illustration of connections in the brain
illustration of connections in the brain

Disruptively green neuromorphic scientific computing leveraging stochasticity

Picture: Dominique Kösters

The massive use of AI algorithms rapidly increases the energy consumption of computing systems. An appealing and possibly more sustainable alternative is provided by so-called neuromorphic paradigms: brain-like computer architectures which can drastically reduce data transfer by co-locating memory and computation. However, it is still unclear how neuromorphic hardware concepts can optimally be deployed in terms of computational capacity and environmental impact.

Algorithms and hardware materials

In this project, we will develop algorithms for neuromorphic hardware with computational material science as a use case. This will pave the way towards solutions for problems in optimization and sampling, creating more fundamental knowledge on next-generation materials for neuromorphic hardware. In this way, we will create a positive feedback loop for the development of both algorithms and materials systems for new neuromorphic hardware. We will focus on the design of stochastic algorithms which directly harness the stochasticity of the underlying materials systems. This is not yet feasible today and would be a real breakthrough.

Many computational tasks can be efficiently solved by stochastic algorithms. For example, instead of carefully including all possible terms in an expression, a systematically improvable solution can be obtained by randomized sampling the most important terms. However, these are deployed on digital hardware. With further miniaturization, stochastic effects inevitably emerge and are currently mitigated. At Radboud University, we have key expertise in both stochastic algorithms, their application in computational science as well as the characterization of stochastic properties of nanoscale magnetic materials. Together with Mainz, which is a key expert in the discovery and growth of stochastic magnetic materials we have all the expertise to discover how synchronized stochastic properties can be harnessed for computations as opposed to being mitigated. Moreover, with the expertise of SURF, Fontys and InfinityQ we are able to explore applications of the results in science and society.

Objectives

The objectives of our research programme are:

  1. Develop new computing algorithms to harness synchronized stochasticity of memristive nanomaterials
  2. Demonstrate implementation of these new algorithms in neuromorphic materials
  3. Demonstrate a neuromorphic advantage for state-of-the art use cases in computational physics

Consortium

To realize these goals, we combine expertise from artificial intelligence (RU-AI), computational (RU-physics) and experimental (JGU Mainz-physics) physics and materials sciences, high-performance computing (SURF), and computer hardware and software engineering (InfinityQ). Together we develop algorithms, implement them in recently discovered ultra-fast, ultra-small and ultra-low-energy stochastic dynamics in magnetic materials, and benchmark (Fontys) our method for runtime and energy-efficiency against traditional computing methods.

Project leader Johan Mentink: “In this project we will demonstrate the potential of stochastic neuromorphic hardware, both to radically reduce the energy consumption and to perform so far inaccessible large-scale simulations in science and beyond.”

 

 

Funding

Partners

The project is led by Johan Mentink from the Institute for Molecules and Materials

Contact information

More information? Please contact our press officers at 024 361 6000, media@ru.nl or the project members.