Probabilistic Value-Deviation-Bounded Integer Codes for Approximate Communication (1804.02317v1)

Abstract: When computing systems can tolerate the effects of errors or erasures in their communicated data values, they can trade this tolerance for improved resource efficiency. One method for enabling this tradeoff in the I/O subsystems of computing systems, is to use channel codes that reduce the power needed to send bits on a channel in exchange for bounded errors and erasures on numeric program values---value-deviation-bounded (VDB) codes. Unlike rate distortion codes, which guarantee a bound on the expected value of channel distortion, the probabilistic VDB codes we present guarantee any desired tail distribution on integer distances of words transmitted over a channel. We extend prior work to present tighter upper bounds on the efficiency for VDB codes. We present a new probabilistic VDB encoder that lowers power dissipation in exchange for bounded channel integer distortions. The code we present takes the peculiar approach of changing the channel bit error rate across the ordinal bit positions in a word to reduce power dissipation. We implement the code table generator in a software tool built on the dReal SMT solver and we validate the generated codes using Monte Carlo simulation. We present one realization of hardware to implement the technique, requiring 2 mm$^2$ of circuit board area and dissipating less than 0.5 $\mu$W.