SPX: Collaborative Research: Automated Synthesis of Extreme-Scale Computing Systems Using Non-Volatile Memory

The project investigates the design of a scalable computing infrastructure that uses nanoscale non-volatile memory (NVM) devices for both storage and computation. The project's novelties are (i) the use of multiple parallel flows of current through naturally occurring sneak paths in NVM crossbars for computation; (ii) the replacement of slow organic expert-driven discovery of flow-based computing designs by automated synthesis techniques for accelerated discovery of novel NVM crossbar designs; and (iii) a pervasive focus on fault-tolerance throughout the design of exact, approximate and stochastic flow-based computing designs. The project's impacts are (i) the design of an end-to-end framework that maps compute-intensive kernels written in a high-level programming language onto nanoscale NVM crossbar designs and (ii) the creation of a new scalable capability to perform exact and approximate in-memory digital computations on fault-prone nanoscale NVM crossbars. The team of computer scientists and nanoscience researchers is creating flow-based computing designs for four benchmark problems: the Feynman grand prize problem, computer vision, basic linear algebra, and simulation of dynamical systems. The automatically synthesized NVM crossbar designs are being evaluated using high-performance simulations and experimental benchmarking in a modern nanotechnology laboratory. Computing using multiple parallel flows of current through data stored in nanoscale crossbars is often fast and more energy-efficient, but the design of such crossbars is highly unintuitive for human designers. The project explores a combination of formal methods for checking satisfiability of Boolean formulae, and artificial intelligence techniques such as best-first search, to automatically synthesize NVM crossbar designs from specifications written in a high-level programming language. The team of computer scientists and nanoscience researchers is pursuing a transformative agenda for extreme-scale computing by leveraging memory devices in NVM crossbars as structurally-constrained fault-prone distributed nano-stores of data, and exploiting the natural parallel flow of current through NVM crossbars for computing over data stored in the distributed nano-stores. The NVM crossbar designs generated from OpenCV, LAPACK, and ODEINT programs are evaluated using the Xyce circuit simulation software and subsequently fabricated for experimental benchmarking. By combining storage and computation on the same device, the project circumvents the von Neumann barrier between the processor and the memory and creates scalable solutions for extreme-scale computing on fault-prone NVM crossbars without introducing substantial changes to the programming model. 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.