Even though memory has no moving parts, it is prone to wear and eventually erosion. The electrical current flowing through the chips eventually causes failure and brings the entire system down. In the case of flash memory, which is also built on DRAM memory technology, that means data loss.
But now Chinese researchers have developed a technique that could greatly improve the durability of next-generation semiconductor memory technology, potentially removing one of the biggest obstacles to its use in high-performance computing and future AI systems.
A research team led by scientists at Xidian University in Xi’an, working with researchers from City University of Hong Kong and Fudan University, has demonstrated more than 10 billion write cycles in a class of materials known as wurtzite ferroelectrics. The research was published in the journal Science.
The result represents roughly a 100-fold improvement in endurance compared with previous generations of memory using the same type of material. Earlier devices based on the technology generally started to deteriorate after about 100 million write cycles, a level considered insufficient for many commercial memory applications.
Wurtzite ferroelectrics have attracted increasing attention as potential materials for next-generation memory because they can switch between two stable electrical states to represent stored data.
One material that has drawn interest is aluminium scandium nitride (AlScN) for a number of reasons. It offers fast switching between states and potentially low power consumption. It can also be integrated with semiconductor manufacturing processes already used in the industry, potentially reducing the need for entirely new production infrastructure.
However, the technology has a major reliability problem. Repeated electrical switching can cause the material’s performance to degrade, limiting the number of times a memory cell can reliably be rewritten.
The Chinese research team appears to have addressed that problem by controlling the movement of nitrogen vacancies within the material’s crystal structure. These atomic-scale defects can migrate during repeated electrical operation and contribute to degradation.
Restricting that movement allows the ferroelectric material to maintain its switching characteristics for substantially longer, according to reports describing the research.
The improvement if it pans out would be significant because future computing systems are expected to require memory capable of supporting extremely large numbers of read and write operations, particularly as it relates to AI computing.
A memory technology capable of surviving billions or tens of billions of write cycles would be much more attractive for high memory reads and write scenarios for its durability and longevity.
Needless to say, this technology is still years away. Researchers did not even give a timeline for when it might appear in the market.
