The Normal Blog – A First Demonstration of Thermodynamic Matrix Inversion

I bet there is huge potential in physical computation but it seems like we'll need post-silicon or massively-3D circuitry to scale it out sufficiently. Bio-electronics and organically grown components are possible but also in pure R&D stages.

Imagine growing a gigantic organic 3D circuit with trillions of electrically charged connections, capable of trillion dimension matrix operations needing a fraction of current silicon based power requirements.

We could give it a great name like Synaptic Architecture, or Neural Fibre or ...

I’m a big fan of Crooks and thermodynamic computing in general, but I’d like to see a plot comparing against traditional methods in terms of real seconds. I don’t expect them to beat GPUs given how mature GPUs are but the big promise is that these methods can converge to a decent result faster than traditional algorithms. I’d like to see some empirical result showing that promise.

As a caveat it’s pretty late so I only skimmed it but I’ll check it out in detail tomorrow

Off topic a bit but does any one know whether matrix inversion is doable with just optics? There is a a small bunch of photonic processors doing matrix multiplication at high speed using (IIUC) reconfigurable mirror arrays. Just wondering whether matrix inversion is possible too.