Recently, technology scaling has enabled the placement of an increasing number of cores, in the form of
chip-multiprocessors (CMPs) on a chip and continually shrinking transistor sizes to improve performance.
In this context, power consumption has become the main constraint in designing CMPs. As a result, uncore
components power consumption taking increasing portion from the on-chip power budget; therefore,
designing power management techniques, particularly memory and network-on-chip (NoC) systems, has
become an important issue to solve. Consequently, a considerable attention has been directed toward power
management based on CMPs components, particularly shared caches and uncore interconnected structures, to
overcome the challenges of limited chip power budget.<div>This work targets to design an energy-efficient uncore architecture by using heterogeneity in components
(cache cells) and operational parameters (Voltage/Frequency). In order to ensure the minimum impact on the
system performance, a run-time approach is investigated to assess the proposed method. An architecture is
proposed where the cache layer contains the heterogenous cache banks in all placed in one frequency voltage
domain. Average memory access time (AMAT) was selected as a network monitor to monitor the performance
on the run-time. The appropriate size and type of the last level cache (LLC) and Voltage/Frequency for the uncore
domain is adjusted according to the calculated AMAT which indicates the system demand from the uncore.<br></div><div>The proposed hybrid architecture was implemented, investigated and compared with the a baseline model
where only SRAM banks were used in the last level cache. Experimental results on the Princeton Application
Repository for Shared-Memory Computers (PARSEC) benchmark suit,show that the proposed architecture yields
up to a 40% reduction in overall chip energy-delay product with a marginal performance degradation in average
of -1.2% below the baseline one. The best energy saving was 55% and the worse degradation was only 15%.<br></div>
Energy-Efficient Run-Time Mapping and Thread Partitioning of Concurrent OpenCL Applications on CPU-GPU MPSoCs