Data-dependent approaches for reducing power consumption

Significant power reduction can be achieved by exploiting variations in the data operated on in a microelectronic implementation. This dissertation demonstrates the magnitude of this opportunity by developing novel, low-power design approaches that exploit data variation in a speech codec, a video codec, and a channel decoder—three of the most important functions in digital communications today. Other applications with similar data exploitation opportunities can also benefit from these approaches.; Much of the fixed-point, two's complement data that is processed by the IS-54 Vector Sum Excited Linear Prediction (VSELP) speech codec, exhibits characteristics of significant sign extension, small magnitude variation, and significant truncation. An intuitive model that relates these data characteristics to switching activity of datapath interconnect is found. A heuristic for reducing energy consumption based on this model is presented and applied in detailed examples herein. Potential energy reduction of about 18% during the two codebook searches, lag search, and synthesis filter computation, the four most computationally intensive modules of this speech codec, is demonstrated with minimal impact on speech quality.; Frames of video data are classified as I-type, P-type, and B-type when processed by an MPEG-2 video codec. Differences in the resilience of each of these frames to Discrete Cosine Transform (DCT) and Inverse DCT (IDCT) approximation during encoding and decoding are studied. Low-power methods for selectively applying approximations to frames in six practical video communication configurations are found for a range of quality constraints. Potential energy reduction of 8 to 25% during DCT and IDCT computation is demonstrated.; Convolutional coded data varies with changes in channel E_b/N ₀ and code rate. Approximate Viterbi decoding achieved by varying truncation length and the pruning threshold of the T-algorithm while employing trace-back memory management is studied for these variations and as maximum acceptable bit-error rate (BER) varies. Potential energy reduction of 70 to 97.5% compared to Viterbi decoding is demonstrated by optimally adapting truncation length and pruning threshold.; Based on the three cases studied in this dissertation, resulting power reduction can indeed be worth the added effort of developing and applying approaches for exploiting data variation.