Materials for Neuromorphic Supercomputers

Many efforts in neuromorphic computing seek extreme energy efficiency for edge devices by leveraging principles of neural information processing. Fewer efforts aim to capture the full complexity and scale of biological brains, potentially implementing neural principles in technological hardware at a scale that surpasses that of the human mind. For this second thrust, which we term “neuromorphic supercomputing”, the materials challenges are quite unique.<br/><br/>Utilizing existing silicon microelectronics is a natural choice for neuromorphic supercomputing. Yet when implementing neuro-inspired circuits, architectures, and algorithms with existing silicon microelectronic hardware, communication bottlenecks emerge as a major limitation. Biological neurons make thousands of connections to near and distant synapses, whereas silicon neurons cannot achieve this level of fan-out due to wiring parasitics. Address-event representation of spikes is employed with a shared communication infrastructure. As the number of neurons grows, that infrastructure becomes overwhelmed, and the rate at which each neuron can spike drops below the rate of biological neurons, rendering the technological approach inferior to the biological one, even for relatively small systems. This is a severe obstacle for neuromorphic supercomputers.<br/><br/>To overcome this barrier, we use light for communication. A neuron spike is represented by a pulse of photons distributed to synaptic connections via nanophotonic waveguides, circumventing wiring parasitics that prevent direct electrical communication. Superconducting single-photon detectors allow communication to occur at the lowest possible light level for extreme energy efficiency. For computation, the photonic components must be paired with electronic devices and circuits. Using superconducting single-photon detectors requires low temperature operation, so other superconducting elements such as Josephson junctions can be used as well. To realize the full spectrum of synaptic, dendritic, and neural functions, a combination of cryo-CMOS and Josephson circuits are uniquely powerful.<br/><br/>The full hardware stack for this neuromorphic supercomputing platform includes conventional CMOS working in conjunction with superconducting electronics, semiconducting light sources, multiple planes of passive dielectric waveguides, and single-photon detectors integrated with those waveguides. To realize this full stack, multiple challenges in materials science remain. First and foremost, we must integrate light sources with CMOS electronics. Such a feat has been a goal of the industry, yet this context provides one crucial advantage: low-temperature operation. Silicon itself is a reasonable light emitter at low temperature, provided the crystalline lattice is modified to contain emissive centers. The optimization of silicon for this unique optoelectronic purpose represents an important objective of this research. A second objective relates to the single-photon detectors. For scalability, we seek to operate at 4.2K. Polycrystalline materials such as NbTiN make excellent single-photon detectors with superconducting transition temperatures above 4.2K, but the yield and detection plateau are sub-optimal. Other amorphous materials such as MoSi and WSi have much higher yield and detection plateau, but they typically must be operated between 1K and 2K. Finding a material with the best of both worlds remains an important goal for the project as well. Finally, to achieve large systems with the number of neurons and synapses as the human brain will require many planes of waveguides and active superconducting devices. Numerous materials integration challenges must be solved to realize this sophisticated microelectronics process at the 300mm scale. If these problems can be solved, the world will have access to a new domain of advanced computational technology for neural information processing, achieving several important fundamental physical limits of cognition.