Content-addressable memory searches by content rather than by address, which makes it valuable for pattern matching, but conventional implementations built from static memory cells are expensive in area and power. A team from the Department of Electrical and Computer Engineering and the Centre for Advanced Semiconductors and Integrated Circuits in the HKU Faculty of Engineering has built an analogue version from two-dimensional materials using just two transistors per cell, where digital designs need up to sixteen.
The device is a flash memory made from molybdenum disulfide, with antimony used as a semimetal contact to overcome the Schottky barrier that normally limits current injection into such materials. Measured performance includes read-out current above 60 microamps per micrometre, an on-off ratio above a billion, energy use below 0.1 femtojoules per search per cell and 36 picosecond latency. Vertically stacking n-type molybdenum disulfide with p-type tungsten diselenide halves cell area and removes the need for peripheral inverters.
On analogue Hamming-distance classification, the task the architecture is built for, the chip ran around a hundred million times faster than a standard processor. The work was published in Nature Nanotechnology. The research group notes that the attention mechanism underlying large language models is essentially a search operation, which positions in-memory search as directly relevant to transformer workloads rather than to niche pattern matching alone.
The target here is the von Neumann bottleneck, the cost of shuttling data between memory and processor that increasingly dominates machine learning energy budgets. Performing the comparison where the data already sits avoids that movement entirely. Large-scale commercialisation still faces packaging engineering challenges, which is the familiar obstacle for two-dimensional material devices moving from laboratory demonstration to manufacturable product.