Scalable Machine Learning on Heterogeneous Computing Platforms

Machine learning has been very successful with human-like performance in image classification, natural language understanding, and games like Go and Chess. This is possible with training on very large data sets over days and weeks on compute clusters with hundreds of processors and special purpose ASICs such as TPUs. Computational Hardware is the key to realizing the full potential of machine learning and eventually (perhaps) a world that is powered by artificial intelligence. The amount of computation required has increased by 300000X since 2012 (beginning of the deep learning era) and is doubling every 3.5 months!! This presents a huge opportunity for hardware companies and computer architecture researchers.

We are taking a fresh and holistic approach to memory subsystem design and optimization is required that is based on exploiting the unique benefits and trade-offs with emerging non-volatile memory and storage technologies.

This project is related to our Heterogeneous memory project.

This project is sponsored by Intel Corporation.