How Multithreading Addresses the Memory Wall

The memory wall is the predicted situation where improvements to processor speed will be masked by the much slower improvement in dynamic random access (DRAM) memory speed. Since the prediction was made in 1995, considerable progress has been made in addressing the memory wall. There have been advances in DRAM organization, improved approaches to memory hierarchy have been proposed, integrating DRAM onto the processor chip has been investigated and alternative approaches to organizing the instruction stream have been researched. All of these approaches contribute to reducing the predicted memory wall effect; some can potentially be combined. This paper reviews several approaches with a view to assessing the most promising option. Given the growing CPU-DRAM speed gap, any strategy which finds alternative work while waiting for DRAM is likely to be a win