2024 International Conference on Solid State Devices and Materials
Sep 1 - Sep 4, 2024Arcrea HIMEJI
2024 International Conference on Solid State Devices and Materials
Sep 1 - Sep 4, 2024Arcrea HIMEJI
[B-5-01 (Invited)]Online Training with Analog RRAM Chips
Qi Qin1, Yue Xi1, 〇Bin Gao1(1. Tsinghua Univ. (China))
Energy efficient, accuracy and learning capability are the key benchmarks for edge devices. Through co-optimize of device and training algorithm, analog RRAM based computation-in-memory (CIM) chip can implement software-comparable accuracy and improve learning with online training. This paper first describes the technologies for optimizing analog RRAM performance. The analog switching characteristics, state number and relaxation are discussed. Then, different integrated scale of analog RRAM CIM chips will be presented. Finally, the paper summaries the online training algorithms for CIM chips, which can improve accuracy of various AI tasks and adapt to new environment.