A Graph-Based Dataflow Architecture for Executing Neural Networks

Abstract

Neural networks are graph-based applications with opportunities to execute many graph nodes concurrently. Recent architectures have responded with massively parallel systems, but scheduling them has proved challenging, often relying on an oracle compiler. Instead, Mythic created an architecture that works on graphs directly: producer/consumer relationships are hardware concepts and parallel execution happens automagically when dependencies are met. This presentation gives a high-level overview of Mythic’s architecture to quickly and efficiently achieve parallelism on a wide variety of neural networks.

Biography

Dave Fick is a full-stack computer engineer, with significant experience in software engineering, computer architecture, digital integration, and full-custom design for analog, digital, and flash circuits. He works towards first-pass success through process-driven methodology which unites all layers of the system. He received his PhD in Computer Science and Engineering from the University of Michigan.
He is the CTO and co-founder of Mythic, a US-based startup that is creating the next generation of AI inference microchips. Mythic builds unique technologies to approach neural network inference.