BEIJING, Feb. 10 (Xinhua) -- Chinese scientists have developed a high-performance computing-in-memory device using a novel engineered material. This breakthrough provides a crucial foundation for advancing next-generation chips that aim to achieve both low power consumption and integrated memory-processing capabilities.
A team from the College of Chemistry and Molecular Engineering at Peking University has successfully tackled longstanding challenges associated with applying ferroelectric materials to advanced chips. Their breakthrough, recently published in the journal Science, involves the creation of wafer-scale, ultrathin and uniform ferroelectric thin films, which were used to produce high-speed ferroelectric transistors.
In conventional chips, the physical separation between memory and computing units forces data to travel continuously, a process that creates a major bottleneck for computational performance.
Ferroelectric materials offer a promising solution due to their inherent ability to retain a "memory" state through rapid polarization switching, potentially allowing memory to be built directly into the computing unit itself, thus enabling true computing-in-memory.
However, integrating these materials into modern chips has been hindered by three significant obstacles: the difficulty in producing large-area films that are both extremely thin and uniform; the severe degradation of ferroelectric properties at such reduced thicknesses; and the challenge of achieving seamless integration with semiconductor layers.
The Peking University team has managed to overcome all these hurdles. "We fabricated an ultrathin, uniform ferroelectric oxide on a wafer, similar to how single-crystal silicon wafers are made," said Professor Peng Hailin, dean of the college.
"Remarkably, even at a thickness of just 1 nanometer, it retains excellent ferroelectric properties. The memory and computation layers are perfectly integrated, and our process is compatible with existing semiconductor production lines," Peng added.
The resulting ferroelectric transistors exhibit exceptional performance. Operating at a low voltage of 0.8 volts, they can write data in just 20 nanoseconds and endure over 1.5 trillion rewrite cycles.
This reliability surpasses the stringent demands of cloud AI computing, and their overall performance exceeds that of current industry-standard hafnium-based ferroelectric systems.
Peng noted that this achievement provides a novel material solution to the intertwined problems of chip power consumption and computing limits.
The scalable production and application of this technology could lead to future electronic devices that are more durable, faster and smarter, Peng explained.
Computing-in-memory chips could dramatically cut energy wasted on data transfer. Furthermore, they would enable complex AI algorithms to run efficiently directly on smartphones, cars and home appliances, reducing dependence on cloud computing, according to Peng. ■
