Takes me back... There was significant hype around those things when they first managed to build them at scale (~15 years ago), because they were promising for low power, high density persistent storage and are also academically interesting: The "concept" of memristors was explored over 50 years ago (they are passive components that couple electrical charge and magnetical flux, just like a resistor does with current/voltage, a capacitor with voltage/charge or an inductivity with current/flux).
But I think the main problem was that they never managed to scale up the clock speeds sufficiently, even though structure size (=> density) was already highly promising from the start.
Maybe in a slightly different history with some discoveries in different orders these could have replaced flash memory in SSDs completely.
But that whole episode thought me that betting on early technology is hard, and always a risky business, because no matter how promising an approach looks, if it turns out that you can not find the necessary improvements in only a single dimension, then the whole thing is kinda doomed and will probably never be competitive (=> a highly relevant insight especially when speculating about things like novel battery chemistries or the like).
> if it turns out that you can not find the necessary improvements in only a single dimension, then the whole thing is kinda doomed and will probably never be competitive
I don't know, we've been working on digital computers since at least the late 1800s. Sometimes technology just takes a while.
That does make it hard to gamble on it if the time horizon is longer than you need to make a profit.
But I don't think we should convince ourselves that a technology that takes longer than 15 years to become profitable is doomed. If we thought like that we'd still be subsistence hunter gatherers.
Absolutely! For the record: I don't think that memristors are doomed to be useless-- we'll have to find out.
My point is just that even with research-tech that sounds absolutely amazing (low power, persistent, high density) you just need to fail on a single dimension for it to basically become irrelevant.
This is also why its so easy for media to overhype research results, which (predictably) results in continuous disappointments and loss of trust (of the public) in science reporting and/or even science in general...
And we only have blue LEDs due to the sheer stubbornness of Shuji Nakamura. If you haven't already heard the tale of its development, head over here and enjoy: https://www.youtube.com/watch?v=AF8d72mA41M
> To address the challenge of EUV lithography, researchers at Lawrence Livermore National Laboratory, Lawrence Berkeley National Laboratory, and Sandia National Laboratories were funded in the 1990s to perform basic research into the technical obstacles. The results of this successful effort were disseminated via a public/private partnership Cooperative R&D Agreement (CRADA) with the invention and rights wholly owned by the US government, but licensed and distributed under approval by DOE and Congress.[3] The CRADA consisted of a consortium of private companies and the Labs, manifested as an entity called the Extreme Ultraviolet Limited Liability Company (EUV LLC).[4]
> Intel, Canon, and Nikon (leaders in the field at the time), as well as the Dutch company ASML and Silicon Valley Group (SVG) all sought licensing. Congress denied[citation needed] the Japanese companies the necessary permission, as they were perceived[by whom?] as strong technical competitors at the time and should not benefit from taxpayer-funded research at the expense of American companies.[5] In 2001 SVG was acquired by ASML, leaving ASML as the sole benefactor of the critical technology.[6]
>By 2018, ASML succeeded in deploying the intellectual property from the EUV-LLC after several decades of developmental research
They never promised high-density. Semiconductor memristors were always fated at staying at a much lower density than the same amount of capacitive memory. And that's before you get into the manufacturing issues and the problem that it loses "data" when read.
Those things where hyped out of nowhere, with lots of blatant lies making into the popular discourse (like that high-density prediction). I don't even know why, because nobody was making any serious bet on them. They are a very interesting design, that may still get some real-world usage (the manufacturing problems are a showstopper right now), but won't ever compete with flash.
Is that actually the case? Or have memristors just proven to be a 'boring' technology that's just quietly replaced other bits and pieces that we don't hear about? A bit like graphene was supposed to be this wonder material, and now it's found in the soles of trail and hiking boots.
> Or have memristors just proven to be a 'boring' technology that's just quietly replaced other bits and pieces that we don't hear about?
As far as I know, they have no application apart from academic toy/reseearch subject right now. And you have to consider that there are a lot of niches for storage technology that they could have taken over (because there is a lot of tradeoffs to make, e.g. latency, bandwidth, persistence, density, power consumption).
We might be just a few breakthoughs from those things replacing flash memory in SSDs, or revolutionizing neural-network accelerator hardware, but I am quite skeptical for now.
Note: I still believe that this (and other stuff i'm skeptical about) is SUPER worthwhile to research and always a huge uphill battle, simply because we have invested hundreds of billions of dollars into improvements of CMOS technology and processes, and collected over half a century of experience with it...
But new tech is to me kinda like a startup-- not every technology is the future, just like not every startup is a unicorn. Investing is still the right move, but you have to be realistic about expectations (which modern media is absolutely not)
This was always a strange point of contention - Intel denied using memristors. I assume there were some sort of patent or trademark issues.
WP:
"Development of 3D XPoint began around 2012.[8] Intel and Micron had developed other non-volatile phase-change memory (PCM) technologies previously;[note 1] Mark Durcan of Micron said 3D XPoint architecture differs from previous offerings of PCM, and uses chalcogenide materials for both selector and storage parts of the memory cell that are faster and more stable than traditional PCM materials like GST.[10] But today, it is thought of as a subset of ReRAM.[11] According to patents a variety of materials can be used as the chalcogenide material.[12][13][14]
3D XPoint has been stated to use electrical resistance and to be bit addressable.[15] Similarities to the resistive random-access memory under development by Crossbar Inc. have been noted, but 3D XPoint uses different storage physics.[8] Specifically, transistors are replaced by threshold switches as selectors in the memory cells.[16] 3D XPoint developers indicate that it is based on changes in resistance of the bulk material.[2] Intel CEO Brian Krzanich responded to ongoing questions on the XPoint material that the switching was based on "bulk material properties".[3] Intel has stated that 3D XPoint does not use a phase-change or memristor technology,[17] although this is disputed by independent reviewers.[18]
According to reverse engineering firm TechInsights, 3D XPoint uses germanium-antimony-tellurium (GST) with low silicon content as the data storage material which is accessed by ovonic threshold switches (OTSes)[19][20] made of ternary phased selenium-germanium-silicon with arsenic doping.[21][22]"
IIRC, performance was fantastic, but they were never able/willing to match the data density and data cost improvements in stacked-NAND flash, and without forcing themselves into the market at competitive rates, nobody wanted to write applications or design hardware suited to their unique strengths as low-latency caches.
There is still, to this day, a numerical niche for these drives, which is being served imperfectly by either normal TLC drives of very large size, SLC cache drives, or DRAM expansion cards connecting to the CPU through a PCIE bus. Just not at the prices they wanted to charge.
But wasn't the potentially transformative market intended to be "persistent DRAM" for instant-on devices removing the distinction between memory and storage, requiring DRAM-like speed rather than NAND-like speed ?
I recall their early R/W speed performance projections being far faster than what they ever achieved with Optane drives.
The products that used a PCIe X4 interface with a block storage protocol layered on top were never intended to deliver the best performance the memory was capable of.
Interesting - I wasn't aware, but even avoiding the PCI bus the performance must have been lacking as that link talks of "memory tiering". I guess this was "mid tier" somewhere between SSD and DRAM, which is a bit of a no-mans land especially in terms of conventional systems architecture ... really just a fast type of storage, or storage cache (a bit like a hybrid SSD-HDD drive).
> graphene was supposed to be this wonder material, and now it's found in the soles of trail and hiking boots.
I mean, that's not because graphene has become a routine part of our material repertoire. It has no reason to be in those things, does nothing, and is just marketing fuel. We may put "graphene" in things, but we are not much closer to using its interesting properties.
I think that Memristors are perfect for use as configuration RAM for FPGAs and FPGA-like things. Something that you want to be able to update, but not frequently, and read all the time.
Of course, then the question becomes one of refreshing their state, like DRAM.
I remember that too; I was very, very interested, but it never materialized. Very disappointing.
That said, I think this is something a bit different, or at least a different application. If my translation of the summary is correct (I'm not very fluent in sciencese), it's basically using them as some kind of matrix multiplier rather than memory. Whether they're making use of power-off data retention at all was unclear to me, but then I just skimmed it.
Interesting, but I was really hoping for fast, persistent memory to appear.
What value does it provide over linking to the original source?
Any website that constantly asks me to login is spammy in by book. It's a for profit website that adds little value other than duplicating information from primary sources and occasionally mangling pdfs with redundant information to advertise themselves.
They also provide a central place to search things, with a richer interface than Google scholar, and have centralized a significant amount of good sources.
There's a reason millions of researchers have joined. That you don't find value or know what they provide is no reason others should not learn the value they add.
That's fine. You also miss all papers not on arxiv. You miss published versions of even the papers on Arxiv (which are often improved versions), and you miss any benefit peer review has on those papers.
I use arxiv nearly every day, and also a few places that get things not on arxiv because the majority of papers are simply not there. Arxiv is paid for by universities paying subscriptions, locked in for five years at a time. It's also funded by Simons Foundation (which may not pay forever) and Cornell and many individual donors. Affiliate groups like professional societies and govts pay huge sums to keep it running. Many companies pay 10's of thousands annually to be members.
Piggybacking on their money while taking affront at a bigger, more comprehensive service, because they dare post an ad, seems somewhat short sighted, but to each his own.
Since a significant number of job postings for researchers as well and communication and networking opportunities are widely used on Research Gate, none of which is present on Arxiv, you are simply missing likely useful contacts and tools for your career. And I write this as a researcher for several decades, long before any of these were live.
As I said, enough people find value at research gate that millions do pay.
Just a bit of context: MRAM exists as an IP for long time, but their promise to execute code directly from it didnt really pan out because of speed. So it competes against flash to store the code that is booted into SRAM and it loses there too because of mostly larger area
I think memristors are technically more closely related to ReRAM (RRAM) than to MRAM. Though ReRAM so far also just unsuccessfully competes against NAND flash.
The brain is running on 20W of power and it has the best LLM, the best robotics control unit, very good sensor integration and all the other exciting stuff which we* want** AIs to have. I'd rather have that than nuclear powerplants feeding data centers.
Not really. Consider private equity’s solution to our doctor shortage. Import foreign trained physicians without requiring additional training. Give them access to an electronic medical record eith AI. Teach them how to click a mouse so they can copy/paste notes.Use digital real time translation to allow anyone to talk to anyone. Have them ‘treat’ 100 patients a day. Pay them next to nothing. When some patient suffers, allow the legal system to crucify the imported ‘doctor’. Deport that one and get the next one.
Yes really, human doctors start a medical degree at what, 17 or 18? Then it's a 4-6 year medical school followed by 3-7 years residency depending on your country and education system.
(This is also why LLMs passing medical exams, though impressive, has not rendered the profession obsolete: LLMs are book smart, but don't have the implicit knowledge that we humans only gain from practical experience).
Actually, no! Producing ATP is ~30% efficient, which is much better than premium consumer solar panels. And while that heat is technically waste, even if it was more efficient we would still generate heat other ways just to keep warm.
What's even more shocking is that conversion of food to energy period is ~90% efficient, which is crazy to me. The fact that you can burn food and measure the energy given off, and that's very close to how much energy you get from eating it- that's insane.
The efficiency of the human body is all over the place. Muscles are only ~30% efficient, and the rest is waste heat... but humans walk using orders of magnitude less power than any walking robot. As far as I know we have never made a powered walking machine that is 10% as efficient as a person. The only way we can beat it is with a carefully balanced, specially-lubricated pair of legs that is leaned downhill on a treadmill and powered by gravity.
To be fair, our food to energy conversion is so efficient because the foods we eat are already in a very energy ready state. Our bodies don't bother with stuff that is harder to convert.
true, but recycling proteins is also pretty amazing. These are the most complex machines we know of, elegant atomic factories that do the seemingly impossible... and you dip them in acid and then you can pop them apart like a string of beads, to be reassembled into a totally new molecular miracle.
I wouldn't. We have been trying[1]; humans are genuinely shocking in this way. Mechanical systems should have so many advantages over humans; springs are 10x better than tendons, motors are 3x better than muscles. It's possible humans evolved walking as a predatory tactic. Even if we develop a super-efficient walking robot, humans are efficient at several speeds, and keep that efficiency while varying their stride and foot placement, and with one of the largest most complex brains of any animal. And the ratio of leg/torso length is pretty variable among humans! Not to mention the flexible spine and swinging hips should be a huge energy sap, but they just kind of... aren't.
Beating human locomotion in the general case is pretty far off. It's a combination of body plan, extreme optimization of joints and energy storage, and really good algorithms.
One killer feature of the human body is synovial fluid. It's very thin, non-newtonian, self-replenishing and contained in particularly low-friction bearing surfaces. It's certainly better than 99.9% of mechanical joints, because these surfaces filter, heal and re-lubricate themselves. Mechanical joints have sticky grease so they stay lubricated without maintenance, and work in the presence of water and grit. It's doubtful that any joint that doesn't heal itself can compete, long-term.
Absolutely correct, evolution will never ever beat a 10 cent ABEC-7 bearing. Skateboards really changed manufacturing by making a specific size of wildly precise bearing incredibly cheap. One day a robot will be able to step directly onto ice skates, but meanwhile I'm looking forward to the blooper reel because it's gonna be funny as hell
Our current society converts chemical energy to electrical, not the other way round. So its the electrical energy user that needs a lossy conversion step.
Fossil fuel plants are chemical -> heat -> mechanical -> electrical. The heat -> mechanical step is the inefficient one, sadly limited to ~40% efficiency even with very fancy machinery. Most other conversions can be above 90% efficient.
The efficiency of gas-fired combined cycle power plants can exceed 60% (lower heating value). And their capex is just over $1/W. Combustion turbines are amazing.
With a SOFC topping cycle they might approach 70-80% efficiency. SOFC with just a combustion turbine (no steam bottoming) could exceed 60%. Granted, SOFCs are direct chemical->electrical conversion, but their waste heat is very usefully hot.
I don't think it's entirely a coincidence that nuclear power plants in the US stopped being built about the same time combustion turbines (by themselves, without the steam bottoming cycle) reached efficiency parity with high temperature steam turbines.
(SOFC = solid oxide fuel cell, which operate around 1000 C.)
("Lower heating value" is based on energy that could be obtained burning natural gas to CO2 and water vapor. An additional 10% could be obtained by condensing the water vapor to liquid, this is "higher heating value".)
But I think the main problem was that they never managed to scale up the clock speeds sufficiently, even though structure size (=> density) was already highly promising from the start.
Maybe in a slightly different history with some discoveries in different orders these could have replaced flash memory in SSDs completely.
But that whole episode thought me that betting on early technology is hard, and always a risky business, because no matter how promising an approach looks, if it turns out that you can not find the necessary improvements in only a single dimension, then the whole thing is kinda doomed and will probably never be competitive (=> a highly relevant insight especially when speculating about things like novel battery chemistries or the like).
reply