Integer-Value Encoding for Approximate On-Chip Communication

Abstract

Approximate computing has been broadly investigated in the computing domain to maintain the benefits of technology scaling. Recently, it has been extended to the communication domain. The idea is to trade transmission quality for energy or/and performance efficiency. Exploiting the integer-value representation of the transmitted signals in parallel buses, we propose two memoryless encoding approaches for approximate communications in order to reduce the integer value deviation. First, restricted to area-constrained applications, we propose a Combined Integer-Value (CIV) coding technique based on the swap and inversion of the input signals. Second, a Crosstalk-Avoidance-based Integer-Value (CAIV) coding technique for applications with a more relaxed area constraint is presented. The optimal mapping of the data words and codewords combined with the selective inversion of the input signals minimize the error magnitude in this coding technique. A comprehensive experimental result using a 65 nm commercial technology evaluates the proposed coding approaches. For example, our proposed CIV coding scheme improves the quality of the received images and the sampled radio communication signals with different modulation up to factor of 28 and 27, respectively, as it compares with conventional transmission of signals with no coding. CAIV coder can improve the data transmission accuracy by about 218 as it compares with the conventional transmission. Furthermore, to assess the applicability of the proposed encoders, we carried out two case studies employing the scale-invariant feature transform algorithm and SOBEL edge detector.

Keywords

• Approximate communication
• on-chip interconnects
• integer-value encoding
• error efficiency
• computer vision