REACT: A Heterogeneous Reconfigurable Neural Network Accelerator with Software-Configurable NoCs for Training and Inference on Wearables
Welcome to the REACT Project website!!
REACT is a NoC-based AI accelerator for wearables that has heterogeneous cores supporting both training and inference. REACT’s architecture being NoC-centric, has the weights, features and gradients distributed across cores, accessed and computed efficiently through software-configurable NoCs. Unlike conventional dynamic NoCs, REACT’s NoCs have no buffer queues, flow control or routing, as they are entirely configured by software for each neural network. This allows to perform online learning for CNN and MLP networks with significalty higher performance and energy-efficiency than state-of-the-art accelerators with similar memory and computation footprint.
As seen above, REACT is a heterogeneous architecture with two types of Cores wihch can efficiently map Convolution layers and Fully Connected layers. The architecture also allows for performing on-chip learning for fully connected layers which allows to increase accuracy of the CNN models by online learning using the on-chip data.
Publication
"REACT : A heterogeneous reconfigurable neural network accelerator with software-configurable NoCs for training and inference on wearables", Mohit Upadhyay, Rohan Juneja, Bo Wang, Jun Zhou, Weng-Fai Wong, and Li-Shiuan Peh, Proceedings of the 59th ACM/IEEE Design Automation Conference (DAC '22) [Paper: link, PDF]
TEAM
CODE Repository
To acces the REACT architecture (System Verilog) codebase, please contact Mohit Upadhyay (mohitu@u.nus.edu)