The Future of Data Storage: Dr. Stuart Parkin on AI, Spintronics & Racetrack Memory
Autria Godfrey and Laila Rizvi discuss how data storage can keep up with the demands of AI, with the physicist Stuart Parkin. Parkin, whose spin valve is used in hard drives and who won the 2014 Millennium Technology Prize, explains how he developed it at IBM and how the field of spintronics grew from it. He describes racetrack memory, a solid state storage technology he proposed 20 years ago, and says it is close to being ready but needs investment.
Transcript
Welcome to Agents of Tech, the podcast that delves into the cutting-edge innovations and pivotal ideas shaping our world. I'm Autria Godfrey coming to you from Washington, D.C. And today we are tackling a crucial challenge. How do we keep up with AI's growing demand for data storage? Joining us today is visionary Dr. Stuart Parkin, whose groundbreaking work is transforming the future of data technology. Dr. Stuart Parkin's work has had enormous impact in revolutionizing data storage technologies with his invention in magnetic disk drives and the development of spintronic devices over more than 20 years ago.
Yes, you know, I think that's actually I'm always telling my PhD students here, we want to do the impossible. And so oh, they think oh, we we know that's impossible. But now I think the impossible is actually what we should strive for.
Remember, if you are enjoying our show, please don't forget to like and subscribe to Agents of Tech as this really helps develop the show and shine a light on cutting-edge research. We thank you.
Joining us now is Laila Rizvi, the director of strategy and research here at WebsEdge, the company behind this podcast. Laila, it's such a pleasure to have you with us today.
Thank you for having me, Autria.
It's such an important topic that we're discussing today, data storage. I mean, this continues to be a critical issue both for business and government. Uh we need solutions and innovations.
Yeah, definitely. And I think the thing about data storage is is it's probably one of the biggest bottlenecks which isn't really being discussed in the mainstream. I mean, we talk about hardware, we talk about the size of large language models and the need for more compute, but we're not necessarily talking about the fact that we need to store more data as these models are doubling in size, you know, every 6 to 12 months. And they are reliant on really large data sets to be more and more successful. Where is that data being stored? And you mentioned Stuart Parkin and magnetic disk drives and that development 20 years ago. We haven't really moved on from that technology and clearly student is Stuart Parkin is leading the way in every kind of decade when it comes to innovation in data storage. So it's going to be really exciting to talk to him, but we're still using his technology which was commercialized when he was at IBM 20 20 years ago or so. Really really critical in that respect.
So let me try to just put this into, you know, perspective here. According to a recent report by data storage company Hitachi Vantara, they are predicting that the average large organization will be doubling the amount of data that they expect to be storing just in the next 2 years. So we're not talking about something in the far distant future. We're talking about the need for being able to store these massive amounts of data. It's coming up, you know, by the year 2026.
Yeah, exactly Autria. And the other thing is which I think some people fail to mention. Obviously data and data storage is critical to training these large AI models, but a fundamental component or characteristic of AI is its ability to share knowledge with limited loss. So you and I when we're sharing knowledge, we share it through words. So you have to interpret the sentences that come out of my mouth. Whereas AI is a digital system that can make copies of itself. That copy is in the form of data. So when it comes to AI progressing and evolving, it needs to be able to have the capacity to store data and thus share knowledge with other AIs or with humans. So that's another level beyond kind of where we are now, which I don't think people are really anticipating, which is the ability for AI to share knowledge which is contingent on data storage, but also the movement of data.
Right. So you know, we mentioned that Dr. Stuart Parkin is the father of spintronics. Can you give us a a couple sentence definition of what that is?
Spintronics is basically the field in which they use the spin of an electron as the basis of the hardware or the computing. So, rather than electric current, you have the spin of the electron which encodes for a one or a zero. And what Dr. Parkin has kind of done over the past couple of decades is develop racetrack memory which is based on the kind of philosophy of spintronics or the concept of spintronics. And racetrack memory, which I imagine he'll go into more detail about, is um is a data storage mechanism which moves the data rather than the hardware. As a result, it's much more efficient, it's much more durable, and it kind of occurs at the time scale of nanoseconds.
And Laila, we have to, you know, of course, address one of the world's largest prizes in the world of technology, the Millennium Technology Prize that Dr. Stuart Parkin won in 2014.
Massive achievement, and it's unsurprising because he's pretty much laid the groundwork for for all of our hardware for computing so far. I mean, with his innovation when it came to magnetic disk drives, we would not be where we are now without that. So, I'm looking forward to see what the next 10 years of Dr. Parkin's kind of innovations lead us to and how they're commercialized.
Right. Well, he is a true visionary, and we are absolutely honored to have him with us now. Let's take a listen. Here's Dr. Stuart Parkin.
I want to get started with, you know, from the University of Cambridge to IBM, now to the Max Planck Institute in Germany. Your career has spanned continents and innovations. Were there any moments early on in your career that made you really stop and think, "Wow, this is it. I'm on to something big."?
Sure. I think in several times in in career, I felt, "Oh, that was a that's a great idea and I should pursue it. So, indeed, yes, I think sometimes it just happens or you feel oh, that's a major that could happen that could have major impact scientifically or technologically.
And then Dr. Parkin, I want to talk more about the spin valve which was kind of the first thing that kicked off everything in spintronics. And it's in every hard drive today. How did How did you come up with that idea and how has it laid the groundwork for other spintronics based technologies?
The spin valve evolved as you said from a very early work. I I used to work on organic superconductivity when I first worked at IBM. They wanted me to work on magnetism. So, I came up with this concept of magnetic multilayers. Very exciting cuz you could engineer new properties. And as a result of that, I started working on on magnetic sandwiches. And then giant magnetoresistance was discovered and then I realized that oh, one could exhibit that phenomena without the need for coupling between the magnetic layers allowing the possibility for sensor devices that could operate at very very low magnetic fields.
And since I worked at IBM Research and IBM Research at that time had a very large company, 10,000 people making magnetic disk drives just a few miles from the research lab. Then I knew that uh the sensing aspect of magnetic disk drives was a major impediment to the future evolution of these disk drives. In fact, people believed that you could not achieve a certain areal density, the number of magnetic bits per unit area in a magnetic disk drive with the existing technology and there were many papers written that this is this is it. This is the final possible capacity of magnetic disk drives. But I knew because of that that a new sensing technology could overcome that barrier. And in fact, the spin valve enabled a thousandfold increase in the storage capacity of magnetic disk drive beyond what many people thought was possible at that time.
Is there anything today that you think people think, you know, we've reached the absolute limit, we can't increase it anymore. Is there any particular technology or area in which you think that's the case and you disagree with the general consensus.
Yes, you know, I think that's actually I'm always telling my PhD students here, we want to do the impossible. And oh, they think oh, we we know that's impossible. But now I think the impossible is actually what we should strive for. What I think is basically beyond the current capabilities of our knowledge, our research. I think that's very interesting.
And in particular, I think the field of atomically engineered materials has so much opportunity to discover new phenomena, new ideas, new concepts for computing, for data-centric technologies in general. But there's huge opportunity. For example, uh I think people, you know, fundamentally believe that we consume too much energy per computing operation. And we could reduce that by maybe 100,000 times to reach the fundamental limit. So, of course in principle, it's possible, but very difficult to imagine how that could be possible whilst doing that with very very low energy. So, of course, uh the brain consumes tiny amounts of power, 20 W versus it's believed perhaps 10 MW to compete with today's technology. So, there's orders of magnitude improvement in the energy consumption of today's technologies that could allow for much more energy-efficient computing in the future with new ideas.
So, I want to come now more onto kind of commercialization of concepts and ideas from discovery to implementation. Um and with the spin valve, what were kind of the challenges? I know you you went from using molecular beam um epitaxy to a more kind of scalable and practical method which allowed for more practical layering of uh materials. What have you learned from that process of taking something from a con- concept to commercialization? And how has that kind of maybe changed from when you originally did it with the spin valve to now?
Oh, that's a very interesting question. I think often it takes a very long time to take a concept, you know, the possibility of achieving something to reality in the sense of mass production. The reason is that in the laboratory you typically make one or two devices and maybe a small number actually work because you don't have the capability of making all the devices to be identical. Whereas today in today's technologies, whether it's storage, memory, or computing, you need to make billions or tens of billions of devices, each one basically identical at tiny length scales. Well, this that's a tremendous technology engineering challenge that takes a lot of time, effort, and money and typically the workings of many companies around the world. The whole enterprise has to work together, which in some ways was why CMOS was so effective. The idea was you would scale every 2 or 3 years by a factor of two in a critical dimension. And so the whole industry worked towards that in concert. Of course, that meant that many other concepts and ideas were not explored. All resources, all the talent went into that one possibility. Well, maybe, you know, two decades ago we could have said, "Oh, maybe there's an alternative where you could do the same same computing with much less energy, but it would require massive investments."
Because you've been able to do the impossible, we now see how your work is enabling other newer technologies like big data, like Facebook and Google, which at one time seemed impossible. So, when you see your work being utilized to help those types of companies, what goes through your mind and do you ever take a step back and think about the scale of it all?
Every day I think, "Oh, it's amazing the technology." I look at my iPhone, I think, "Oh, that's fantastic." The display that, you know, 10 years ago it's just totally inconceivable. And this is actually we can imagine the evolution of nearly all of these technologies with the amount of investment in technology engineering and so many bright, intelligent people working on these technologies. Of course, they will evolve. And typically in computing systems, computing power has increased, let's say, a thousandfold every decade. And it's very difficult to imagine what that would then entail, what would then be possible. And some Some people can see, and then they're the ones who make, you know, maybe make more of the breakthroughs than others. And so, I think that's extraordinary how technology continues to evolve, and I think largely speaking does tremendous good, bring tremendous value to the world.
Talking about technology evolution, obviously data storage and processing speeds and data access is kind of the the lifeblood of all modern technology today. Um and especially AI where it's pushing its limits. And although, you know, magnetic disk drives and the existing technology we have is incredibly important, the ability to increase storage capacity is is now reaching a limit, I guess. So, what challenges and opportunities do you see in meeting um technologies like AI's massive data demands with innovations maybe in spintronics and memory storage?
Approximately 70% of all digital data today is magnetic disk drives. However, that technology, it's 50-60 years old. And 20 years ago I proposed a new technology. I called it racetrack memory, which I believe has the potential to match the storage capacity of magnetic disk drives, but would be approximately a million times faster to access the first bit, would have a much greater volumetric density, and would consume much less energy. So, I think this would enormous impact. And you know, people will say, "Oh, you've worked on this for 20 years." But 20 years uh we had the idea, I had the idea. The reason for that idea was that IBM sold its disk drive business. And for some reason they kept me. And so, I thought, "Oh, what can I do to make a better disk drive?" And that's how I came come with the concept for racetrack memory. and the basic concept is you store a series of magnetic bits in a tiny magnetic racetrack, which is a nanowire, and then you move the information around using very novel concepts in physics that were known at the time. And in this way you can move the information to individual reading and writing devices and created the a memory storage device that would be totally solid state and would basically be 100 times have 100 times the capacity of any technology today, any solid state memory technology today. So, I think that is um remarkable if we could achieve it. And over the last 20 years, we're very close to achieving that. So, it's taken some time, but along the way we've actually because we had this I had this one focus, then we've been we've discovered new physics and new materials that make the concept maybe 10 or 100 times better than it was 20 years ago. So, I think it's we're very close to bringing this technology, I think to be a real mainstream technology in the but it might take a few more years.
I think that's so interesting also because the focus right now from, you know, everyone in terms of the hype has been we need more compute compute, more compute, more we need to improve the compute, but we can't make transistors any smaller, right? So, I think it's interesting that when there's not as much of a focus on increasing processing speeds for data access and storage when in a lot of systems like self-driving cars or electric vehicles, you need to be able to access the data quickly, right? Um so, I think maybe we need a change in focus or direction from the industry.
There is a tremendous focus on a high access times between memory and computing. This is called high bandwidth memory, which is the essence of the computing systems needed for artificial intelligence. So, there are interesting ways of improving existing technologies and making memory and logic essentially have shorter paths to transfer of the data, but at the same time, AI requires enormous amounts of data. And, you know, AI has already used up all the data in the world. We need to create more data, and then we need to be able to store it and access it in ways that are much more energy efficient. So, racetrack memory is exactly the technology that is needed for this application, and that's a completely spintronic technology. I believe just if there were enough investment, it could become a mainstream technology that would help enormously in making AI much more even more powerful than it is today, and it must use energy. Problem with AI is it's consuming vast amounts of energy. So, we need to focus on this.
I'm sure after having worked on this for 20 years, this is the question you get all the time, but probably hate to get, but how close do you think we are to seeing this racetrack memory being something that can be implemented?
I think we're very close. We just published a paper in Science a few months ago where we demonstrated racetracks that are basically a tiny tiny dimensions, and we were able to show that we could move and position domain walls at the same dimensions that would be needed for a technology a mainstream technology. So, I believe we're very close. It just requires some investment. Of course, the investment would be enormous, but nevertheless, the investment is totally possible, and it's the same on the same sort of magnitude that's required just to increment the, you know, from one CMOS technology generation to the next. So, I think one should instead of doing that, maybe invest in racetrack, and then you'd have a technology that would be enable much less energy computing, and would be more high-performing, and would consume much less space.
Is racetrack memory compatible with current systems and architectures that companies are using?
It's compatible in the sense we could use the same tools that are used today to make magnetic random access memory, which is also a technology that I've put forward at IBM, and which has become a mainstream technology from proposal 1995 to 2019, you know, TSMC, Samsung, GlobalFoundries, all of these, Intel, all of them have used this technology. But, however, so I think racetrack we could make we can make it with existing tools. However, the architecturally you probably would change the architectures.
That kind of brings me on to my next question, which is about neuromorphic compute computing and energy-efficient brain-inspired technology, which you're you're a big kind of advocate for as well. Um can you tell us more about how kind of three-dimensional architectures like neuromorphic computing and also three-dimensional memory storage um could be a game-changer for AI and computing?
You know, the brain is very interesting because the elements of neurons are connected, each one by 10,000 connections, the synapses. And the information is basically stored in those arrays, not one bit per synapse, but, you know, 20,000 bits per 20,000 synapses, that allows for uh redundancy and error correction. So, I think this is very interesting, very difficult to imagine ways of doing that using CMOS-like technologies. So, you have to imagine new ways of building structures which are innately three-dimensional with multi-fold connectivity. And this is something that uh is extremely challenging, and I'm not sure there are many solutions today. But, of course, uh it's possible. We know that. The brain, therefore, though it seems impossible, there are many ways, you know, of of doing it, I think. So, we're working. We have I have programs on 3D mostly 3D racetracks, and we've built very interesting three-dimensional devices that have unique properties.
Would we be able to force everyone to switch to what something which is a completely different world, even if it is more effective? I think it's difficult to convince everyone to, you know, scrap the stuff that you're doing on the digital side and go back to kind of analog?
The foundries, each one foundry costs 10, 15, 20 billion dollars each. So, So is a massive investment. If you have another technology that maybe you could do with each foundry it costs much less than I think you would be able to develop more rapidly and have more impact. So I I I think there is there are many ways in which maybe now is the time to imagine going beyond CMOS technologies which are really close to the limit of what is possible to alternatives, but that requires again, you know, a lot of lateral thinking, going beyond what the current thinking you know the what current possibilities, but of course it's totally possible because we know the brain has these remarkable functions, but maybe it's only possible biologically. I mean it's not possible with inorganic materials. So then maybe we need more, you know, maybe we should work more between the the biologists and the neuroscientists.
Dr. Parkin, my final question for you today. You know, you spent your career shaping the future of technology and your contributions are have just been incredible. What advice would you give to young up-and-coming scientists who dream of making the same type of groundbreaking discoveries that you have?
You know, you don't want to do incremental improvements unless that gives you the focus understanding something so well you come up with a breakthrough idea. I think one, you know, you have to you have you mustn't follow the crowd in many ways I believe. So you have to have your own independent concepts, ideas, and follow them even though it would be very difficult because maybe you're the only one with that approach. And it was true in in even in racetrack in the early days people were not convinced. Probably many people not even today. And nevertheless, we made so much progress along the way with new physics. I think that's been a very exciting journey.
So I think doing something different also I think you know, sort of interdisciplinary science at the intersection of chemistry, physics, biology, engineering is also very very interesting because every subdiscipline has their own language, their own way of thinking, and sometimes that's very confining. So I think working with other scientists from other disciplines is very interesting and may give you many insights. Also, I Some to some extent, it's a very interesting question, how much do you really need to know about something? Unless you go into all the details, maybe get lost in the in the in the weeds, and you really need a big picture in some sense to to So, you can have a concept some over overriding concept that you want to follow and imagine you could how to do it. And then, I think you can achieve something beyond what others have done, but I I agree it's not so straightforward.
Yeah, maybe one of the points I'd like to make is, going back to your new technologies, I think it's very interesting, let's say quantum technologies. That's an analog technology beyond digital, and huge interest in quantum computing. However, mhm, and that operates typically at very low temperatures. What's very interesting is to come up with alternative memories that would operate at those low temperatures. So, we have a program, I think it's a fantastic idea to use very interesting new concepts in superconducting, in particular, superconducting spintronics, intersection of those two fields.
How do you go about coming up with completely novel concepts or ideas? I know you've spoken a lot about uh the intersection between disciplines. Is that somewhere that you sometimes look?
Because I think that's a that's a very good point. Actually, quite frankly, for me, uh I like having like discussions with my students and postdocs, and then they, you know, if we discuss data they've taken, then often I think you have a gives you a new insight. You certainly you certainly need to talk to people that have ideas. I think talking with individuals somehow is more gives you more uh ideas than just reading lots of papers.
Thank you so much for your time today. We really appreciate it. This has been fascinating, so thank you.
Appreciate it. Thanks so much.
Very enlightening. Um, you know, I think that the the racetrack memory in particular, um, given how long he has worked on that and how close he thinks we are to having it be, you know, a real-world application and be implemented into is very exciting.
I have not seen racetrack memory once on like my timeline. It's all just about more chips, more compute, etc. And this is actually coming coming into play, I think from 2019. It started to be adopted by bigger players, but I just think he's on to a lot of big stuff. We've reached a limit with what is quite dated technology now. We can't make transistors any smaller and we need to look towards like completely novel avenues.
I think it'll be interesting to see the role that this racetrack memory plays in the future of data storage. That does it for today's episode. If you enjoyed our discussion with Dr. Stuart Parkin, don't forget to like, subscribe, and share as this really helps us to improve the show for you. Thanks again for joining us today. Until next time, goodbye.
Goodbye.
Agents of Tech is brought to you by WebsEdge. Our special thanks to our studio engineers Adam Deen and Cody Winn, editor Abbie Harries, and our producer Cath Sheehan.