SNIA Experts on Data

Breaking the Memory Wall with MRAM

• SNIA • Episode 30

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 35:25

The compute curve is exploding, but memory is not, and that gap is now one of the biggest constraints in AI infrastructure. We unpack the “memory wall” and why it shows up so clearly in AI inference, where time to first token, token throughput, and unpredictable demand can make yesterday’s architectures feel suddenly brittle.
 
 We’re joined by Dr. J Metz, Chair of the SNIA Board of Directors, and Jack Guedj, one of three co-chairs of the new SNIA Compute, Memory and Storage Community MRAM Alliance Special Interest Group to talk about MRAM (magnetoresistive random access memory), and why the MRAM Alliance joined forces with SNIA. The goal is straightforward: bring persistent memory conversations into the same room as storage standards, system design realities, and the messy trade-offs that appear at scale. When a “medium” cluster can mean 100,000 GPUs, you cannot treat memory, storage, networking, protection, and security as separate puzzles.
 
We dig into what makes MRAM interesting for modern systems: very low latency reads, strong performance potential for inference, persistence without power, and the possibility of reducing power draw by eliminating refresh overhead. Hear how to connect the dots between the memory wall, AI inference performance, and discover why persistent memory is becoming a system-level priority.
 

SNIA is an industry organization that develops global standards and delivers vendor-neutral education on technologies related to data.  In these interviews, SNIA experts on data cover a wide range of topics on both established and emerging technologies.

About SNIA:

Fresh News And Quick Intros

SPEAKER_01

All right. Welcome to the podcast, everybody. I'm super happy because we've got fresh news. May not be totally fresh. You may have seen some of it, but we're going to dive into why this is important news. Uh, so my name is Eric Wright. I'm the host of the SNEA Experts on Data podcast here. Uh, and I'm excited to have two fantastic humans with me to talk about the MRAM Alliance, how this relates to SNEA, what's happened, what's happening. Uh, and I guess to kick things off, uh, we'll just do a quick intro just to say who you are, then we'll jump into the subject matter. We'll start with you, Jay.

SPEAKER_00

Sure. Uh so my name is Jay Metz. I am the chair of the board of directors for SNEA. And uh I've been doing this for uh for quite a while and basically all things storage.

SPEAKER_01

Nice. And Jack, let's uh for folks that are brand new to you, which hopefully they aren't because uh uh you got some fantastic stuff you've already shared in the world. And uh so let's let's introduce you for folks that are new. Sure.

SPEAKER_02

Uh my name is Jack Gage. I'm uh chairman of the board of UMEM, uh an MRAM-based uh chip company. And I'm also managing director of HITEN, which is an advisory and investment uh firm.

SPEAKER_01

Now, the interesting thing about all of the work that's gone on in standards, we've got incredible innovation, we've got the industry is struggling with what is the most starved resource today. And certainly uh memory is uh top of that list, aside from dollars to purchase more memory. But we quite literally have a lot of real problems in delivering on the supply that's needed to do stuff. And also, as that's occurring, there's an incredible amount of innovation happening about how do we better optimize what we have in place, what other innovations do we make that can maybe defer the need or remove the need to go down sort of traditional memory growth. Many, many things are impacting a lot of stuff in in the industry. But what's great about seeing organizations that are standards-based, that have really, really good broad customers and clients who are contributing to those communities is now we're bringing more and

Why MRAM Fits SNIA’s Mission

SPEAKER_01

more people with a common goal of making systems better so that we can make people better. And uh, Jay, if you want to introduce what is it that we want to talk about today, how what is the MRAM alliance in relation to SNEA now, and why is this important to SNEA and its community members?

SPEAKER_00

Oh, absolutely. So um one of the things that we're trying to do inside of SNEA is uh look for an end-to-end holistic way of approaching problems that we're dealing with in the data centers now. So obviously we're a storage uh ecosystem uh standards organization. And so we've developed an awful lot of uh standards that people have they they use inside of their um, you know, inside of the data centers, they use inside of their laptops, we use inside of uh in tribes, uh different technologies that deal with uh data movement, data protection, and so on. So um it made perfect sense that once the you know the MRAM group was uh does its development inside of JEDIC, which is another standards organization. Um but because of the fact that it's a persistent memory solution, it made perfect sense to fit in with some of the other things that we're doing inside of SNEA. One of the the initiative that we're creating is something called uh uh storage AI, which is the approach for how do we handle all of the storage responsibilities that go across the board for AI-based workloads. That's not the only thing we do, but that is one of the things that is incredibly important to a lot of our members as well as the general conversation going on across the planet. And so new memory technologies, especially persistent memory technologies, fit in perfectly. So we're we're looking to try to take stuff that is going on inside of MRAM and extrapolate that into other areas that it's related to.

SPEAKER_01

And I guess that's probably why it's important for us to learn what it is that made the MRAM Alliance so interesting and such a fantastic fit. So, Jack, I'd like to actually start from the the first principles. What is the problem that that the org is you know here to solve and and how does it relate in in the work that you've been doing?

The Memory Wall Meets AI Inference

SPEAKER_02

Yeah. Um well, first of all, the uh memory wall, as it's now uh more widely known, has been in existence for a while. You know, um five plus years, if not more. Since 2012, the computer's been growing at a rate of a thousand X since 2012. Memory has only grown 30x. So there is a huge deficiency between the rate of innovation in the memory and the rate of innovation in the computer. So more than ever now, companies need to consider any new types of memories and architecture as well as interfacing in order to tear down that memory wall. Um and there MRAM is a is ideally suited for uh AI inferencing. Um I can tell you later on, I've got eight reasons why MRAM is ideally suited for uh AI inferencing. Uh so you know, all of this made that we felt we needed a broader audience to get beyond the existing uh uh memory companies that have been doing pretty much the same thing and improving it uh year over year, but with the same technology, and then be able to open to the world a new technology, again, that's ideally suited for AI inferencing.

SPEAKER_01

Well, it's an interesting thing because it's a in effect that memory wall creates a Jevons paradox that we have seen a lot of stuff that's moved into changes and innovations around KV caching and how do we better do sort of like storage tier offload where we can still keep things intact but just move it down to lower tiers. But then again, even in the memory space, how do we better, you know, build memory and target memory capabilities based on these new workloads? And inference fundamentally shifted everything because you know, I I I don't want to be like every LLM out there. The the paradigm has shifted, yeah. Like, but we have actually seen that inference is everywhere and it's very unpredictable with how it's being consumed, and it will be for a long time. So this is why you know you must be excited because you were already the world arrived to where you are, Jack. Like, like, like, you know, Sans and Jay, you've probably seen this as well with a lot of things. Like, we've been developing these ideas, and all of a sudden people are like, hey, you know, how do we actually look to what's out there today and and put it into the place in the most optimal way? So what do you see as like now the sync?

SPEAKER_02

You know, we we started this group uh three years prior to joining SNEA um as a small group and growing a little bit, but we felt we needed a bigger platform, which we are you know extremely happy since uh you know we joined SNEA in January uh was the broading. Uh and that's what we're looking for. We're looking for ways to reach out to system companies, we're looking for ways to get more collaboration, because at the end of the day, more collaboration, the market is huge, uh, and more collaboration will yield uh a better solution, better system solution uh for those AI uh you know memory issues.

SPEAKER_01

You had me at collaboration, Jay. You've you and I have been at this for a long time as well, through various communities. So what do you

Seeing The Whole System Puzzle

SPEAKER_01

what do you see as like kind of the immediate win with creating this and creating the SIG? And also across the other parts of the SNEA community, where do you see the things plugging in that will both help what Jack and the SIG are doing as well as bring that this whole ecosystem together?

SPEAKER_00

So I think that if you're looking at the life cycle of a workload, and you know, from a very practical perspective, especially for those of us who are geeks and we love to figure out how things work, um you hear people talk about things like time to first token, you know, uh you hear people talk about uh direct access into storage and accelerated this and accelerated that. But what winds up happening is that you you run the risk of focusing on a very small part of a larger puzzle with a lot of moving parts. I believe Jack um kind of uh mentioned it, but I want to reinforce the fact that a lot of these things are are not done in isolation. You know, we we have we have the data movement that goes from one place to the other. It's not it's not just that you have an accelerator or a processor, it's not just that you have memory, it's not just that you have KV cache and and whatever buzzword happens to be popular in the next, you know, the next month or so. It has to do with how all these things relate together, right? And we we see ourselves with a number of different alternatives that are coming out because there are limitations at these kinds of scale. And I don't think that people necessarily understand exactly what kind of scale we're talking about. You know, um in in certain AI workloads, for instance, it is not uncommon to talk about small, medium, and large size clusters, but people have no clue what that actually means, right? So for in in my day-to-day job, for instance, where I am, you know, we consider a medium-sized cluster to be 100,000 GPUs. And that's just the GPUs. That's not counting everything that goes along with it, you know, the storage, the networking, the real estate space you have to put in in place for it, right? Um and these kinds of workloads are very, very, very demanding, but they're also extremely sensitive, right? So you have to put in a lot of protection, you have to put in a lot of of security in the right places, and there's so many different things to make the whole thing work. It's it's sort of like saying, well, I'm gonna remove a single lug nut from my car and expect it to work right. And that's just not a very wise thing to do. Well, it's the same thing inside of these data centers, inside of its workload. And AI is only one workload, right? So what we're trying to do is say, look, as we start to figure out, you know, how do we get a better, you know, a better piece of the puzzle, where does it fit into the overall, you know, uh the overall ecosystem, overall, you know, technology architecture, you know, there are consequences, there are trade-offs that have to be made. And what we do at SNEA is we look at those consequences and we work with other organizations like JEDEC and a few others to make sure that this actually is a smooth transition from you know a more traditional data center type of environment to a very workload-specific environment. So it's very important to us that we figure out um, you know, in advance so that people don't learn the hard way, you know, what the trade-offs are going to be and what they need to be.

SPEAKER_01

Yeah, it was funny. I was a in a DR discussion with somebody about uh a steward solution, and the question came up that I I hear that there's an issue with like kind of a full like rack row level failure, that that could cause a problem. I said if you're uh an NCP or if you're uh a neo cloud and you lose a a whole row and you have no other backup for it, yes, that's a problem. Like we the localized things that we're thinking about don't come into play the same way when you, as you said, like kind of scale it out. Where we're not thinking about rack level and server U level. We're talking about row scale and data center scale stuff. So, you know, Jack, as you've been developing and and you know, looking at where these technologies apply and what are the use cases that match. What is the what's what's in the MRAM Alliance portfolio of hardware tips and tricks, you know, that has created now what will be a really strong opportunity to build around stronger and faster yet at scale type of memory implementations.

SPEAKER_02

Yeah, there um as Jay mentioned, there um you know, there's performance issues or performance metrics uh like time to force token and then uh you know token throughput. There's also uh power related uh constraints. Right now, those racks are just uh you know gobbling power, and you need uh uh close to a nuclear power plant to power the new data center. So it's it's it's going crazy. Um and you know, one of the two of the uh eight reasons MRAM is ideally suited for AI inferencing is one, it's fast, it's really fast, and low latency. Uh it is much faster at a bit cell level than a DRAM. And uh the latency because it's a direct read, just like an SRAM, uh, is really low. So you can get latency on the first token less than 10 nanoseconds, uh, which is blazing fast. And you can run that MRAM you know with with an engine, you can run it basically maxing out the process frequency. And it doesn't matter which process it is. So uh so you can get really fast throughput and uh and really low latency. And also this memory is able to retain data uh without power. So you can uh and you don't have any refresh, you don't have any of this overhead, so you can achieve uh a very low power with that. So uh as Jim

Speed Latency And Power With MRAM

SPEAKER_02

mentioned, uh we're totally aligned with with the SNEA uh even prior to joining SNEA, which is we want system companies to learn more so they can do a better job at fault and not learn by their mistakes. We want we want to reduce the number of mistakes they're making and speed up their time to adoption and time to market. Um and right right now we're embarking to um uh uh it's it's a new initiative we have, which is on the system design side to be able to, by application, explain and give guidelines to companies on how they can incorporate MRAM and uh you know what things they should be careful in doing their system design. And you know, we'll expand that to webinars, we'll expand that to FAQs so uh we can get uh the word out and again, as Jay's saying, make people's life easier.

SPEAKER_01

Yeah, it's great that you know I'd say the education is one of the most powerful things that we even if you go looking for individually, I may go dig and find some things and I'll individually learn some things, but inevitably when I get to SDC and you have extended conversations with other people who've also done their own research, then now you're actually working in that collaboration and there's a persistence in the collaboration versus me like I'm just gonna build one thing for my specific purpose and I'm in and out. But being able to move this in and have that SIG approach means you get you know the right backing as far as governance and and everything to keep the the lights on for the org, but also just now you get access to all these fantastic technologists who are as excited as heck about this stuff as us, which you know maybe that tells you something about us, but I'd say I'm pretty damned excited about what the heck is in front of us. Like especially when you get to low power, you know, we're seeing way more things around inference at the edge. And we actually know what the edge is now. It's not not what it was eight, 10, 12 years ago when somebody invented it. And thank God we didn't land on fog. But here we are. Like we are legitimately seeing small footprint, small power footprint, inference capabilities that are gonna need this type of memory access and persistence because you solve one problem, you create 10 others.

SPEAKER_02

Uh and so it sounds to me like bringing the eye close to the edge is also a power reduction because now you don't have to transmit as much. And we know the the wired and wireless transmission, especially wireless transmission, uh tend to burn a lot of power.

SPEAKER_01

So, Jay, because this is obviously certain we'll say we're AI focused in the discussion, but do you find is that like kind of the biggest conversation right now inside SNEA and with some of the like your your own like other community partners? Yeah, yeah. Like we obviously we're we're a bit hyper-focused on inference in in a few of the discussions we have, but I don't think that we're alone. I think this is like a very strong collective problem we're challenging, we're trying to solve together.

SPEAKER_00

Oh, I I I think you're right. I mean there's it it would be disingenuous to say that AI didn't bring up a lot of the conversations. Um however, I do find it is somewhat fractal. Right. So it's just it's just one stage amongst many in that in that level of discussion. Um you know, ultimately what we're trying to figure out is is how to make data movement more efficient overall, period, regardless of whether it's for AI or something else. Um and and the everything that goes around it, you know, everything that sort uh uh surrounds that question is fair game for conversation, right? So so when you think about storage, and I and I believe that one of the big problems that we have is that a lot of people who are not storage oriented don't really understand what storage means, right? And storage has one job. Give me back the correct bit I asked you to hold on to for me when I ask for it, and everything that goes around that is all part of what we have to do. Now, different people are gonna focus on different things, right? Um, you know, if you if you're focusing on the capacity of a drive, you know, um the bit that I ask you to hold on to is what you're gonna remember. If you're going to be an application person, you want to be the correct bit, right? If you're an end user, it's when I ask for it. But if you're happening to be, you know, a lot of the members inside of SIA,

Storage Means Giving Data Back

SPEAKER_00

not only do they have to deal with all that, but it's give the back, right? Give me back that bit. And that word give is one of the things that is the most difficult to do correctly and efficiently. Because every time you have to give, you you have to find the trade-off about accuracy and resiliency and security and reliability. And the best bit that you have to send is the one you never sent, right? So if I don't have to send an IO, then I am that's great. But if what I'm doing is I'm just basically kicking the responsibility to someone else, not so much. So it's important to note that we all have to take a look at all these different uh these these concentric circles of influence with the the technical architectures. And every single stage along the way, you've got the physical aspect of it, you've got the networking aspect of it, you got the memory, and so on. You know, it's like somebody standing on Times Square, yo, I got your MRM over here, and I got your VDR over here, and I got your gold watches right here. So, so what we need to do is we need to say, all right, you know, how are all these things supposed to be working in concert together? And that means that it used to be that SNEA did some of this, and NVMexpress did some of this, and OCP did some of this, and Jeddek did some of this, and DMTF did some of this. It's not it's now the problems are so big and so involved that SNEA works on a large portion of it, JEDEC works on a large portion of it, OCP. He works in a large portion of it. So even something as focused as some as MRAM has implications beyond SMIA, beyond Jetic, in other areas. And you need to be able to have the conversation at all of the different levels if you're going to have any credibility. Right. Right. And so, you know, Jack's talking about wanting to get people involved in doing this. He's like, yes, because of the fact that the the tail of influence is extremely long now.

unknown

Yeah.

SPEAKER_00

Right. And that's it's something that that if you're not careful, you could find yourself on the business end of a customer hissy fit. And I don't think anybody really wants that. This is true.

SPEAKER_01

Yeah. So Jack, uh with that, you know, what what is what are some of the other use cases that are maybe not necessarily AI centric, but like what are the more you know enterprise use cases and other use cases that you've already seen come out or that you are seeing as opportunities for for MRAM as a spec?

SPEAKER_02

Yeah. First of all, Jay is is is right on. So um, you know, we're we're totally in line. Can I have that in writing? I want to I want to I want to pass that on to someone. Yeah. And you know, on the on the give difficulty to accurately and efficiently uh pass on the data, uh that is reason number three, MRAM is ideally suited for AI interesting. It is one of the newer memory that is the most reliable. Um so reliability is extremely good, and it's also uh uh very high endurance. And so, you know, going to the uh question you asked about applications, it does have a wide range of applications because of that, uh, including automotive right right now. You see uh companies, uh the big um automotive chip companies uh making AI-based uh sorry, automotive MROM-based chips, not only for AI, but for a wide range of applications, including microcontrollers, um staying on the on the reliability uh and also the fact that inherently, because MRAM is not a transistor-based technology, MRAM is uh sits between two metal

MRAM Beyond AI And Into Space

SPEAKER_02

layers on top of the active layer. So um because of that, it is very resilient to uh uh radiation. So when people are talking about uh data centers in space, uh it is a lot easier when you have a technology that is uh resilient to uh radiation and soft errors. Uh but in terms of application, it's pretty wide. I mean, it goes from you know smart cameras, and there are some right now being designed uh with uh with MRAM uh to um applications like ARVR, uh ultra low power application, because you can shut down the system, you know, say you have a smartwatch and you're sleeping, you don't need that memory necessarily to be uh to be live. It's the your watch is not doing very much, but right now uh it's hard to shut it down because you don't have a memory that is fast and that retains the data. So with this, you know, you can really play and be able to have different levels of um you know the uh of uh of sleep modes where you can gradually and gracefully uh lower the power. But that scales all the way to data center because since this memory is uh is bling fast for inferencing, um you can scale that all the way to data centers.

SPEAKER_01

Well, and we're seeing more of you know just the right patterns as we learn. Yeah, we we couldn't necessarily have predicted how the right patterns and the storage patterns will go for these workloads. And even as we go. So definitely durability, the endurance of the actual hardware, the the fact that you can subject it to pretty harsh conditions, you know, that's again, these are things that they may have seen like unnecessary for most of the use cases we have two years ago, even you know, probably even even six months ago, it would have been like, oh, it's interesting. But when you dig in and you look, just in in a few minutes, you've already hit yeah, this is a very, very important technology. And uh so yeah, super exciting. So, what do you see now? What what's the best thing for folks that do want to get involved? Uh Jack, I'll start with you. What's the the way that you want to engage via the SIG? And then uh Jay, I'll have you kind of close up and talk about how to connect.

SPEAKER_02

Yeah, well, uh first of all, SNEA is the platform, and uh you can find us through the the SNEA platform. Uh we have a booth, uh SNEA booth at the uh in the exhibit hall of uh FMS, which is going to happen uh early August, uh future of memory and storage. And we will have demos at the SNEA booth. Uh so you know if you want to learn more about what we're doing, I will give a talk uh at FMS. We will have uh chat with the experts. So that's a good forum if you have time uh during the day or Wednesday evening to learn more. Uh and uh we will also be present with multiple talks at the uh SDC. Uh and then later on this year we'll have the Global Forum. So many ways to find out more about the MRAM Alliance and uh and how you can join and how you can get involved either on the uh overall alliance SIG or one of the SLEP SIG we have, uh subgroups. Uh as I talked about earlier, the system design with MRAM, memory interfaces, that's another point. Uh basically in this, you know, give back uh accurately and efficiently uh transfer the data back, uh, so memory interfaces so that we can play along because it's a multi-memory uh type system in order to optimize the architecture. Uh and then we also have a roadmap track uh which kind of goes over what's coming in the future.

SPEAKER_01

Fantastic. Well, looking forward to FMS. I'll be there. Uh strangely enough, they allowed me to speak at it. So I don't know how who they might have a low bar there, but uh so it's gonna be exciting. And one can wait.

SPEAKER_00

I'm uh I'm in that same boat.

SPEAKER_01

So we can all enjoy. We will be there to uh put you through the test, Jack. We're gonna put you through the technical exam. Uh but Jay, what what what's ahead

How To Join The SIG

SPEAKER_01

and based on this and what what do you recommend for folks when you know how do they engage both with SNA and with Jack and the folks through this SIG?

SPEAKER_00

Yeah, well, Jack um uh was absolutely spot on about the upcoming events. We've got a number of things that are going on. I mean, as we're doing this recording, um we've got the the FMS presentations that are coming up, we've got the SNEA Developer Conference in September, the AI infrastructure summit, the supercompute. SNEA is going to be at all of these things. But beyond that, um there's an you know, the ongoing information that we have at SNEA.org is incredibly important for people to be able to keep the updates and see the webinars that Jack was talking about and the uh in the integration that's going on, and to participate in the storage AI community is absolutely incredibly important. Um the the key thing I think that is is really missing from a lot of the discussion, and I believe SNEA is stepping up for this, is that a lot of times people are just sort of saying, Well, this is what's going on, these main big companies are kind of pushing this on us. I don't really have a voice. This is the opportunity to have a voice. This is the chance for people to say, you know what, I want to have some thoughts in in in the open discussion, and that's exactly what SNEA offers. You have the opportunity to share your voice, share your ideas, and even influence the way that the direction works for the entire ecosystem. And so by joining, you know, SNEA.org or SNEA.ai, if you want to be specifically for the AI stuff, both of those are great options, even beyond the events that are coming up in the next few months.

SPEAKER_01

It's gonna be, you know, I think we're gonna see more and more of these smaller events and and smaller get-togethers because it's it's eventorama right now with all these AI use cases coming up. So I'm glad to see fantastic minds coming together. Looking forward to seeing both you both at FMS and at SDC. And for folks, of course, you can check out the links below. We've got links to, and if you just do a search for MRAM Alliance, the beauty part, it'll take you right to the main page. It'll pull you into the SNEA org. Get on in. If you're not already a member of SNEA, then uh, well, you're a wrong person. Go do that thing. Go get in, go get in charge in SNEA. There's a fantastic group of humans. Uh and rare.

SPEAKER_02

I have to say it is um also affordable as SNEA as if we're great. So, of course, smaller code makes it very affordable to join SNEA.

unknown

Yeah.

SPEAKER_01

Yeah. And then it's very startup friendly. It's super important because nowadays you know, the last thing you want is to join, you know. I mean, there are obviously sort of very vendor-aligned things that end up having to happen to a lot of these hardware communities, and then next thing you know, they're burning their entirety of their budget just to be able to participate with an upstream you know player. And that's tough, you know, where this is much more broad ecosystem coverage and international too. So that's the other thing is you get these are there's regional events, there's community members from all over, and it doesn't end when SDC closes the the gate at the hotel, it continues online from that moment onwards. So there's lots of ways to keep engaged. So I'm I'm excited. All right. Well, so let's just do a last little quick recheck. So, Jack, where do folks find you if they want to get connected with you? And uh you can feel free to say at SNA.org is a great place to find you because I imagine there's a few places to go.

SPEAKER_02

Yeah, I would say at SNA.org uh or SNA.ai. That's probably the best place to find us. Or you know, you can Google us or you can use AI to find us.

SPEAKER_01

That's it. It's always neat to be consumed by the very thing that we create. Uh Jay, for for folks that want to uh get connected with you and uh how do they do it?

SPEAKER_00

I think the probably the best way to do that is to um uh as as Jack said, you know, um I can be reached through SNEA.org, but also on LinkedIn and the usual social media challenge at channel chat channels, excuse me. Um, you know, YouTube. I've got um I uh I'm starting to get back into doing some more blogging on my my own site, jmets.com. So you'll see some other things in a variety of different places, but um I'm sure you'll probably hear from me somehow.

SPEAKER_01

And also uh for folks that didn't get a chance to see it, Jay did a really great presentation on sort of the some of the stuff that SNE is handling at uh Tech Field Day, which is uh uh a sort of a vendor, a non-vendor vendor-ish community of practitioners, but it was a great presentation. It's available online. So if you just do a search for Jay Met's uh Tech Field Day, you'll find a lot of content. Look for the most recent one. Uh so that was a great presentation, really kind of diving into the problem. And that that reminded me that as smart as all of those amazing humans are that we're surrounded by, we forget that things have changed. And we have this sort of like assumption that everybody just is caught up on how AI consumes storage. And it was wild to see some of the questions that I wouldn't have expected from really, really super smart people that have been in the industry a long time going, oh, right, we we need to keep educating and we need to keep learning. So uh I'm looking forward to learning from from both of you. So, Jack, Jay, thank you very much. And uh for folks, of course, again, follow the links, get involved with SNEA, and uh we'll see you all on the next Experts on Data podcast and live at FMS and at SDC.

SPEAKER_02

Thank you. Bye.