- 期刊
鐵電記憶體於類神經網路計算之應用
Ferroelectric Memory Transistors for Neuromorphic Computing
摘要
隨著近年來AI產業的興盛，面對未來處理大量數據的需求，使用記憶體內運算來突破von Neumann瓶頸是目前重要的研究方向之一。為增加AI系統運算效率，須減少記憶單元的面積來增加晶片的記憶體容量，以避免須不斷自記憶體晶片搬移資料之情況我們在記憶體內運算電路中運用鐵電記憶體取代傳統隨機存取記憶體（SRAM），除了元件面積小之外，還擁有多位階（Multi-level）儲存能力，更適合應用於AI類神經網路。本實驗室開發了模擬平台CIMulator，針對記憶體內運算之AI系統進行訓練與預測，也驗證了鐵電記憶體在AI應用的可行性：此元件在可操作溫度範圍内表現相當優異。
並列摘要
With the rapid advancement of the Al industry in recent years, and the need to process a huge amount of data in the near future, compute-in-memory (CIM) breaks through the von Neumann bottleneck and became one of the promising research areas nowadays. To increase the overall computing speed in AI systems, it is necessary to reduce the memory cell size to fit as much data on the computing chip as possible, so as to avoid unnecessary off-chip data transfer. We explored the use of ferroelectric memory to replace traditional SRAM in CIM circuit macros. In addition to the small device area, it supports multi-level operation, which makes it an excellent candidate for use in neural networks. Our research team has developed a simulation platform that performs training and inference simulation for AI systems based on CIM. Using this platform, we have also verified the feasibility of ferroelectric memory for Al-the new device exhibits excellent performance in its operational temperature range.