Research
Threshold Logic and Artificial Neuron Design
Motivation: Digital design using conventional CMOS technology has been optimized over four decades to improve its performance, power, and area (PPA), leaving little to no room for further improvement. Therefore, the investigation of non-CMOS technology and the development of radically different computing primitives is warranted to design digital systems of the future with high energy efficiency and high throughput.
We investigate threshold logic as an alternative to Boolean logic and its hardware implementation in the form of an Artificial Neuron (AN) to design digital ASICs and custom hardware accelerators.
Related Publications:
Processing In Memory (PIM)
Motivation: The latency and energy consumption of data-intensive or data-centric applications is dominated by the movement of the data between the processor and memory. In modern CPU/GPU systems, about 60% of the total energy is consumed by the data movement over the limited bandwidth channel between the processor and memory. Processing in Memory (PIM) is an effective technique to overcome the memory bottleneck and provide an energy-efficient platform for execution of the data-intensive applications.
Related Publications:
[1] Gian Singh, Ankit Wagle, Sarma Vrudhula, and Sunil Khatri. “CIDAN: Computing in DRAM with Artificial Neurons.” In 2021 IEEE 39th International Conference on Computer Design (ICCD), 349–56. Storrs, CT, USA: IEEE, 2021. [Video]
[6] Gian Singh, Ayushi Dube, and Sarma Vrudhula. "Energy-Efficient and Low-Latency Computation of Transcendental Functions in a Precision-Tunable PIM Architecture." In IEEE Computer Society Annual Symposium on VLSI, Knoxville, TN, USA, July 1-3, 2024. (accepted paper.)
[7] Gian Singh, Ayushi Dube, and Sarma Vrudhula. "A High Throughput, Energy-Efficient Architecture for Variable Precision Computing in DRAM." IFIP/IEEE International Conference on Very Large Scale Integration, 2024. (Accepted Paper)
CNN Hardware Accelerator
Motivation: Deep neural networks (DNNs) have become the dominant framework for machine learning and artificial intelligence applications. The significant energy consumption and the massive environmental impact of training and inference of such DNNs may soon overshadow their societal benefits. Deploying these networks on battery-constrained edge devices requires a highly energy-efficient platform.
Related Publications: