Abstract
For decades, conventional computers based on the von Neumann architecture have performed computation by repeatedly transferring data between their processing and their memory units, which are physically separated. As computation becomes increasingly data-centric and as the scalability limits in terms of performance and power are being reached, alternative computing paradigms are searched for in which computation and storage are collocated. A fascinating new approach is that of computational memory where the physics of nanoscale memory devices are used to perform certain computational tasks within the memory unit in a non-von Neumann manner. Here we present a large-scale experimental demonstration using one million phase-change memory devices organized to perform a high-level computational primitive by exploiting the crystallization dynamics. Also presented is an application of such a computational memory to process real-world data-sets. The results show that this co-existence of computation and storage at the nanometer scale could be the enabler for new, ultra-dense, low power, and massively parallel computing systems.
Introduction: The von Neumann Bottleneck
Modern computing faces a significant challenge known as the von Neumann bottleneck. In standard architectures, the central processing unit (CPU) and memory are physically distinct, requiring data to be constantly shuttled back and forth. As applications become more data-intensive—such as real-time sensor processing and large-scale machine learning—this data movement accounts for a disproportionate share of energy consumption and latency.
To address this, researchers are exploring "computational memory," where the memory devices themselves perform mathematical operations. By moving beyond the paradigm of memory as a passive storage bucket, this approach enables massively parallel processing directly within the memory array.
Methods: Harnessing Crystallization Dynamics
The core of this demonstration is the use of Phase-Change Memory (PCM). PCM devices store data by switching a chalcogenide glass material between an amorphous state (high resistance) and a crystalline state (low resistance). The "secret sauce" of this approach is the exploitation of crystallization dynamics: when a "SET" electrical pulse is applied, the amorphous region shrinks, and the device's conductance increases.
The researchers developed an algorithm that translates binary stochastic processes into specific electrical pulses. By modulating these pulses based on the "collective momentum" of multiple data streams, the resulting change in device conductance becomes mathematically equivalent to an estimate of the correlation between those streams. This allows the physical accumulation of atoms in the crystal lattice to perform the integration required for statistical tasks.
Experimental Results
The team implemented this system on a prototype platform featuring a 90nm CMOS chip with 3 million PCM devices. They evaluated the system using two primary datasets:
- Synthetic "Einstein" Dataset: The system processed one million artificial binary processes, successfully identifying the ~9.5% that were mutually correlated to reconstruct a 1000x1000 pixel image of Albert Einstein from what appeared to be pure noise.
- Real-World Weather Data: Using rainfall data from 270 NOAA weather stations, the computational memory identified geographical clusters. The results were highly consistent with standard k-means clustering algorithms but achieved with significantly less computational overhead.
The authors estimate that a dedicated computational memory module could accelerate correlation detection tasks by up to 200x compared to high-end conventional systems using POWER8 CPUs and Nvidia P100 GPUs.
Conclusion and Future Outlook
This large-scale experimental demonstration shows that computational memory can effectively perform high-level computational primitives by exploiting nanoscale physics. While hurdles such as device variability and limited dynamic range (saturation) remain, the results indicate that co-locating computation and storage is a promising path toward ultra-dense, low-power, and massively parallel computing systems. Future work will focus on extending these principles to other primitives, such as matrix-vector multiplication, and improving device reliability.
Summary updated