應用於近記憶體資料處理系統之高頻寬三維堆疊動態記憶體控制單元

Abstract

大數據是現今最熱門的話題及研究趨勢，而分析這些數據就需要使用非常大量的記憶體存取去讀取及運算這些資料，為了應付多位使用者的存取，記憶體需要高頻寬以及高密度。而在大數據的時代，移動資料所消耗的能量也隨著資料量成長到無法忽視的階段，面對大量資料的儲存及讀取，高頻寬三維堆疊動態記憶體是個十分具有潛力的解決方法，利用矽導孔(TSVs)技術的三維堆疊動態記憶體比起傳統的隨機存取記憶體多出至少32倍的輸出/輸入腳位，為了利用三維堆疊動態記憶體所帶來的高容量及高頻寬，我們需要一個設計一個新的記憶體控制單元。在這篇論文中，我們提出了一個可以充分利用三維堆疊動態記憶體好處且有效率使用能源的記憶體控制單元。利用指令的重新分配和記憶體的交錯排列來平行化多位使用者所發出的指令，而自我刷新(self-refreshing)則可以節省記憶體刷新所耗費的能量，運用近記憶體資料處理來縮短資料搬移的路程可以有效減少資料傳輸的能量消耗。使用這個記憶體控制單元使整體的頻寬有66.8%的提升，而執行時間也減少了40.08%，在能量消耗方面也有27.18%的減少。

Nowadays, big data becomes one of most popular topics in the world. Analyzing these data needs large amount of memory accessing. For the requests of multi users, the memory need high bandwidth and high density. The power of moving data also needs to be considered in the big data generation. High density 3D-stacked DRAM is the potential solution for the big data storage. By applying the through-silicon-vias (TSVs) technology in 3D-stacked DRAM, the I/O pins between logic tiers and DRAM are at least 32x larger comparing the conventional DRAMs. In order to utilize the 3D- DRAM lager capacity and memory bandwidth, a new advanced 3-D memory controller is needed. In this paper, we present an energy-efficient DRAM controller to fully utilize the benefit offer for 3D-Stacked DRAM. The controller uses command rescheduling and rank interleaving to parallel the commands from multi users, and the self-refresh can save the power for 3D-stacked DRAM refresh. Using near-data processing short the data path of transfer and makes the energy consumption decreased. The controller can realize the bandwidth improvement by 66.8%, and the execution time improvement by 40.08%. The energy consumption also decrease 27.18%.