| Project Detail |
Integrating vision and cognition to power neuromorphic computing
The field of computing with light has seen significant advancements over the past few years, with applications ranging from neuromorphic computing to quantum computing. However, it has been difficult to integrate all of these functionalities onto a single chip. The EU-funded RESPITE project will combine vision and cognition on a single chip. This will enable in-sensor neuromorphic computing with unparalleled performance levels, thanks to the use of superconducting Joule switches as neurons, multi-level phase change memory elements as synaptic weights, and superconducting single-photon detector arrays as retinas. Scalability, ease of fabrication, and compatibility with low-cost cryostats, high-Tc superconductors, quantum applications, and on-chip learning architectures will set this platform apart.
Computing with light using integrated optics has seen huge progress over the last 3-4 years in multiple fields such as neuromorphic computing, quantum computing and on-chip data storage. This has created a vast ecosystem that relies on high-speed reconfigurations of nanophotonic circuits (such as their use as synapses or routing applications) and ultrafast yet high-resolution, low-power photodetection. Currently, it is impossible to combine all these functionalities into an integrated platform that fits onto a single chip. In RESPITE, by utilizing our newly invented superconducting Joule switches as neurons, multi-level phase change memory elements as synaptic weights, and superconducting single-photon detector arrays as retina we will demonstrate a novel platform which combines vision and cognition on a single chip. This new platform will allow in-sensor neuromorphic computing with unprecedented performance levels. The platform will have attoJoule switching power consumption, sub-nanosecond latency, and high compactness (3000 neurons and >100K synapses on <5 mm2). Unlike other superconducting neuromorphic technologies, our new platform will be scalable, easy to fabricate, and compatible with low-cost cryostats, high-Tc superconductors, quantum applications, and on-chip learning architectures – making it a game changer for a wide range of users and disciplines. |
