Analog Compute-in-Memory at Mythic

Abstract

AI and many other applications have opportunities to build systems that merge memory and computing into a unified structure in ways which yields significant improvements in energy efficiency, performance, and cost. In these scenarios, “moving the compute to the memory” makes sense because the applications have large amounts of data to process and relatively simple operations to perform, which makes it not too difficult to create specialized processing near the memory when traditionally moving the data to the main system processor would be slow and inefficient. These approaches have a  have a wide variety of applications as well as approaches. On one end of the spectrum, systems which have processing inside of an SSD to perform searches inside the drive itself, and on the other end of the spectrum, systems that have analog compute performing mathematics directly on the bitlines of the memory arrays. In this talk we will provide an overview of many of these approaches as well as the methods to their madness.

Bio

Dave Fick co-founded Mythic with Mike Henry in 2012, a company building next-generation intelligence processing units (IPUs) for edge and cloud inference. Mythic has raised $86M of venture capital to combine analog computing, dataflow architecture, and compute-in-memory to provide the best combination of performance and energy efficiency available. Dave received his PhD from the University of Michigan under David Blaauw and Dennis Sylvester at the Michigan Integrated Circuits Lab.