| Geometrical scaling of critical semiconductor dimensions while keeping electric field constant improves logic delay and power consumption, but degrades global and semi-global interconnect delay. Repeater insertion and higher aspect-ratio wires partially alleviate interconnect delay, but adversely impact power consumption, temperature, and crosstalk noise. Although a number of interconnect design techniques have been proposed, they primarily target: (1) one or the other design metric, not all relevant metrics simultaneously and/or (2) random or worst-case traffic conditions instead of real-workload traffic, which is important because interconnect behavior (energy, delay, temperature, etc.) depends upon traffic value characteristics.;To address the shortcomings of previous work, we present a value-aware interconnect design methodology that permits simultaneous optimization of multiple metrics in a prioritized manner tailored to target application requirements. Towards this end, we first survey existing commonly-used interconnect design techniques and analyze their influence on different interconnect circuit parameters and design metrics. From among these, we focus on wire spacing and data encoding as examples of two techniques to which we apply our value-aware methodology for optimization of interconnect energy, average delay, and peak wire temperature in the context of the SPEC CPU2k benchmarks. Next, we describe and compare two value-aware techniques that optimize energy savings within area constraints: partitioned hybrid encoding and value-aware wire spacing. Finally, we present three encoding techniques with increasing flexibility to adapt to traffic value characteristics: (1) static, which employs a single encoding mode, fixed at design time based on traffic value characteristics, in all cycles; (2) dynamic, which chooses the most suitable encoding mode in any given cycle from among two or more predetermined modes; and (3) adaptive, which selects the most appropriate encoding mode in any given cycle from among a set of modes that changes with traffic conditions at run time. These techniques provide increasingly greater benefits, with our static technique easily outperforming previous, more complex, but value-oblivious, dynamic encoding techniques. |
