Artificial intelligence (AI) and deep learning (DL) are the engines that are rapidly powering innovation across industries from healthcare to autonomous vehicles and agriculture. By 2020, IBM projects that the world’s volume of digital data will exceed 44 zettabytes [1]. Organizations that recognize the value of their data for decisions and actions are turning to DL systems that can rapidly ingest, accurately interpret, and quickly provide key data insights from the volumes of new data generated now and in the future.

Enterprises are increasing investment in AI research and innovation, with related patents growing more than 30% and academic papers by 13% during the last decade [2]. Arguably, only one kind of enterprise will survive and thrive in the future— the data-driven enterprise. Highly performant and scalable DL systems must excel at data ingestion, data training, and verification, and deliver inferences and classifications while handling the growing demands of DL in the organization. The DL algorithm accuracy may improve by increasing the neural network size, and data quantity and quality used for the model training step. However, this accuracy improvement comes with a significant increase in computational complexity and increased demand on DL compute, storage, and network resources.