Date: May 21, 2018 Source: EE Times (
click here to go to the source)
AUSTIN, Texas -- Startup Revolution Computing here has novel ideas both for a machine-learning accelerator and for holding down the costs of designing it. The company aims to produce for as little as $12 million its first chip based on a new architecture tuned for a broad range of data analytics algorithms.
The Revolver chip will speed up a basket of algorithms that analyze the structured databases that most businesses maintain. For example, it believes that it will boost x86 server performance more than tenfold on decision trees, random forests, ensemble methods, support vectors, and gradient boosting.
The algorithms serve a wide variety of use cases including fraud detection for banks, ad matching and recommendation engines for retailers, and predictive maintenance for industrial users.
The chip also will run deep-learning jobs. However, Revolution co-founder Rodney Hooker notes that many businesses don't maintain the large unstructured image and video datasets suited for deep learning.
The startup is forging partnerships with vendors of data analytics software such as Fair Isaac Corp. Hooker believes that another new startup, SambaNova, also will try to deliver both hardware and software to accelerate for businesses a similarly broad set of machine-learning algorithms.
Revolution will implement a version of the Predictive High-Performance Architecture Research Mavens (PHARM) technology developed over the last decade at the University of Wisconsin under Mikko Lipasti, also a co-founder of Revolution along with one of his PhD students, Mitchell Hayenga, who serves as the startup's chief executive.
PHARM defines a way to build out-of-order processors without front-end fetch-and-decode blocks or large register sets. Instead, it defines the concept of a crib in which instructions are held until they are ready for processing.
