CISE-ANR: FET: Small: ASTROCOMP- Advanced Stochastic Magnetic Tunnel Junction Architectures for Scalable Neuromorphic Computing

About This Grant

Artificial intelligence (AI) needs random numbers for its complex computing tasks. A random number is a number randomly chosen from a given distribution or a set of numbers, where there is no relationship between numbers chosen one after the other. There is an increasing demand for drawing random bits from specific distributions such as uniform, Gaussian, and exponential, which are critical for efficiently solving complex computing tasks. For example, Gaussian distributions play a key role in generative and diffusion models used for large artificial intelligence applications. However, generating trillions of numbers from these distributions in a fashion that is truly random and ultra-fast within fractions of a second is a challenging task using the current hardware. The need to innovate new types of hardware based on emerging devices to generate these random numbers at scale is now critical. This project aims to develop such hardware based on new type of devices called stochastic magnetic tunnel junctions. The stochasticity in these devices helps at the physical level to generate random bits from distributions of interest with high computational efficiency. This project will fabricate stochastic magnetic tunnel junctions integrated with commercially developed silicon wafers with high yield and reliable switching and test this new hardware for random number generation in AI tasks. This project will involve students at different educational levels, e.g., graduate and undergraduate, and enhance hardware prototyping platforms of interest for future application-relevant translation. This project is organized into three objectives. The first objective is to iteratively optimize the stochastic magnetic tunnel junctions fabrication process on wafers with similar surface properties as the commercially planarized wafer. The device properties will inform circuit designs for small scale demonstrations of problems covered in the following objectives. The second objective is the co-integration of stochastic magnetic tunnel junctions developed in the back-end-of-the-line of a foundry-planarized memory characterization array. The crossbar array consists of 20,000 crosspoint memory cells partitioned into four banks of 5,000 devices in a 180nm process, with transistor-based circuit multiplexing, control and compliance infrastructure. The third objective is the development of peripheral circuits that enable statistical sampling from uniform, gaussian and exponential distributions. The project team is in a unique position to achieve high yield and reliability by leveraging pre-existing knowledge in stochastic magnetic tunnel junctions fabrication at SPINTEC (France) and foundry-planarized mid-production wafers with exposed vias developed specifically for memory characterization by members at George Washington University (USA) in a previous program. This will enable iterative refinement of the stochastic magnetic tunnel junctions process over a fabrication interface with minimal surface roughness by the French team followed by an automated characterization schedule with probe card instrumentation developed by the American team. The resulting demonstrators will illustrate how physical randomness and stochastic switching can be harnessed for tasks such as data generation, classification, and time-domain inference, offering a better alternative to purely digital or software-based probabilistic computing approaches. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.