CNN Accelerator

CNN accelerators are specialized hardware designed to significantly speed up and reduce the energy consumption of convolutional neural network (CNN) inference, crucial for deploying AI on resource-constrained devices. Current research focuses on optimizing various aspects, including efficient data movement through novel memory mapping strategies and dataflow architectures, exploiting sparsity in network weights and activations via pruning and quantization techniques, and developing hardware-aware training methods. These advancements are vital for enabling real-time AI applications in areas like embedded systems and IoT, where low latency and power consumption are paramount.

Papers

May 17, 2022

A Silicon Photonic Accelerator for Convolutional Neural Networks with Heterogeneous Quantization
Febin Sunny, Mahdi Nikdast, Sudeep Pasricha
Convolutional Neural Network Photonic Accelerator CNN Accelerator Silicon Photonics Heterogeneous Quantization

May 2, 2022

Zebra: Memory Bandwidth Reduction for CNN Accelerators With Zero Block Regularization of Activation Maps
Hsu-Tung Shih, Tian-Sheuan Chang
Pruning Method Tiny ImageNet Major Challenge Bottleneck Activation Map CNN Accelerator

March 30, 2022

L^3U-net: Low-Latency Lightweight U-net Based Image Segmentation Model for Parallel CNN Processors
Osman Erman Okman, Mehmet Gorkem Ulkar, Gulnur Selda Uyanik
Neural Network U Net CNN Accelerator Small Object Segmentation

CNN Accelerator

Papers

A Silicon Photonic Accelerator for Convolutional Neural Networks with Heterogeneous Quantization

Zebra: Memory Bandwidth Reduction for CNN Accelerators With Zero Block Regularization of Activation Maps

L^3U-net: Low-Latency Lightweight U-net Based Image Segmentation Model for Parallel CNN Processors