Inferencing, Fast and Slow with Ultra-scaled Phase-Change Devices

A major challenge for deep learning inference is the high energy demand required to retrieve large amounts of synaptic weight data from memory. One promising approach to address this is the use of conductance-based devices, such as non-volatile phase-change memory, to develop chips with stationary synaptic weights. However, two key obstacles remain: enhancing the computational capabilities and increasing the energy efficiency of these devices. INFUSED tackles both issues through groundbreaking device innovation. By utilizing the physics of ultra-scaled materials, it pushes energy efficiency closer to its theoretical limits. Moreover, it introduces dual neurally-plausible temporal dynamics, combining fast adaptive responses with slow, gradual conductance changes. This reimagines traditional neural network elements like the perceptron for more energy-efficient AI inference.INFUSED specifically aims to:☞ Develop a breakthrough device design method to encode slow weights efficiently by reducing electrical contacts and active volumes to unprecedented scales—below 10 nm², and using 1-nm-thick van der Waals films. These weights are programmable with energy in the tens of femtojoules, offering up to a 100x improvement over current industry devices.☞ Design a reconfigurable volatile memory to encode fast weights within the slow weights. This innovative method uses field-effect properties of the ultra-scaled volumes. These weights exhibit non-linear temporal dynamics, consuming energy in the picoWatt range, with timescales spanning microseconds to hours, adaptable to AI tasks.☞ Demonstrate a mixed hardware-software implementation of biologically inspired algorithms leveraging fast and slow dynamics. New neural architectures are developed and benchmarked against mainstream neural networks, as well as against hardware substrates like parallel computing architectures (e.g., GPUs), in complex vision tasks.