The Best AI Development Tool - Deep Learning Framework Caffe-MPI with the Best Speedup Ratio
BEIJING, Sept. 26, 2017 /PRNewswire/ -- Recently in the 2017 AI Computing Conference (AICC), Mr. Chu Xiaowen, assistant professor at Hong Kong Baptist University, delivered a keynote speech on testing the current mainstream deep learning frameworks and emphasized that compared to other mainstream framework, the open source program Caffe-MPI led by Inspur has the best speedup ratio and achieves the best performance in GoogLeNet network model.
According to the data mentioned in his speech, the Caffe-MPI 2.0 developed by Inspur uses ImageNet data set on 16 P40 GPU in 4 nodes. With the use of GoogLeNet model (Batchsize = 128), training performance of 6 cards reaches 5886 frames per second (FPS), which is 14.2 times higher than a single card. With the use of ResNet model (Batchsize= 32), training performance of 16 cards reaches 1372 FPS, which is 15.34 times higher than a single card. The expansion efficiency reaches 96%. It surpasses CNTK and MXNet to become the deep learning framework having the best speedup ratio and the best performance in the current GoogLeNet models.
The test report from Hong Kong Baptist University
Mr. Chu Xiaowen, also expressed that among the current mainstream deep learning frameworks being tested, Caffe-MPI from Inspur shows the best speedup ratio. It is believed that Caffe-MPI from Inspur will be the best option when users need to expand Caffe to multi-card environment.
The excellent speedup ratio of Caffe-MPI 2.0 comes from the innovation of Parallel Algorithm. Caffe-MPI is designed on two communication models: the inter-card communication within the nodes of GPU and the overall communication within the nodes of RDMA, and they can be realized with the adjustment in NCCL 2.0. This design greatly reduces the pressure on network communication and overcomes the bandwidth imbalance between PCIE and network in traditional communication models, which best suits for the current high-density GPU server.
Meanwhile, Caffe-MPI 2.0 is also designed to achieve deep neural network model with layers of computing and asynchronous communication among them. It combines the data that are waiting to be communicated in different layers and starts to communicate when the quantity reaches a certain amount, effectively avoiding the communication delay when the data quantity is relatively small in the individual layer. Besides, Caffe-MPI 2.0 also provides a better cuDNN compatibility, which means that users can seamlessly employ the latest cuDNN version to achieve the greater performance improvements. These design details enable Caffe-MPI 2.0 to realize the near-linear expansion in cluster training within the currently popular deep learning frameworks.
Caffe-MPI is the first Caffe deep learning computing framework with parallel clusters. Based on Berkeley Caffe structure and developed by Inspur, it has openly released all the codes on Github. Its advanced MPI technology optimizes Caffe with data parallelism and aims to solve the efficiency problems exist in deep learning computing model training. Recently, the version 2.0 has been updated in the community and all the users can download freely from the following link: https://github.com/Caffe-MPI/Caffe-MPI.github.io/tree/master
The 2017 AI Computing Conference (AICC) was sponsored by CAETS and undertaken by Inspur, aiming to concentrate on current needs of AI and its future development. Inspur is focusing on computing innovation to promote a sustainable development of AI industry by cooperating with companies, users, experts and developers working in AI computing and application fields. Dozens of well-known experts from home and abroad as well as thousands of professionals from various sectors participated in this seminar with a theme of "AI Computing Innovation".
About Inspur
Inspur is the leader in intelligent computing and ranked top five in worldwide server manufacturing. We provide cutting-edge hardware design and deliver extensive AI product solutions. Inspur provides customers with purpose-built servers and AI solutions that are Tier 1 in quality and energy efficiency. Inspur's products are optimized for applications and workloads built for data center environments. To learn more, http://www.inspursystems.com.
SOURCE Inspur Electronic Information Industry Co., Ltd
Share this article