Shared Virtual Output Queuing: Maximizing NoC Switch Utilization
TimeTuesday, July 12th2:15pm - 2:30pm PDT
Location2012, Level 2
DescriptionModern mobile SoC integrates throughput-oriented processors such as GPU/NPU/ISP for interactive gaming and AI camera experiences, which requires increasingly higher memory and NoC bandwidth for high-throughput computation. The switch contention by Head-of-Line (HoL) blocking is the major bottleneck limiting NoC bandwidth in memory channel-interleaved systems. In order to resolve the HoL blocking and maximize switch utilization, we evaluate the previous proposals and their limitations, and propose Shared VOQ enabling to 1) share a single buffer with packets heading to different destinations and 2) consider all matching availability between sources and destinations to maximize switch utilization. Our evaluation shows that Shared VOQ achieves up to 10.3% throughput improvement comparing to the conventional queue with up to 14.6% area overhead and negligible timing impact. It is also showed that achievable throughput by shared VOQ can be nearly reached to theoretical maximum.