REACT: A Heterogeneous Reconfigurable Neural Network Accelerator with Software-Configurable NoCs for Training and Inference on Wearables
TimeTuesday, July 12th11:15am - 11:37am PDT
Location3007, Level 3
Event Type
Research Manuscript
In-Package and On-Chip Communication and Networks-on-Chip
Physical Design and Verification, Lithography and DFM
DescriptionThere is a need for on-chip learning to improve model accuracy on personalised user data and preserve privacy. This work proposes REACT, an AI accelerator for wearables that has heterogeneous GIGA and nano cores supporting both training and inference. REACT’s architecture is fully distributed and NoC-centric, with 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. REACT’s online learning realises upto 10% accuracy improvement, and is upto 55x faster and 520x more energy efficient than state-of-the-art accelerators with similar memory and computation footprint.