Stefano Ambrogio, Pritish Narayanan, Hsinyu Tsai, Charles Mackin, An Chen, Robert M. Shelby, Geoffrey W. Burr, IBM Research-Almaden
We provide a summary of recent progress in hardware acceleration of AI algorithms, such as training Fully-Connected Networks based on large Phase-Change-Memory arrays. Crossbar arrays of weights encoded as conductances can provide orders of magnitude increases in speed and energy efficiency with respect to state of art CPUs and GPUs.