Virtualizing Deep Neural Networks for Memory-Efficient Neural Network Design
NVIDIA
arXiv:1602.08124 [cs.DC], (25 Feb 2016)
@article{rhu2016virtualizing,
title={Virtualizing Deep Neural Networks for Memory-Efficient Neural Network Design},
author={Rhu, Minsoo and Gimelshein, Natalia and Clemons, Jason and Zulfiqar, Arslan and Keckler, Stephen W.},
year={2016},
month={feb},
archivePrefix={"arXiv"},
primaryClass={cs.DC}
}
The most widely used machine learning frameworks require users to carefully tune their memory usage so that the deep neural network (DNN) fits into the DRAM capacity of a GPU. This restriction hampers a researcher’s flexibility to study different machine learning algorithms, forcing them to either use a less desirable network architecture or parallelize the processing across multiple GPUs. We propose a runtime memory manager that virtualizes the memory usage of DNNs such that both GPU and CPU memory can simultaneously be utilized for training larger DNNs. Our virtualized DNN (vDNN) reduces the average memory usage of AlexNet by 61% and OverFeat by 83%, a significant reduction in memory requirements of DNNs. Similar experiments on VGG-16, one of the deepest and memory hungry DNNs to date, demonstrate the memory-efficiency of our proposal. vDNN enables VGG-16 with batch size 256 (requiring 28 GB of memory) to be trained on a single NVIDIA K40 GPU card containing 12 GB of memory, with 22% performance loss compared to a hypothetical GPU with enough memory to hold the entire DNN.
March 1, 2016 by hgpu
