Open Source Summit North America 2017

With the emergence of new powerful hardware accelerators, such as Graphical Processing Units (GPUs) and Programmable Field Gate Arrays (FPGAs), more and more artificial intelligence (AI) has appeared in our daily lives, from Siri, Alexa, language translation, image recognition, to self-driving cars. The cognitive era has truly begun. In this presentation, we will look at the key characteristics of cognitive workloads, including the deep-learning workloads that employ neural networks to solve complex AI problems. Popular software stacks and frameworks supporting these workloads, such as TensorFlow and Caffe, will be discussed. We will then present some performance data to show how well typical cognitive workloads, such as neural network model training, run on GPUs and FPGAs that are available on public clouds, such as IBM SoftLayer, Amazon Web Services (AWS), Microsoft Azure, and others. Finally, we provide some recommendations on what public clouds should provide, in terms of hardware and software images, to optimize for cognitive workloads.