Analogue Phase Change Computational Memory with High Precision Reads and Energy Efficient Writes

Computing has long relied on a separation between where data is stored and where it is processed. In a standard computer, the brain of the machine must constantly shuttle information back and forth to the memory, a journey that consumes vast amounts of energy and slows down performance. To overcome this bottleneck, scientists have been exploring a different approach called in-memory computing, where the memory itself performs the calculations. This method is particularly promising for artificial intelligence, which requires massive amounts of mathematical operations to recognize patterns and make decisions. The ideal memory for this task would be able to hold a continuous range of values, much like a dimmer switch rather than a simple on-off light, allowing it to mimic the complex connections found in the human brain. However, building a device that is small enough to pack millions of these units together, while remaining precise enough to calculate accurately and efficient enough to run without overheating, has proven to be a formidable challenge.

A team of researchers at IBM Research and the Italian National Research Council has now introduced a new design for a memory device that addresses these conflicting demands. They focused on a type of memory known as phase-change memory, which stores information by switching a special material between two physical states: a disordered, glass-like state and an ordered, crystal-like state. By carefully controlling how much of this material is in the glassy state, the device can hold different levels of electrical conductance, effectively storing a number. The researchers' goal was to create a version of this memory that uses very little power to write new information and can be read with extreme precision, all while fitting into a tiny space.

The core of their innovation lies in the physical architecture of the device. Instead of the traditional vertical shape used in most memory chips, which resembles a mushroom, they engineered a structure with an ultra-thin layer of the phase-change material, only about 5 nanometers thick. This layer sits between a bottom electrode and a dielectric spacer, but crucially, it rests on top of an even thinner, non-insulating projection layer made of amorphous carbon. This projection layer acts as a continuous bridge that spans the entire width of the device. When the researchers apply a short, intense burst of electricity to switch the material, they melt a small circular region of the phase-change layer right above the bottom electrode, turning it into the glassy state. The electrical current then flows through this small glassy circle and the surrounding projection layer, which are connected in parallel. Because the projection layer is continuous and highly conductive, it dictates the overall electrical behavior of the device, ensuring that the current remains low and stable.

This specific arrangement allows the device to achieve a level of precision that was previously difficult to reach. The researchers demonstrated that they could program the device to hold distinct conductance values with an accuracy approaching six bits, meaning it can reliably distinguish between 64 different levels of resistance. In practical terms, this high precision is essential for performing the complex matrix calculations required by artificial intelligence. Furthermore, the design drastically reduces the energy needed to write data. The team measured the current required to switch the device and found it could be as low as 90 microamperes, a significant drop from the hundreds of microamperes typically needed by older designs. This reduction is achieved because the ultra-thin geometry confines the heat generated by the electrical current so tightly that it melts only the necessary tiny volume of material, making the heating process incredibly efficient.

Beyond efficiency and precision, the new device solves a major problem known as temporal instability. In many memory technologies, the stored value tends to drift over time or fluctuate due to random electrical noise, which introduces errors into calculations. The researchers found that their design, with its continuous projection layer, significantly suppresses these fluctuations. They observed that the drift in the stored values was reduced by a factor of twenty compared to conventional devices, and the random noise was cut by a factor of four. This stability means that the device can hold its programmed state reliably, even at elevated temperatures, ensuring that the calculations performed by the memory remain accurate over time.

To prove that these individual devices could work together in a real system, the team built a small grid, or crossbar array, containing sixty-four of these units. They connected each memory cell to a transistor that acts as a switch, allowing them to control and read the cells individually. They then tested the array by programming it with a set of weights and performing a mathematical operation known as a matrix-vector multiplication, which is the fundamental building block of neural network inference. The results were striking: the hardware system achieved an effective computational precision of more than five bits. This level of accuracy is a substantial improvement over current state-of-the-art systems, which typically manage only three or four bits of precision. The success of this experiment suggests that the new architecture is not just a theoretical concept but a viable path toward building dense, low-power, and high-precision computing systems that could one day power the next generation of artificial intelligence.

The researchers also explored variations of this design to push the performance even further. By modifying the bottom electrode so that it penetrates directly into the stack of materials, creating an edge-contact configuration, they were able to reduce the programming current to approximately 50 microamperes. This change increased the local power density at the contact point, making the heating even more efficient. Throughout these experiments, the team used standard, undoped materials that are compatible with existing manufacturing processes, indicating that this technology could be integrated into current production lines without requiring entirely new fabrication methods. While the study focused on a specific proof-of-concept, the results provide a clear roadmap for overcoming the trade-offs between density, energy, and precision that have long hindered the development of in-memory computing.