NVIDIA is looking to help advance the field of artificial intelligence with a new Facebook partnership that they have just announced today. The partnership sees NVIDIA lending a hand in development with Facebook's Caffe2 deep learning A.I. framework which Facebook had created in part to help understand and better manage the world's growing set of information and data. Since A.I. information processing is an intensive task that requires loads of power to work efficiently NVIDIA worked with Facebook to essentially modify Caffe2 so that it would be fully compatible with NVIDIA's GPU Deep Learning platform.
To this end NVIDIA states that Caffe2 is now capable of utilizing the platform which includes the NVIDIA Deep Learning SDK libraries, which in itself includes the cuDNN, cuBLAS, and NCCL libraries. With this compatibility developers who are using Caffe2 to create large-scale distributed training scenarios and to build machine learning applications will notice a high increase in performance thanks to the support for multi-GPU acceleration. NVIDIA explains the technical details of the new acceleration benefits but what it breaks down to is that it allows developers to complete training of A.I. models much faster than they were able to before now, and as an example of how much faster NVIDIA paired their DGX-1 supercomputer with Caffe2 for the this purpose and got some interesting results.
According to NVIDIA the Caffe2 trains A.I. models seven times faster with their DGX-1 supercomputer, something that was made possible thanks to the two companies working together to bring compatibility for NVIDIA's GPU Deep Learning platform to the Caffe2 Deep Learning framework. NVIDIA displays this with the help of a chart which showcases a comparison between same training with varying amounts of GPUs on three different network types, which are the Inception v3, VGG16, and ResNet50 networks. Other terms of the partnership were not mentioned but NVIDIA does note that the partnership will see the DGX-1 supercomputer be the first system to offer Caffe2 with these optimized changes that they have made for deep learning. NVIDIA also states that they'll eventually be adding the Caffe2 training to their DLI Curriculum beginning on May 8th when they kick off their GPU Technology conference, where developers will be able to get some hands-on time with the technology.