Mahesh Subedar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mahesh Subedar has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Charles Hong
UC Berkeley, United States of America
1 shared publication - Chien-Yi Yang
University of California San Diego, La Jolla, USA
1 shared publication - Flavio Ponzina
San Diego State University, San Diego, CA, USA
1 shared publication - Georges Gielen
ESAT-MICAS, KU Leuven, Leuven, Flemish Brabant, Belgium
1 shared publication - Grace Dinh
UC Berkeley, United States of America
1 shared publication - Jiacong Sun
Department of Electrical Engineering (ESAT), Microelectronics Circuits and Systems (MICAS) Division, KU Leuven, Leuven, Belgium
1 shared publication
Latest Publications
- 23 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (5) , 2222-2235
SunPar : An Analytical Design Space Exploration Framework Modeling Performance Uncertainty in Sparse AI Accelerators
Sparsity-aware techniques, traditionally applied to weight tensors, are increasingly utilized for activations in AI accelerators to enhance the energy efficiency and performance. While sparse accelerators promise significant performance gains, the impact of unpredictable activation sparsity on the performance variation and system uncertainty remains largely overlooked and so far has never been quantitatively analyzed, though it is a critical aspect for time-sensitive applications. Existing analytical models for sparse accelerators fail to capture the performance uncertainty due to their reliance on deterministic modeling of average performance, which assumes consistent hardware behavior and neglects variations arising from diverse sparsity patterns and levels. In this article, we propose SunPar, an open-source modeling framework for benchmarking and evaluating the actual performance variation of sparse AI accelerators under sparsity uncertainty. We categorize uncertainty in the sparsity patterns into two types: intratensor uncertainty (IntraU) and intertensor uncertainty (InterU). By constructing an analytical behavioral model, our framework can evaluate the performance variation in the form of probability density functions (PDFs), providing both the average performance and the performance uncertainty range with different confidence levels. Evaluation of SunPar across four sparse accelerators demonstrates its high accuracy, with a mismatch of <8% in the average performance estimation and <0.2% in performance variance estimation compared with the hardware results. In addition, two case studies provided for real-world applications based on image datasets (ImageNet) reveal a significant divergence between Pareto-optimal hardware choices with and without awareness of the performance uncertainty, underscoring the necessity of uncertainty-aware frameworks for designing sparse AI accelerators in time-sensitive applications.
- 23 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (5) , 2222-2235
SunPar : An Analytical Design Space Exploration Framework Modeling Performance Uncertainty in Sparse AI Accelerators
Sparsity-aware techniques, traditionally applied to weight tensors, are increasingly utilized for activations in AI accelerators to enhance the energy efficiency and performance. While sparse accelerators promise significant performance gains, the impact of unpredictable activation sparsity on the performance variation and system uncertainty remains largely overlooked and so far has never been quantitatively analyzed, though it is a critical aspect for time-sensitive applications. Existing analytical models for sparse accelerators fail to capture the performance uncertainty due to their reliance on deterministic modeling of average performance, which assumes consistent hardware behavior and neglects variations arising from diverse sparsity patterns and levels. In this article, we propose SunPar, an open-source modeling framework for benchmarking and evaluating the actual performance variation of sparse AI accelerators under sparsity uncertainty. We categorize uncertainty in the sparsity patterns into two types: intratensor uncertainty (IntraU) and intertensor uncertainty (InterU). By constructing an analytical behavioral model, our framework can evaluate the performance variation in the form of probability density functions (PDFs), providing both the average performance and the performance uncertainty range with different confidence levels. Evaluation of SunPar across four sparse accelerators demonstrates its high accuracy, with a mismatch of <8% in the average performance estimation and <0.2% in performance variance estimation compared with the hardware results. In addition, two case studies provided for real-world applications based on image datasets (ImageNet) reveal a significant divergence between Pareto-optimal hardware choices with and without awareness of the performance uncertainty, underscoring the necessity of uncertainty-aware frameworks for designing sparse AI accelerators in time-sensitive applications.
- Open Access
- 27 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-9
Multi-Objective Software-Hardware Co-Optimization for HD-PIM via Noise-Aware Bayesian Optimization
- 27 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-9
Open AccessMulti-Objective Software-Hardware Co-Optimization for HD-PIM via Noise-Aware Bayesian Optimization
- Open Access
- 28 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 209-224
DOSA: Differentiable Model-Based One-Loop Search for DNN Accelerators
In the hardware design space exploration process, it is critical to optimize both hardware parameters and algorithm-to-hardware mappings. Previous work has largely approached this simultaneous optimization problem by separately exploring the hardware design space and the mapspace—both individually large and highly nonconvex spaces—independently. The resulting combinatorial explosion has created significant difficulties for optimizers. In this paper, we introduce DOSA, which consists of differentiable performance models and a gradient descent-based optimization technique to simultaneously explore both spaces and identify high-performing design points. Experimental results demonstrate that DOSA outperforms random search and Bayesian optimization by 2.80 × and 12.59 ×, respectively, in improving DNN model energy-delay product, given a similar number of samples. We also demonstrate the modularity and flexibility of DOSA by augmenting our analytical model with a learned model, allowing us to optimize buffer sizes and mappings of a real DNN accelerator and attain a 1.82 × improvement in energy-delay product.
- 28 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 209-224
Open AccessDOSA: Differentiable Model-Based One-Loop Search for DNN Accelerators
In the hardware design space exploration process, it is critical to optimize both hardware parameters and algorithm-to-hardware mappings. Previous work has largely approached this simultaneous optimization problem by separately exploring the hardware design space and the mapspace—both individually large and highly nonconvex spaces—independently. The resulting combinatorial explosion has created significant difficulties for optimizers. In this paper, we introduce DOSA, which consists of differentiable performance models and a gradient descent-based optimization technique to simultaneously explore both spaces and identify high-performing design points. Experimental results demonstrate that DOSA outperforms random search and Bayesian optimization by 2.80 × and 12.59 ×, respectively, in improving DNN model energy-delay product, given a similar number of samples. We also demonstrate the modularity and flexibility of DOSA by augmenting our analytical model with a learned model, allowing us to optimize buffer sizes and mappings of a real DNN accelerator and attain a 1.82 × improvement in energy-delay product.