Malleable NoC: Dark Silicon Inspired Adaptable Network-on-Chip
Haseeb Bokhari1,a, Haris Javaid2, Muhammad Shafique3,c Jörg Henkel3,d and Sri Parameswaran1,b
1School of Computer Science and Engineering, University of New South Wales, Sydney, Australia.
2Google Inc., USA.
3Chair for Embedded Systems, Karlsruhe Institute of Technology, Karlsruhe, Germany.
Network on Chip (NoC) has been envisioned as a scalable fabric for many core chips. However, NoCs can consume a considerable share of chip power. Moreover, diverse applications are executed in these multicore, where each application imposes a unique load on the NoC. To realise a NoC which is Energy and Delay efficient, we propose combining multiple VF optimized routers for each node (in traditional NoCs, we have only a single router per node) for efficient NoC for Dark Silicon chips. We present a generic NoC with routers designed for different VF levels, which are distributed across the chip. At runtime, depending on application profile, we combine these VF optimized routers to form constantly changing energy efficient NoC fabric. We call our architecture Malleable NoC. In this paper, we describe the architectural details of the proposed architecture and the runtime algorithms required to dynamically adapt the NoC resources. We show that for a variety of multi program benchmarks executing on Malleable NoC, Energy Delay product (EDP) can be reduced by up to 46% for widely differing workloads. We further show the effect on EDP savings for differing amounts of dark silicon area budget.
Full Text (PDF)