Soumya Joshi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Soumya Joshi has a total of 40 co-authors. Here are the top co-authors by number of shared publications:
- Linga Reddy Cenkeramaddi
University of Agder
5 shared publications - Mekala Bindu Bhargavi
BITS Pilani-Hyderabad Campus
4 shared publications - Jitesh Choudhary
Centre for Development of Advanced Computing, Guwahati, India
3 shared publications - M. B. Srinivas
Department of Electrical and Electronics Engineering, Birla Institute of Technology and Science—Pilani, Hyderabad 500078, India
3 shared publications - Sri Parameswaran
The University of Sydney
3 shared publications - Aveen Dayal
Department of Artificial Intelligence, Indian Institute of Technology at Hyderabad, Hyderabad, India
2 shared publications
Latest Publications
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 434-437
BRILU: BRAM-Optimized LU Decomposition With RISC-V-Based Hardware Accelerator
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 434-437
BRILU: BRAM-Optimized LU Decomposition With RISC-V-Based Hardware Accelerator
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 476-479
Enhancing Branch Prediction Accuracy and Efficiency in RISC-V Processors Using NovelBP
Branch Prediction (BP) is a critical technique in modern processors to minimize pipeline stalls and enhance performance. Conventional predictors like 2-bit saturating counters work well for simple patterns but struggle with complex branch behaviors. This work presents NovelBP, a state-of-the-art perceptron-based branch predictor enhanced with XOR-based hashing and loopback buffers, implemented in a 5-stage in-order RISC-V pipelined processor. Designed in Verilog and validated on an FPGA, NovelBP shows improved prediction accuracy by 2–3% while reducing power consumption. Analysis using MPKI (Mispredictions Per Kilo Instructions) and CycWpPKI (Cycles Wasted Per Kilo Instructions) metrics confirms its scalability and effectiveness in branch-intensive workloads, making it suitable for real-time applications. The results highlight NovelBP as a scalable and efficient branch prediction solution.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 476-479
Enhancing Branch Prediction Accuracy and Efficiency in RISC-V Processors Using NovelBP
Branch Prediction (BP) is a critical technique in modern processors to minimize pipeline stalls and enhance performance. Conventional predictors like 2-bit saturating counters work well for simple patterns but struggle with complex branch behaviors. This work presents NovelBP, a state-of-the-art perceptron-based branch predictor enhanced with XOR-based hashing and loopback buffers, implemented in a 5-stage in-order RISC-V pipelined processor. Designed in Verilog and validated on an FPGA, NovelBP shows improved prediction accuracy by 2–3% while reducing power consumption. Analysis using MPKI (Mispredictions Per Kilo Instructions) and CycWpPKI (Cycles Wasted Per Kilo Instructions) metrics confirms its scalability and effectiveness in branch-intensive workloads, making it suitable for real-time applications. The results highlight NovelBP as a scalable and efficient branch prediction solution.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 488-491
Enhancing Functional Coverage Closure in Network-on-Chip Systems With Reinforcement Learning
As the demand for designing and validating complex technologies increases, the need for computationally efficient methodologies becomes crucial. Network-On-Chip (NoC) verification, with its vast architecture of numerous routers and links, presents significant challenges when performed using traditional methods, delaying the entry of products into the market. This paper presents a streamlined and lightweight solution that uses Reinforcement Learning to automate test vector generation. By optimizing the verification process, this approach significantly speeds up coverage closure, ensuring thorough validation and improving overall efficiency.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Embedded Systems Letters
- Vol. 18 (5) , 488-491
Enhancing Functional Coverage Closure in Network-on-Chip Systems With Reinforcement Learning
As the demand for designing and validating complex technologies increases, the need for computationally efficient methodologies becomes crucial. Network-On-Chip (NoC) verification, with its vast architecture of numerous routers and links, presents significant challenges when performed using traditional methods, delaying the entry of products into the market. This paper presents a streamlined and lightweight solution that uses Reinforcement Learning to automate test vector generation. By optimizing the verification process, this approach significantly speeds up coverage closure, ensuring thorough validation and improving overall efficiency.
- Open Access
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 43113-43135
A Scalable Fault-Tolerant and Congestion-Aware Q-Learning Routing Framework for Network-on-Chip Systems
Efficient and adaptive routing is critical to the performance and reliability of modern Network-on-Chip (NoC) architectures.With the increasing size of NoC meshes, it has become essential that routing algorithms in modern NoCs must be inherently fault-aware and capable of dynamically adapting their routing decisions. This paper presents a Q-learning-based optimal-path routing framework for fault-aware Network-on-Chip (NoC) architectures. The proposed method formulates NoC routing as a Markov Decision Process and introduces a compact, state representation with a reward function designed to be adaptive, enabling a light Q-table which is independent of the size of the mesh, with low memory overhead suitable for hardware deployment. The routing policy is trained offline on a 5 × 5 mesh and evaluated during inference on larger meshes (8×8, 12×12, and 16×16) without retraining. Inference-based results demonstrate comparison of these mesh sizes on the basis of the delivery rate and Hop-Counts required to reach the destination. For hardware-based implementation, a state computation unit has been designed, and the trained Q-table is utilised to test the feasibility of the proposed algorithm on FPGA boards. The simulation results validate that a fault-aware, scalable, optimal, and practically feasible routing algorithm has been proposed.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 43113-43135
Open AccessA Scalable Fault-Tolerant and Congestion-Aware Q-Learning Routing Framework for Network-on-Chip Systems
Efficient and adaptive routing is critical to the performance and reliability of modern Network-on-Chip (NoC) architectures.With the increasing size of NoC meshes, it has become essential that routing algorithms in modern NoCs must be inherently fault-aware and capable of dynamically adapting their routing decisions. This paper presents a Q-learning-based optimal-path routing framework for fault-aware Network-on-Chip (NoC) architectures. The proposed method formulates NoC routing as a Markov Decision Process and introduces a compact, state representation with a reward function designed to be adaptive, enabling a light Q-table which is independent of the size of the mesh, with low memory overhead suitable for hardware deployment. The routing policy is trained offline on a 5 × 5 mesh and evaluated during inference on larger meshes (8×8, 12×12, and 16×16) without retraining. Inference-based results demonstrate comparison of these mesh sizes on the basis of the delivery rate and Hop-Counts required to reach the destination. For hardware-based implementation, a state computation unit has been designed, and the trained Q-table is utilised to test the feasibility of the proposed algorithm on FPGA boards. The simulation results validate that a fault-aware, scalable, optimal, and practically feasible routing algorithm has been proposed.
- Open Access
- 5 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 5746-5763
Scalable Network-on-Chip Design for FPGA Implementation
Network-on-Chip (NoC) technology has become a fundamental solution to address communication bottlenecks and scalability challenges in modern System-on-Chip (SoC) architectures. This paper presents a fully automated and scalable NoC design framework capable of generating synthesizable Verilog code for diverse topologies, including Mesh, Torus, Butterfly Fat Tree (BFT), and Custom configurations. The framework, based on a wormhole router architecture, automatically configures interconnections, routing paths, and network parameters according to user specifications, eliminating manual intervention and reducing design time errors. The generated NoC architectures were synthesized and verified on a Xilinx Virtex-7 Field Programmable Gate Array (FPGA) using Vivado, demonstrating design adaptability and efficiency. Experimental results show up to ∼ 12, 500 speed-up for Mesh, ∼ 11, 500 for Torus, ∼ 10, 800 for BFT, and ∼ 10, 200 for Custom topologies compared with manual approaches. The proposed design was evaluated for throughput, latency, packet delivery ratio (PDR), and energy per packet. Results show that the BFT topology delivers the highest PDR (98.2%) and the lowest energy consumption (1.80 nJ), confirming the framework scalability. Overall, the AutoNoC framework offers a reconfigurable, topology-independent, and FPGA-compatible solution for scalable SoC communication.
- 5 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 5746-5763
Open AccessScalable Network-on-Chip Design for FPGA Implementation
Network-on-Chip (NoC) technology has become a fundamental solution to address communication bottlenecks and scalability challenges in modern System-on-Chip (SoC) architectures. This paper presents a fully automated and scalable NoC design framework capable of generating synthesizable Verilog code for diverse topologies, including Mesh, Torus, Butterfly Fat Tree (BFT), and Custom configurations. The framework, based on a wormhole router architecture, automatically configures interconnections, routing paths, and network parameters according to user specifications, eliminating manual intervention and reducing design time errors. The generated NoC architectures were synthesized and verified on a Xilinx Virtex-7 Field Programmable Gate Array (FPGA) using Vivado, demonstrating design adaptability and efficiency. Experimental results show up to ∼ 12, 500 speed-up for Mesh, ∼ 11, 500 for Torus, ∼ 10, 800 for BFT, and ∼ 10, 200 for Custom topologies compared with manual approaches. The proposed design was evaluated for throughput, latency, packet delivery ratio (PDR), and energy per packet. Results show that the BFT topology delivers the highest PDR (98.2%) and the lowest energy consumption (1.80 nJ), confirming the framework scalability. Overall, the AutoNoC framework offers a reconfigurable, topology-independent, and FPGA-compatible solution for scalable SoC communication.
- Open Access
- 27 October 2025
- Published by Springer Nature in Discover Computing
- Vol. 28 (1) , 1-20
Enhancing fault-tolerant application mapping in network-on-chip through transformer network based reinforcement learning approach
Network-on-Chip (NoC) is a promising technology that uses a network of interconnected processing elements to improve performance and reduce communication costs in complex systems. However, faults can occur in NoC systems due to various factors such as manufacturing defects, aging, and environmental factors. To address this challenge, fault-tolerant techniques are necessary to ensure the reliable and efficient operation of NoC systems. One such technique is fault-tolerant application mapping, in which application cores are assigned to processing elements to reduce the system performance degradation caused by the failures. However, the application mapping for NoC systems, including fault tolerance, is a complex task. This paper presents a novel approach consisting of a transformer-based reinforcement learning algorithm for application mapping in mesh topology based NoC. This allows NoC to communicate efficiently in the presence of faults. The proposed approach is compared with other state-of-the-art techniques. The findings demonstrate that the proposed approach outperforms the average communication cost encompassing 2, 5, and 7 percent more than RL-MAP, NMA, and PSO, respectively, and also has less run-time. Also, it is observed that the introduced technique excels against the remaining techniques with respect to the average network latency decrease by 3, 5, and 6 percent for RL-MAP, NMA, and PSO correspondingly. The approach in this method is thus helpful in the design of NoC systems that are energy efficient as well as reliable for targeted applications.
- 27 October 2025
- Published by Springer Nature in Discover Computing
- Vol. 28 (1) , 1-20
Open AccessEnhancing fault-tolerant application mapping in network-on-chip through transformer network based reinforcement learning approach
Network-on-Chip (NoC) is a promising technology that uses a network of interconnected processing elements to improve performance and reduce communication costs in complex systems. However, faults can occur in NoC systems due to various factors such as manufacturing defects, aging, and environmental factors. To address this challenge, fault-tolerant techniques are necessary to ensure the reliable and efficient operation of NoC systems. One such technique is fault-tolerant application mapping, in which application cores are assigned to processing elements to reduce the system performance degradation caused by the failures. However, the application mapping for NoC systems, including fault tolerance, is a complex task. This paper presents a novel approach consisting of a transformer-based reinforcement learning algorithm for application mapping in mesh topology based NoC. This allows NoC to communicate efficiently in the presence of faults. The proposed approach is compared with other state-of-the-art techniques. The findings demonstrate that the proposed approach outperforms the average communication cost encompassing 2, 5, and 7 percent more than RL-MAP, NMA, and PSO, respectively, and also has less run-time. Also, it is observed that the introduced technique excels against the remaining techniques with respect to the average network latency decrease by 3, 5, and 6 percent for RL-MAP, NMA, and PSO correspondingly. The approach in this method is thus helpful in the design of NoC systems that are energy efficient as well as reliable for targeted applications.
- 20 December 2024
- Published by Springer Nature
- p. 133-146
DAN: Deep Neural Network-Based Application Mapping for Optimized Network-on-Chip Design
- 20 December 2024
- Published by Springer Nature
- p. 133-146
DAN: Deep Neural Network-Based Application Mapping for Optimized Network-on-Chip Design
- 17 December 2022
- Published by Springer Nature
- p. 277-291
MANA:Multi-Application Mapping onto Mesh Network-on-Chip using ANN
- 17 December 2022
- Published by Springer Nature
- p. 277-291
MANA:Multi-Application Mapping onto Mesh Network-on-Chip using ANN
- 17 December 2022
- Published by Springer Nature
- p. 515-526
Implementation and Analysis of Convolution Image Filtering with RISC-V Based Architecture
- 17 December 2022
- Published by Springer Nature
- p. 515-526
Implementation and Analysis of Convolution Image Filtering with RISC-V Based Architecture