Traditionally, the Central Processing Unit or CPU has been the brain of a computer. In recent times, however, Graphics Processing Unit or GPUs are emerging as a platform of choice for massively parallel computations, such as deep learning. But, writing a program for GPU is significantly harder than writing one for the CPU. Further, while a significant fraction of modern computing today happens in the public cloud computing infrastructures (e.g., Amazon’s EC2), GPUs are not easy to deploy to the cloud. These present significant predicaments in fully harnessing compute capabilities of modern GPUs.
Arkaprava and his team’s focus has been to make GPUs more easily programmable and deployable in the cloud platforms through co-design of the hardware and the software. More recently, they are working on how the hardware can provide feedback to the programmer to help her to write correct GPU programs.
Arkaprava Basu, Eric Van Tassell, Mark Oskin, Guilherme Cox, Gabriel Loh. “Single instruction multiple data page table walk scheduling at input output memory management unit”. US Patent App. 15/852,442, 2019.
Seunghee Shin, Michael LeBeane, Yan Solihin, and Arkaprava Basu. “Neighborhood-aware address translation for irregular GPU applications”. In Proceedings of the 51st Annual IEEE/ACM International Symposium on Microarchitecture (MICRO-51). Fukuoka, Japan — October 20 – 24, 2018, pp 352-363.