I think the question that comes to mind is whether it's possible to take GPU-styles architecture and give it 100x more power or more by replacing bits with approximate voltage levels that are "fuzzy" but statistical guarantees to their performance, along with gates that allow the values to be filtered back to zeros and ones.
The thing is the "neural architectures" seem to have been more or less failure through requiring the user to accept one particular neural net structure while the standard GPU seems to have succeeded through being generic-enough to use for a variety of tasks. So some sort of analogy-GPU should also have a similar generic model - but of course it seems likely that the creator of this stuff will want to impose their special model.
Edit: More or less like the D-wave "quantum computer" except not milking quantum hype and being reconciled to being understood as "massively analogue"