[11a-PB5-1~21] FS.1 Focused Session "AI Electronics"

[11a-PB5-1] Evaluation of MTJ Arrays for Binary-Weight Neural Networks

〇(M1C)Yuya Kato¹, Tatsuya Yamamoto², Tomohiro Ichinose², Takayuki Nozaki², Shinji Yuasa², Sumito Tsunegi^1,3 (1.Kyutech, 2.AIST, 3.Neumorph center)

Keywords:

spintoronics,Non-volatile memory,Binary neural network

Binary-weight neural networks (BwNNs), which constrain weights to -1 and +1, are promising for edge AI hardware because they reduce memory requirements and computational costs. Magnetic tunnel junctions (MTJs), capable of non-volatile storage of two resistance states corresponding to parallel and antiparallel magnetization configurations, are attractive candidates for physically implementing binary weights. In this study, a 64 × 64 1T-1MTJ memory array fabricated by integrating MTJs onto a CMOS array produced using a TSMC 0.35 µm process was evaluated. Cells exhibiting a magnetoresistance (MR) ratio greater than 50% were defined as valid. The measured valid-cell yield reached 92.4%. The coefficients of variation were 6.5% for the resistance in the parallel state and 7.4% for the MR ratio, indicating relatively uniform characteristics across the array. Even under practical measurement conditions including CMOS circuit resistance, the average MR ratio remained 79.4%, and the resistance distributions of the two magnetic states were clearly separated. These results demonstrate the potential of the proposed MTJ array as a binary-weight storage device for BwNN applications. Further results on its implementation in BwNNs will also be presented.