Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks |
Presents Eyeriss, a novel hardware accelerator designed for deep convolutional neural networks (CNNs) with a focus on energy efficiency |
The authors recognize the growing demand for efficient hardware to accelerate deep CNNs due to their increasing use in applications like image and speech recognition. Eyeriss is designed to address this need. |
-
Reconfigurable Architecture: Eyeriss is built as a reconfigurable accelerator, enabling it to support different network configurations, layer sizes, and data types. This flexibility makes it suitable for a wide range of CNN models.
-
Energy Efficiency: The paper highlights Eyeriss' energy-efficient design, which is achieved through careful consideration of data movement and computation. The authors discuss techniques for minimizing data movement and exploiting data locality to reduce energy consumption.
-
Memory Hierarchy: Eyeriss employs a memory hierarchy that efficiently stores and accesses intermediate data to minimize the energy required for data transfer.
|
The authors provide an evaluation of Eyeriss, comparing its energy efficiency and performance to other state-of-the-art CNN accelerators. They demonstrate that Eyeriss achieves competitive performance while consuming significantly less energy. |
The paper's contribution lies in introducing a reconfigurable accelerator architecture that is well-suited for deep CNNs, emphasizing energy efficiency, and demonstrating its practical effectiveness. |