The 14-Bit Breakthrough, the 8-Bit Bet

A Bangalore team published a molecular memristor in Nature last year with 14-bit precision — 16,384 distinct conductance states in a single memory cell, IISc reported, a result the paper called a leap forward in computing efficiency. The startup spinning out of IISc's Center for Nano Science and Engineering is now trying to put that research into a chip you can actually buy. Their target: 8-bit resolution on TSMC's 22-nanometer process, with a 4-bit power-saving mode for ultra-low-demand inference. A minimum viable product is planned within a year, EE Times reported.

That is not a failure. It is the manufacturing process imposing a tradeoff the prototype never had to make.

The technology is a ruthenium azo-aromatic molecular film that stores and computes in the same place — eliminating the energy shuttle between memory and processor that plagues conventional chip architectures. For convolutional neural networks running on standard hardware, the majority of inference energy is spent not on the math but on moving data on and off DRAM, a problem the field calls the memory wall. The Nature results demonstrated 4.1 trillion operations per second per watt using this approach.

The ruthenium film enables what chemists call zeroth-order kinetics: switching at a constant, predictable rate regardless of starting state, which produces the linear symmetric conductance updates that neuromorphic training requires. The Nature paper describes the mechanism as producing "linear symmetric self-selecting 14-bit kinetic molecular memristors." Conductance states have been tested after seven to eight months with no measurable drift. Endurance was demonstrated to one billion rewrite cycles against roughly one million for conventional flash.

The CeNSE startup will license the patents from IISc. The 22-nm TSMC node is an older, cheaper process — sensible for a device competing on energy efficiency rather than raw compute speed. Ruthenium supply is a practical concern; the team is exploring iron and cobalt alternatives for later versions, EE Times reported.

The 8-bit target is a deliberate engineering compromise. Many edge inference tasks run adequately without 14-bit precision. The question is whether the molecular film can be redesigned to work inside standard foundry tooling at production yield — a fundamentally different challenge from demonstrating the physics in a Bangalore lab. The Nature paper answers the first question. The startup is betting it can answer the second.