Another benefit of such research is that hardware groups on a budget might be able to build analog accelerators on older nodes, like Skywater's.

Old or new, I want to see them done for training, not inference. We have to get training to real-time at good rates on hardware small labs can afford. Once that happens, many opportunities open up.