Computing systems are used to solve computational problems through the use of appropriate algorithms, ideally those that provide desired levels of performance or energy-efficiency, under given constraints (e.g., correctness/precision or cost). These choices among algorithms for solving particular computational problems are often influenced by architecture, microarchitecture, circuit, and device properties of computing system platforms.
This talk overviews a thread of research on the interaction between computational problems, algorithms, architectures, semiconductor devices, and correctness. I will overview the design and implementation of a machine language and machine architecture for precisely capturing (and executing) possibly-error-tolerant, algorithm-independent compute problem specifications. The relation between tolerances to arithmetic deviation and preliminary theoretical bounds on encoding efficiency will be outlined, along with an analytic framework for capturing the interaction between properties of algorithms for solving given problems, and (semiconductor) device-level properties of substrates on which they eventually execute. To tie these ideas together, I will describe ongoing work on a 96-processor prototype, and will give an overview of a completed small-scale “precursor” system which integrates 16 processors with half a terabyte of flash memory while consuming under 5 Watts.
Phillip Stanley-Marbell is a research staff member with IBM Research—Zurich, Switzerland, where his research activities revolve around computing architectures for future device technologies. Prior to joining IBM in 2008, he was a post-doctoral researcher at TU Eindhoven in the Netherlands, having completed his Ph.D. in ECE at CMU in 2007. Before and during graduate school, he held intern positions at Bell-Labs (in the mid nineties) and NEC research labs. He maintains several open source hardware and
software projects, and is the author of a programming language textbook published by John Wiley & Sons in 2003.
Back to the seminar page