China-led team reports 10 billion write cycles in ferroelectric memory

Creating faster memory and storage with greater capacity while keeping energy requirements low is a long-standing goal in computing. In recent years, research has increasingly focused on wurtzite ferroelectrics, polar crystal materials such as aluminum scandium nitride, or AlScN. The material is compatible with semiconductor manufacturing processes already in use and offers low energy use and fast switching speed, but until now reliability has been the main challenge. Tests of previously developed AlScN chips failed after around 100 million write cycles, according to TechRadar.

The researchers investigated why the AlScN chips topped out at 100 million cycles. They found missing nitrogen atoms, or nitrogen vacancies. Wang Ruiqing, a doctoral researcher at Xidian University and one of the authors of the research paper, compared the ferroelectric material to a neatly planted cornfield. “Think of ferroelectric material as a neatly planted cornfield,” Wang said, “with nitrogen vacancies representing spots where seedlings are missing.” The defects were not the problem by themselves; their behavior during switching was. Rather than remaining stationary, the nitrogen vacancies moved and even grouped together. That created “pathways” that allowed electricity to leak.

Upon discovering this, the team designed a structure to minimize the nitrogen vacancies and slow the deterioration of the material. Wang said previous researchers knew that devices failed and could observe some symptoms, but no one had been able to explain at the atomic scale what was moving, how it was moving, and how that movement ultimately caused failure.

With the new structure, the team demonstrated about 10 billion writing cycles, around 100 times the previous limit. TechRadar described the development as one of the most significant breakthroughs in memory development, and said it could make dense, low-power memory chips a realistic possibility for future demands of ever-scaling AI processing. If successful, AlScN-based RAM and storage could form the backbone of future server farms and provide the framework for greater large language model and artificial intelligence adoption, according to the report.