The world of artificial intelligence (AI) is rapidly evolving, and with it, the need for more efficient and powerful computing hardware. A recent breakthrough in chip technology from researchers at The University of Hong Kong (HKU) has the potential to revolutionize AI computing, addressing a critical challenge in the field: the von Neumann bottleneck. This bottleneck, a result of slow and energy-intensive data transfer between memory and processors, has been a major hurdle in the development of faster and more energy-efficient AI systems.
The HKU team, led by Professor Can Li, has developed a novel analogue content-addressable memory (CAM) using two-dimensional (2D) MoS2 flash memories. This innovation is a significant leap forward in chip architecture, offering a minimalist design that performs complex searches directly where the data is stored. By utilizing MoS2, a 2D material that is tens of thousands of times thinner than a human hair, the team has achieved a remarkable reduction in chip footprint and power consumption.
One of the key challenges in this development was overcoming the 'Schottky barrier,' a wall that restricts electron flow when connecting 2D materials to circuits. The HKU team addressed this by using antimony (Sb), a semimetal, as the contact electrode, effectively dismantling the barrier and creating a smooth, low-resistance pathway. The result is a device that is incredibly fast and efficient, with a high read-out current and an ON/OFF ratio exceeding 10^9.
The fabricated analogue CAM array demonstrated a record-breaking energy consumption of under 0.1 femtojoules (fJ) per search per cell, and a latency of just 36 picoseconds. This is an astonishingly short time, considering that light travels only 0.3 millimeters in one picosecond. The device's performance is a testament to the team's innovative approach and their ability to overcome technical challenges.
The team's work also showcases the potential for 3D heterogeneous integration, where N-type MoS2 and P-type WSe2 are vertically stacked within a single CAM cell. This architecture not only halves the area required but also eliminates the need for peripheral inverter circuits, further reducing power consumption. When tasked with calculating the analogue Hamming distance for machine learning classification, the new analogue CAM performed the task approximately 10^8 (100 million) times faster than a standard CPU, scoring exceptionally high accuracy across multiple datasets.
From an application perspective, search operations are incredibly valuable, according to Professor Li. The 'attention mechanism' that serves as the foundation for Large Language Models (LLMs) is essentially a search process. The team is exploring how to use this high-performance hardware to implement search mechanisms for large AI models, potentially revolutionizing the way AI systems operate.
This breakthrough also paves the way for ultra-compact AI chips that operate locally on edge devices. Tasks like smartphone facial recognition could be completed instantly on the device without uploading private data to the cloud. While large-scale commercialization still requires overcoming engineering challenges in packaging, this research has successfully proven the principle, providing a clear blueprint for next-generation, high-performance AI hardware.
The research article, titled 'Sb-contacted MoS2 flash memory for analogue in-memory searches,' was published in Nature Nanotechnology. The paper outlines the team's innovative approach and the potential impact of their work on the future of AI computing. As the field of AI continues to evolve, breakthroughs like this one from HKU are crucial in pushing the boundaries of what is possible, offering a glimpse into a future where AI systems are faster, more efficient, and more accessible.
