Projects
SHAKTIMAAN : An open-source DNN accelerator
Advisor: Prof. Pratyush Kumar
Jun 2019 - May 2020
- Architected a systolic-array based DNN accelerator
- co-processor to SHAKTI processor
Cache coherence protocol for Shakti C-Class processor
Advisor: Prof. Kamakoti
Summer 2018, Spring 2019
Implemented cache-coherence protocol for multi-core SHAKTI processor
- MOESI protocol
- Both snoop-based (Summer 2018) and directory-based (Spring 2019)
Efficient convolution on GPUs by exploiting Quantization
Advisor: Prof. Pratyush Kumar
Fall 2018
Implemented CUDA kernels to efficiently perform convolution in GPUs by Quantization. In quantized convolutional filter maps, weight values are redundant and can be shared, thereby reducing memory used. Overall, it was found that weight sharing did not provide any speedup over baseline model, owing to increase in book-keeping instructions needed for execution of convolution.