Her research has resulted in several publications at premier computer architecture venues (such as ISCA, HPCA, ASPLOS, IISWC, and IEEE Micro), one published patent, and three patent applications (funded by SRC). Marzieh’s work has been nominated for the best paper award at HPCA’20 and IISWC’19. She has also received the John A. Stankovic Graduate Research Award from the Computer Science Department at the University of Virginia for outstanding research.