Polarization-coupled complete operators in sliding ferroelectric semiconductors for native in-memory attention computing
This paper presents a sliding ferroelectric semiconductor-based in-memory computing hardware that intrinsically couples linear and nonlinear operators via polarization to execute complete Transformer attention primitives, including native Softmax, with ultra-low latency and energy for real-time generative AI.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Modern artificial intelligence, particularly the systems that write stories, translate languages, and generate images, relies on a mathematical process called attention. This process allows the machine to decide which parts of a sentence or an image are most important to focus on at any given moment. To make these decisions, the computer must perform two distinct types of calculations simultaneously: simple addition and multiplication, which are straightforward, and complex, curved mathematical functions that act like a filter to normalize the results. Currently, computer chips are built to handle the simple math very well, but they struggle with the complex filtering. They must move data back and forth between memory, where information is stored, and a separate processor that handles the complex math. This constant shuffling of data slows the system down and consumes vast amounts of energy, creating a bottleneck that limits how fast and efficient these intelligent systems can become.
A team of researchers at Nanjing University and other institutions has developed a new type of computer chip that solves this problem by performing both the simple and complex calculations in the exact same place where the data is stored. They created a tiny device made from a special material called ReS2, a layered semiconductor that can be slid to change its electrical properties. Inside this single device, they integrated two different pathways for electricity to travel. One pathway moves electricity vertically through the material, while the other moves it sideways along the surface. Both pathways are controlled by the same internal switch, known as ferroelectric polarization, which acts like a permanent memory setting that can be flipped back and forth with a tiny electrical pulse.
When the researchers apply a voltage to flip this internal switch, it changes how easily electricity flows through both pathways at the same time. The vertical pathway responds in a straight, predictable line, making it perfect for the simple addition and multiplication tasks. The sideways pathway, however, reacts in a curved, exponential way, naturally performing the complex filtering math that usually requires a separate processor. Because both pathways are governed by the same internal switch, the device can execute the entire attention process in one single step without moving any data. The researchers built a small grid of these devices, arranged in a specific pattern, and demonstrated that it could successfully run the core logic of a modern language model. The system worked with incredible speed, completing operations in 600 picoseconds, which is less than a billionth of a second, and used an astonishingly small amount of energy, measuring just 320 attojoules per operation.
The team did not just stop at building the device; they proved it could learn and adapt. By adjusting the physical size of the pathways within the device, they could tune how sensitive the complex math was, allowing them to design the chip for different types of tasks. They tested the system with a closed-loop setup where the chip received text input and generated a response in real time. When prompted with the phrase "I am hungry," the hardware instantly computed the necessary attention weights and generated the continuation "need pizza now." This demonstrated that the device could handle the full cycle of understanding and generating language without external help. Furthermore, the researchers used the measured performance of their tiny device to simulate a much larger system. In these simulations, the chip successfully translated electrical signals from a mouse's brain, which represented different emotional states like fear or excitement, into written language. The simulation showed that this approach could scale up to handle complex, real-world tasks involving sensory data and language generation.
This work represents a fundamental shift in how computer hardware is designed for artificial intelligence. Instead of trying to make the existing linear math faster, the researchers embedded the missing complex math directly into the memory material itself. By doing so, they removed the need for the energy-intensive data shuffling that has plagued previous generations of chips. The device retains its settings for over a year without power and can be rewritten millions of times, proving it is stable enough for practical use. The success of this experiment suggests that the future of efficient artificial intelligence may not come from better software or faster processors, but from materials that can natively perform the complete range of mathematical operations required for thinking, all within a single, tiny unit of memory.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.