The future of artificial intelligence could depend as much on memory as processing power, and today’s QAMSS lecture explored how a new generation of materials may transform both. Dr. Markus Hellenbrand of the University of Cambridge outlined how advances in resistive switching technologies could overcome some of the most pressing limitations facing modern AI hardware.
The talk focused on the growing strain AI places on conventional computing systems, where existing complementary metal-oxide-semiconductor (CMOS) memory technologies face difficult trade-offs between speed, endurance, energy consumption and volatility. As AI models become larger and more complex, these bottlenecks are increasingly limiting performance and efficiency.
Dr. Hellenbrand presented research into novel memory materials and devices designed specifically for next-generation AI applications. Central to the work is resistive switching, a technology that allows memory devices to change resistance states in ways that can store and process information more efficiently than traditional architectures.
By engineering materials from the atomistic level upwards, the research team has developed quasi-analog memory cells capable of high endurance, improved uniformity and tuneable volatility. These properties could support emerging forms of in-memory computing, where data is processed directly within memory hardware rather than transferred repeatedly between processor and storage.
The lecture highlighted the potential of these devices for neuromorphic AI systems, which aim to mimic the efficiency and adaptability of the human brain. The work represents a significant step toward faster, lower-power AI hardware capable of meeting rapidly growing computational demands.
"We are in a very exciting time for this type of research. The semiconductor industry has keenly taken note, and initial products exist, but at the same time, many challenges remain for us to solve," Hellenbrand commented.