抄録
Three-dimensional (3D) memory stacking technology is one of the most promising technologies providing massive memory bandwidth, and thereby provides an opportunity to overcome performance bottlenecks in memory systems. In this paper, we analyze the performance of well-known efficient memory components such as stream buffer, L2 cache, and victim cache in order to exploit the massive memory bandwidth of the 3D stacked memory system. This paper also presents a method to determine the optimal memory capacity with the best performance efficiency (a.k.a. performance improvement efficiency (PIE)) and proposes a cooperative cache memory (CCM), which prefetches adaptively according to the memory access pattern in order to exploit the massive memory bandwidth of a 3D stacked memory system. The proposed CCM with a 512 KB L2 cache and 256 KB stream buffer with the proposed prefetching mechanism can deliver superior performance by about 13% over a conventional 8 MB L2 cache.
