Transcription of Eyeriss: A Spatial Architecture for Energy-Efficient ...
{{id}} {{{paragraph}}}
eyeriss : A Spatial Architecture for Energy-Efficient Dataflowfor Convolutional Neural NetworksYu-Hsin Chen , Joel Emer and Vivienne Sze EECS, MITC ambridge, MA 02139 NVIDIA Research, NVIDIAW estford, MA 01886 {yhchen, jsemer, Deep convolutional neural networks (CNNs) arewidely used in modern AI systems for their superior accuracybut at the cost of high computational complexity. The complex-ity comes from the need to simultaneously process hundredsof filters and channels in the high-dimensional convolutions,which involve a significant amount of data movement. Althoughhighly-parallel compute paradigms, such as SIMD/SIMT, effec-tively address the computation requirement to achieve highthroughput, energy consumption still remains high as datamovement can be more expensive than computation. Accord-ingly, finding a dataflow that supports parallel processing withminimal data movement cost is crucial to achieving energy-efficient CNN processing without compromising this paper, we present a novel dataflow, calledrow-stationary(RS), that minimizes data movement energy con-sumption on a Spatial Architecture .}
ative analysis of the energy costs associated with data movement and the impact of different types of data reuse. (Section VII) II. SPATIAL ARCHITECTURE Spatial architectures (SAs) are a class of accelerators that can exploit high compute parallelism using direct commu-nication between an array of relatively simple processing engines (PEs).
Domain:
Source:
Link to this page:
Please notify us if you found a problem with this document:
{{id}} {{{paragraph}}}