Dr. Joshua Booth, Assistant Professor of the University of Alabama in Huntsville's Computer Science Department, was awarded a 5-year National Science Foundations CAREER grant to start on April 1, 2021. The total intended award amount is $479,230.

This project is jointly funded by CAREER Software and Hardware Foundations HPC program and the Established Program to Stimulate Competitive Research (EPSCoR).

Fast, Energy Efficient Irregular Kernels via Neural Acceleration hopes to improve upon high performance computing (HPC), reducing development time and costs while supporting graduates and undergraduates as they engage in cross-disciplinary involvement.

HPC is an area of computer science that tries to utilize the most computational power out of computer systems. Long-running simulations, such as climate simulation and drug analysis, need to utilize all the computational power possible, and many of these simulations require algorithms that computer science categorizes as irregular kernels. Irregular kernels get their name due to the irregular nature they access memory, and irregular kernels are known for performance issues even with hand-tuning by experts. At the same time, the computer system architecture, i.e., the pieces of hardware we use to run computations, are changing. The common general processor in most people's laptops and computers are evolving into a set of specialized hardware such as GPU and neural devices. This specialized hardware tends to have optimal performance for its particular computational area. However, no such specialization hardware really exists in the mainstream for irregular kernels. However, neural networks, which can be run on specialized neural devices, are a popular computing model. There exists a number of overlapping computing characteristics between many irregular kernels and neural networks.

This work is looking at transforming irregular kernels (or at least pieces of them) into neural networks that can be run on these specialized neural devices. In addition to possible performance gains, this style of computing irregular kernels could be done in less time and by modern computer scientist that commonly have a strong background in neural networks.