Navigation and service

Neuromorphic Compute Nodes (PGI-14)

Nature has been building information-processing machines – computers – for billions of years, far eclipsing the ~100 years of computer engineering by modern humans. Nature evolved these machines through trying and failing with an estimated billion different species and experiments. It would be difficult for even human ingenuity to reproduce this exhaustive search. The Peter Grünberg Institute for Neuromorphic Compute Nodes (PGI-14) aims to explore ways to selectively copy the tricks that evolved in biological information processing systems. Our goal is to build more energy-efficient hardware, solve today’s intractable problems, and even attempt to shed light on the underlying compute principles used by biology (brains) while possibly inventing some new ones of our own.

With the ambitious goal of making more efficient and more capable computing systems, we are in an era where “software-hardware co-design” is an absolute necessity. The entire compute stack must be addressed simultaneously, punching through the layers of abstraction we have built up in the past decades and working in cross-disciplinary teams. Evolution is the original co-designer: changes in DNA can impact everything from the large-scale system architecture, the timing and connections of neural circuits, to low-level cell (devices) functionality and diversity. Evolution turns every knob at the same time, building end-to-end optimized information processing engines. We need to do the same and we are uniquely poised at the Peter Grünberg Institute to accomplish this, through tight collaborations from materials and devices (e.g., PGI-7 and PGI-10), hardware architectures in PGI-14 (us), neuromorphic software ecosystem and algorithms (PGI-15), large-scale supercomputing (JSC), and neuroscience research and understanding (e.g., INM-6).

We wish our R&D to have a positive impact on society and so we seek out practical application areas including scientific computing, machine learning, intractable NP-hard problems, and signal processing. Some of these are highlighted here and in past publications. We especially utilize novel emerging technologies – memristors – which offer analog tunability, data persistence, and highly non-linear dynamics of great potential. It has even been argued that brains are built of memristors [L. Chua, et al. (2012)]. We combine these emerging technologies with modern CMOS elements to build practical circuits and integrated chips for hands-on demonstrations in our laboratory.