2023-07-012024-05-16https://scholars.lib.ntu.edu.tw/handle/123456789/670047在最先進的深度神經網路人工智慧中需要大量的乘積累加運算 (multiply and accumulate operations)以及數據移動。如果以CPU/GPU為基礎的傳統運算架構中運行人工智慧，大量的數據移動所消耗的能量可高達處理數據所需能量的100倍，這個問題也被稱為memory-wall。因此，記憶體內運算 (In-memory computing, IMC)在人工智慧應用領域吸引了大量的注意，並被認為很有前景可以解決memory-wall的問題，記憶體內運算可將人工智慧需要的乘積累加運算直接在記憶體矩陣中完成，並達到更高的能源運用效率。過去的文獻已經展示了利用CMOS或是非揮發性記憶體（如SRAM, DRAM, ReRAM, MRAM, PCM等等）的記憶體內運算架構。 由於磁阻式隨機存取記憶體（MRAM）具有零待機漏電流、較高的寫入/讀取速度、較高的能源效率、與CMOS製程的相容性、高耐用性(endurance)、高資料保存時間(data retention)，以及較高的整合密度(integration denstity)，相較於其他非揮發性記憶體，磁阻式隨機存取記憶體有很大的機會應用在記憶體內運算中。然而，大多建構於非揮發性記憶體只能應用在具有固定權重參數（static weight parameters）的神經網路中，因為較高的記憶體寫入能耗以及有限的耐用性。因此展示出具有超低功耗、在裝置內學習（on-device learning）並且可頻繁變動權重參數的記憶體內運算晶片是非常有挑戰性的。 State-of-the-art deep neural network (DNN) based AI computing systems demand numerous multiply and accumulate (MAC) operations and data movement. In conventional computing architectures used in GPU/CPU, the massive data communication energy could be almost two orders higher than data processing itself, known as “memory wall”. In-Memory Computing (IMC) for AI system has attracted tremendous attention as a promising solution due to its capability to perform MAC computation directly within the memory array to achieve orders of magnitude improvement in overall system energy efficiency. Different types of IMC designs based on either CMOS or post-CMOS non-volatile memory (NVM) technologies have been demonstrated, such as SRAM, DRAM, ReRAM, MRAM, PCM, etc. Among the NVM based IMC designs, the approach based on magnetoresistive random access memory (MRAM) is very promising due to its zero standby leakage, high write/read speed & efficiency, compatibility with CMOS fabrication process, scalability, superior endurance, excellent retention time, and high integration density, etc. However, most NVM based IMC, including MRAM, only targets for neural network inference with static weight parameters due to high memory writing energy or limited endurance. It is exceedingly challenging to demonstrate in real IMC chip with ultralow power, on-device learning capability with frequent weight updates.深度神經網路;人工智慧;記憶體內運算;非揮發性記憶體;磁阻式隨機存取記憶 體;自旋轉移力矩;自旋軌道力矩;Deep neural network (DNN); Artificial intelligence (AI); In-memory computing (IMC); Non-volatile mem