AI3

Algorithmic Innovations for Accelerated Implementations of Artificial Intelligence (AI3) focuses on developing algorithmic design choices for Artificial Intelligence/Machine Learning (AI/ML) models, analyze their tradeoffs, and determine optimal choices for mapping AI applications to heterogeneous platforms (CPU+AI accelerators such as GPUs, and FPGAs).

Projects under this category focus on improving the runtime performance --- execution time, energy, throughput, latency, or a combination of these metrics, of AI algorithms. They focus on developing, evaluating, and selecting algorithmic and application mapping design choices to obtain optimal mapping of AI applications to heterogeneous platforms: CPU + AI accelerators such as GPUs, FPGAs, and near-memory processors.

Energy-efficient, Near-Memory CMOS+X Architecture for Hardware Acceleration of DNNs with Application to NextG Wireless Systems

This is a collaborative project between Prof. Sarma Vrudhula (Arizona State University) and Prof. Marwan M Krunz (University of Arizona). The project focuses on developing a hardware/software co-design framework titled vMC, that will enable development of low-latency, energy-efficient DNN inference accelerators. To evaluate and demonstrate the real-world applicability of our proposed project, the project will use DNN architectures for NextG wireless systems as our driving application. PI Kuppannagari’s focus is to develop novel algorithms and application mapping optimizations for 3D Convolutional Neural Network (CNN) algorithms.

Funding Source: National Science Foundation Fuse2 Program

Portable Library for Reinforcement Learning on Heterogeneous Cloud Cyberinfrastructure

This project is in collaboration with Prof. Viktor Prasanna (University of Southern California). It focuses on creating a user friendly library for deploying Deep Reinforcement Learning applications on a cloud of cluster consisting of CPU+Accelerator nodes.

Funding Source: National Science Foundation CSSI Program