i'm currently more intereseted in nin-memory analog computing, such as with phase change memory or resistive ram memory. it gives a speedup of over 1,000x a GPGPU and consumes 1/1,000 of the power.
A big limitation of it, however, is that while it's great for implementing a trained neutal network (forward propagation), i don't see a way to get that kind of 1,000x improvement in the training / back propagation phase.
it might be possible thought. many years ago i came up with a circuit for an analog adaptive control circuit, which is kind of similiar but integrates to in memory compute but integrates the training on the same plane.
maybe people could work on circuits that combine / synthesize the two approaches; neuromorphic + in memory compute.