| 摘要: | 自旋轉移矩磁阻式隨機存取記憶體（Spin-Transfer-Torque Magnetic Random-Access Memory, STT-MRAM）在新興記憶體中具有顯著的潛力。然而，除了會遭遇因老化引起的磁穿隧接面（Magnetic Tunnel Junction, MTJ）中時間依賴介電擊穿（Time-dependent Dielectric Breakdown, TDDB）的問題之外，在溫度變化時也會使流經MTJ的電流大小不一，導致錯誤產生與可靠度下降。這些現象在STT-MRAM記憶體內運算（In-Memory Computing, IMC）的模式中更為嚴重。為了解決這些挑戰，本文提出了一個具有熱感知能力的早期老化檢測和補償架構，在運算模式下監控記憶體運算的結果，並在檢測出有老化的可能性後予以補償。
我們的框架引入了幾個關鍵創新。首先，我們使用平行（Parallel, P）狀態的MTJ，調整感測放大器中參考產生器，以正確的運算邏輯AND與OR。其次，它具有早期老化檢測機制，同樣也是於參考產生器中使用P狀態MTJ，實時識別老化問題並記錄在查找表（Lookup Table, LUT）中以便補償。第三，我們設計了老化容忍機制，當在LUT中檢測到地址時，會透過調整運算單元進入感測放大器的電流來抵消由老化引起的電流上升。接著，我們主要都是以P狀態MTJ來做設計，包含參考產生器中的參考值以及使用P-P狀態檢測早期老化，有效地減輕了熱效應並防止相關錯誤。最後，我們的框架提供了即時且高效的解決方案，提高記憶體系統的可靠度和壽命。通過早期檢測老化和補償，解決由於老化和溫度變化導致的準確性下降，確保了穩定的STT-MRAM記憶體內運算。
;Spin-Transfer-Torque Magnetic Random-Access Memory (STT-MRAM) holds significant promise for on-chip memory, yet faces pronounced challenges arising from time-dependent dielectric breakdown (TDDB) of aging effects and thermal effects in the magnetic tunnel junction (MTJ). This phenomenon introduces concerns of endurance degradation and defect formation, particularly impactful in the context of In-Memory Computing (IMC) with STT-MRAM. This paper proposes a comprehensive early aging detection and tolerance framework with thermal awareness to address these challenges, designed to dynamically monitor memory cells during computing modes.
Our framework introduces several key innovations. First, we designed the reference generator of the sense amplifier in Parallel (P) state MTJ to compute the logic AND and OR operations correctly. Second, we proposed an early aging detection mechanism that utilizes P state MTJ in the reference generator to identify aging issues in real-time, recording in a Lookup Table (LUT) for tolerance. Third, we developed an aging tolerance mechanism that activates upon detecting addresses in the LUT, adjusting the current of the computing cells to counteract the current enhancement caused by aging, instead of the adjustment of the reference generator. Next, our design mostly used the P-state MTJ, including the reference in the reference generator and the P-P state early aging detection, effectively mitigating thermal effects and preventing related errors. Finally, our framework provides a real-time and efficient solution that enhances the reliability and lifetime of the memory system. By enabling early detection and compensation for aging, we address the accuracy degradation caused by aging and temperature variations, ensuring stable in-memory computing with STT-MRAM. |
