Skip to main content

Neural Acceleration for General- Purpose Approximate Programs

Pages from p105-esmaeilzadehCACM features exciting work by Hadi Esmaeilzadeh (UW CSE Ph.D. alum now on the faculty at Georgia Tech), Adrian Sampson (UW CSE graduating Ph.D. student), Luis Ceze (UW CSE faculty), and Doug Burger (Microsoft Research and UW CSE affiliate faculty).

CMOS scaling is no longer providing gains in efficiency commensurate with increases in transistor density. Today, we can choose any two of performance, energy efficiency, and generality at the expense of the third. One approach to gaining performance and energy efficiency at the expense of generality is the use of GPGPUs and FPGAs. This paper explores another approach: approximate computation, or trading off accuracy in order to gain performance and energy efficiency. The core idea is to learn how a region of approximable code behaves and automatically replace the original code with an efficient computation of the learned model.

Read the paper here.