New memristor better mimics synapses (opens in new tab)

According to [1], "[b]ecause the production of high-quality semiconductor materials was still decades away, Lilienfeld's solid-state amplifier ideas would not have found practical use in the 1920s and 1930s, even if such a device had been built."

It seems like there are parallels. Chua theorized about the memristor in the 1970s, but the first one wasn't created until the 2000s. A question I don't have the answer to: a transistor is to a vacuum tube as a memristor is to what?

[1] https://en.wikipedia.org/wiki/Transistor

HP/HPE has patents on them and has supposedly made a ton in their labs, but hasn't seemed to get the costs down to make them commercially viable, and may have gotten trapped into some (interesting, but diverting) rabbit holes in the process not helping them commercialize the existing work, such as believing they need to reinvent the computer almost entirely from scratch just because of the memristor [0].

[0] "The Machine". There's a lot of weird stuff around on it. Official hype page of the (now cancelled, presumably) project: https://www.labs.hpe.com/the-machine (Makes some of IBM's Watson marketing almost look sane in comparison, doesn't it?)

Depends on what you mean by physically proven. An ideal memristor is physically impossible but devices whose behavior is dominated by its memristance function over an operating range are currently available for purchase. I would say that constitutes a practical memristor.

HP Memristors are suspended nano-wires surrounded by a floating ring of titanium oxide insulator forming a coaxial line whose impedance/resistance is a function of the cross section of the oxide ring. Current causes the ring compress or stretch giving a memristance function of roughly a parabola which saturates when the oxide is fully stretched or compressed. Last I looked, there are a lot of legal battles currently ongoing and patent litigation involving HP and memristors which is tying the technology up.

Other Memristors commonly work by dendrite formation along electric fields but this method is pretty dang slow. Charge trapping semi-conductor junctions have been proposed but I haven't seen any myself.

This new memristor from U of M works by controlling Lithium diffusion via an electric field. The amount of Lithium available controls the formation of lattice networks in the substrate which in turn controls the conductivity of the material.

Need to catch a ride, but here's a quick link where you can make your own memristors and see for yourself. Basically requires two dissimilar metals and an oxidizer:

https://makezine.com/2011/11/03/how-to-homemade-memristor/

Most of the materials used for this purpose have properties where the resistivity changes as a function of the charge which has gone through it. You can reset the material by a thermal cycle or by applying a larger current through it.

Solid state non-volatile storage technologies such as NOR and NAND flash memory use something called floating-gate transistors [1]. Making these transistors is highly complex, and substantial amounts of research efforts have gone into R&D to improve NOR and NAND technologies.

A real issue with these memory technologies is wear. There is a finite amount of times a NOR/NAND gate can be written, and the implementation of wear leveling prolongs the lifespan of the memory device without solving the root issue.

Optane memory (originally called 3D XPoint) starts getting closer to the theoretical ideal of a memristor. According to [2], Optane allows a "memory cell to be written to or read without requiring a transistor." It's still not entirely sure what's the enabling technology, but the general consensus is that Optane is a form of resistive RAM, or ReRAM [3]. Not quite the memristor, but it's edging toward it.

Why memristors? Imagine you're given a piece of conductive material which changes its resistivity based on net current flow over time. It's just a chunk of material, so manufacturing memory cells goes from etching transistors to material deposition. Also, you can likely make the memristors smaller than today's transistors, meaning higher storage densities. Additionally, the memory should be bit-addressable unlike block-addressed NAND. Also, the memristor is fundamentally an analog device, meaning there's the potential for interesting innovations in storage technology (analog storage elements?). Finally, energy consumption should be even lower than transistors.

Takeaways: - NOR/NAND storage is transistor-based and will eventually be replaced with neuromorphic technologies such as memristors. - Optane storage exhibits expected properties of a memristor, but it's not quite a true memristor.

[1] https://en.wikipedia.org/wiki/Flash_memory [2] https://www.anandtech.com/show/9541/intel-announces-optane-s... [3] https://en.wikipedia.org/wiki/3D_XPoint

In theory memristors should be faster than flash and with significantly better write endurance. Additionally, the memristor can be used as the basis of computing. Meaning you could potentially create a device which had programmable logic like an FPGA but could reconfigure itself much faster and operate at the speed of a dedicated ASIC (CPU/GPU, etc.)