Abstract
A key bottleneck in using phase change memory (PCM) for training artificial neural networks is the abrupt nature of the melt-quench process (amorphization), which does not allow gradual reset. Here we demonstrate analog reset (partial amorphization) in PCM by applying sub-nanosecond programming pulses. Intermediate-level reset states are enabled by reducing the pulse width below the dominant thermal time constant of the PCM, which is on the order of a few nanoseconds. We show gradual change in PCM resistance as a function of number of (sub-ns) reset pulses with 50 intermediate states. Our unique scheme allows fine-tuning the resistance with sub-ns pulses of constant amplitude, which can significantly reduce the programming complexity in training neuromorphic hardware.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call