High-Performance Approximate On-Chip Communication

Abstract

Interconnects do not gain from technology scaling as the computation units do. There is still no substitute for copper interconnects to meet the conductivity requirements and reduce dielectric permittivity. It is becoming more expensive to guarantee the interconnect's reliable error-free operation as the gains of technology scaling are diminishing. The on-chip communication is currently a bottleneck of the high-performance integrated circuits. Fundamental improvement of on-chip communication is essential. Approximate computing has been thoroughly investigated for the computation sub-systems. Recently, approximate computing has been extended to the communication domain. The idea is to trade transmission quality for energy or/and performance efficiency. Approximate communication, on the one hand, provides a practical solution to maintain the benefits of the technology scaling; and, on the other hand, exploits the gap between the accuracy requirements of the applications and that provided by communication sub-systems. Besides, the unfulfilled potential of full-system approximation can only be exploited by approximating all sub-systems, including the communication. However, prerequisites for the development and efficiency characterization of effective on-chip codings and approximate techniques are precise energy and delay models. Traditionally, delay and energy models do not consider variations, resulting in inaccurate and only partially useful models. The misalignment effect, that is, the variation in the delay of the transmitted signals has been conventionally overlooked. In this dissertation, the first high-level misalignment-aware energy and delay models for narrow multi-segment global interconnects are proposed to address the problem of conventional models. The proposed models have been evaluated using simulations. Results show a significant improvement in the accuracy of the energy and delay estimations. Three different techniques are proposed to exploit approximate computing's considerable potential in the communication domain. First, two memoryless encoding approaches for approximate communications to reduce the integer-value deviation are presented. Restricted to area-constrained applications, we propose a Combined Integer-Value (CIV) coding technique based on the input signals' swap and inversion. Moreover, a Crosstalk-Avoidance-based Integer-Value (CA-IV) 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 input signals minimizes the error magnitude in this coding technique. Second, taking into account the challenges of classical wave-pipelining in error-free operation, the concept of Stochastic Wave-Pipelining (SWP) communication is proposed. It relaxes the limiting constraints of the classical wave-pipelining and allows its practical implementation. Besides, extra performance improvement by accepting a tolerable error is achievable. Unlike most existing approximate techniques, SWP has the quality to be combined with other approximate techniques. Third, an Alternating Bit-Truncation (ABT) technique is introduced. ABT sets a certain number of LSBs to zero. The inactive lines are used as virtual shielding to shield the data lines. The inactive lines do not contribute to dynamic energy consumption, resulting in a very power-efficient technique. Besides, depending on the number of truncated lines, ABT reduces the crosstalk noise that leads to energy saving and performance improvement. This technique can be implemented with almost no overhead. A comprehensive experimental result validated the proposed approximate communication approaches. The proposed approximate communication methods' applicability is also studied in Network-on-Chip (NoC), where they are compared with a state-of-the-art compression technique. Results show the superiority of proposed techniques for different design criteria.