Chip multiprocessors (CMPs) have emerged as a primary vehicle for overcoming the limitations of uniprocessor scaling, with power constraints now representing a key factor of CMP design. Recent studies have shown that the on-chip interconnection network (NOC) can consume as much as 36% of overall chip power. To date, researchers have employed several techniques to reduce power consumption in the network, including the use of on/off links by means of power gating. However, many of these techniques target dynamic power, and those that consider static power focus exclusively on flit buffers. In this paper, we aim to reduce static power consumption through a comprehensive approach that targets buffers, switches, arbitration units, and links. We establish an optimal power-down scheme which we use as an upper bound to evaluate several static policies on synthetic traffic patterns. We also evaluate dynamic utilization-aware power-down policies using traces from the PARSEC benchmark suite. We show that both static and dynamic policies can greatly reduce static energy at low injection rates with only minimal increases in dynamic energy and latency.
展开▼
