New memristor better mimics synapses | Hacker News Reader

New memristor better mimics synapses

news.umich.edu

90 pointsmangoleaf18 comments

At the device level, variability remains a significant issue for all memristor types. Scaling feature sizes down is likely to make variability worse. https://semiengineering.com/integrating-memristors-for-neuro...

Was the transistor theoretically proposed similar to the memristor was? Wondering if people had as much hype for the transistor eventually appearing as we are about memristors.

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

Electroplating baths have a "memory" of how much current has passed through them. Edison used them as a way of measuring DC power use.

https://ethw.org/Electric_Meter

So the memristor has been physically proven? I thought there was skepticism over whether it was even possible to make them.

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?)

I heard an (albeit unsubstantiated) rumor that Meg Whitman sold off exclusive use to memristors to a company that's just sitting on them rather than turning them into a product. Hence why the machine started to divorce itself from memristors, but then pretty quickly was canceled.

I believe that there are lots of professionally-run science scams.

It starts with an idea that looks good on paper and a bunch of scientists investigating it. At one point it gets picked up by the management of a company and commercialization is being started while the technology is not ready yet. Now you have high-risk development of a new technology in the framework of commercial management. Failing is no option anymore. During the development it becomes clear to some insiders that the goals cannot be achieved. To those are three options. Either they quit, they shut up or they upsell what they have. Plans become bigger and bigger and the deadline shifts later and later. The later you are, the more sunk costs you have and the bigger your plans need to be to recoup the costs and keep your department alive. This is what I consider the when it becomes a scam. It started with science but instead of burying the dead horse it is paraded around town and management applauds while those who know do not talk.

Being one of the scientists who thought they were doing good by initially starting the research must be indescribably painful. I despise when good will is turned on its head. It turns good people into bitter and regretful pessimists. Rather than fight, they quit, like you mentioned - thinking its better not to play twisted games. Only later in life do they realize they should have fought like mad, because no evil should be given quarter, lest it eat you up for years to come. Let's just say I have experience in this type of regret :-/

A lesson for all: the bigger man does not quit, does not keep quiet, does not pacify in the face of evil. Contrary to the truthisms about being the bigger man - fighting for what is right is almost always the right move.

It is often said one should be mature and not "play the game." But the truth is, evil affects us all and isn't confined to isolated games. If you don't stop it locally, it will spread like a poison globally.

Learn from my mistake. Play the game, fight, dont quit.

Quite common in tech too.

Theranos, IBM "Watson" (in reality a family of products sold as one "AI" thing leeching the branding of the chess playing AI, and very underperforming), the "quantum computing" thing...

(I'd add Tesla "auto-but-not-really-auto" pilot...)

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.

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/

Here is a better DIY link: http://sparkbangbuzz.com/memristor/memristor.htm

And here's a where I did asked previous questions on the topic: https://electronics.stackexchange.com/questions/76097/what-a...

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.

What are the merits of memristors over solid state NAND and Intel optane technology?

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

Flash cells are analog as well - depending on the gate charge the resistance/current flow is different.

That's used for MLC/TLC.

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.)

Reply on news.ycombinator.com