Utilizing Tech - Season 7: AI Data Infrastructure Presented by Solidigm - 06x05: Great AI Needs Great Storage with Ace Stryker of Solidigm

Episode Date: March 18, 2024

Many people involved in artificial intelligence don't spend much time considering storage infrastructure. That's the topic of this episode of Utilizing AI, which features Ace Stryker of Solidigm discussing the role of flash storage in AI infrastructure with Frederic Van Haren and Stephen Foskett. Considering the cost of GPUs and the rest of the AI stack, idle time is the enemy. That's why it's critical to have a low-latency storage layer to support tasks like training. Hosts: Stephen Foskett, Organizer of Tech Field Day: ⁠⁠⁠https://www.linkedin.com/in/sfoskett/⁠⁠⁠ Frederic Van Haren, CTO and Founder of HighFens, Inc.: ⁠⁠⁠https://www.linkedin.com/in/fredericvharen/⁠ Guest: Ace Stryker, Director of Product Marketing at Solidigm: https://www.linkedin.com/in/acestryker/ Follow Gestalt IT and Utilizing Tech Website: ⁠⁠⁠⁠⁠https://www.GestaltIT.com/⁠⁠⁠⁠⁠ Utilizing Tech: ⁠⁠⁠⁠⁠https://www.UtilizingTech.com/⁠⁠⁠⁠⁠ X/Twitter: ⁠⁠⁠⁠⁠https://www.twitter.com/GestaltIT⁠⁠⁠⁠⁠ X/Twitter: ⁠⁠⁠⁠⁠https://www.twitter.com/UtilizingTech⁠⁠⁠⁠⁠ LinkedIn: ⁠⁠⁠⁠⁠https://www.linkedin.com/company/Gestalt-IT Tags: #UtilizingAI #FlashStorage #AI @Solidigm @UtilizingTech

Many people involved in artificial intelligence don't spend much time considering storage infrastructure. That's the topic of this episode of Utilizing AI, which features Ace Stryker of Solidigm, discussing the role of flash storage in AI infrastructure with Frederick Van Haren and myself. Considering the cost of GPUs and the rest of the AI stack, idle time is the enemy. That's why storage matters for AI. Welcome to Utilizing Tech, the podcast about emerging technology from Tech Field Day, part of the Futurum Group. This season of Utilizing Tech is returning to the topic of artificial intelligence, where we will explore the practical applications and the impact of AI on technological innovations in enterprise IT.

I'm your host, Stephen Foskett, organizer of the Tech Field Day event series, and joining me today is my co-host, Mr. Frederick Van Haren. Welcome to the show. Thanks for having me. So we've been talking about a lot of topics in AI, and we were at AI Field Day. One of the topics that comes up quite a lot, well, I guess anytime I'm talking, is storage. What do you think? Is storage relevant to AI or not?

Well, that's a good question. I mean, surprisingly, buying just GPUs doesn't cut it, right? Those GPUs need a lot of data. And so you have to store that data, not only storing it, but also have fast access with the lowest latency. Hard drives are really not the ideal solution for this. So we have to look at different technologies such as SSDs and NVMEs. And that brings us back to storage, right? How do we assemble all these pieces? Well, I think that there's also a question about basically the hierarchy from storage to data infrastructure to data platforms and structured data and then into the models. And I think it's an open question whether the low-level storage media actually matters to AI. Right. I think if you ask a data scientist about data,

they're not really going to talk about storage. They're going to talk about data pipelines. And those data pipelines consist of multiple stages. And there are multiple efficient solutions for every one of those stages. And in some cases, SSD, sometimes you don't need that performance and latency, but every data pipeline is different. So there's a need for all of them. Well, I suppose that's true. And I suppose that there are some applications

that do need higher performance storage. That's why we decided today to invite on as a guest, Ace Stryker, Director of Product Marketing at Solidigm, who presented at AI Field Day and addressed this exact question. Welcome on board. Hey, thank you so much for having me, guys. I'm really excited to be here. So as part of the Solidigm presentation, I think it was you actually who came right out and said it, you know, is storage important for AI? And where and when and why and all that? Because of course, it's not a monolithic question. So I guess give us a lay of the land here. Where does storage, where's the relevance for storage to AI? Yeah, it's a great question. Thanks for asking. And it's certainly not the first thing that folks think of, right? When you're designing an AI server, when you're thinking about the

infrastructure needed to support developing and deploying AI models, our research suggests 60 to 90% of the spend in an AI server is going to compute, right? And so that is the 800 pound gorilla, right? And that's the first and primary concern for someone designing a system for this kind of work. But as Frederick says so well, GPUs aren't very useful if they're not fed data, right? That's a key part of the process is ingesting the raw data from whatever your source is. You know, let's take a large language model as an example, right? You may use Common Crawl or, you know, similar sources like that. That can be petabytes of raw data that needs to be stored somewhere and then cleaned up and tokenized and taken through the data prep

process, which induces heavy read activity and some write activity as well. And then as you go through, we can talk about the stages of the workflow in more detail if you want. But training, and in particular within training, you have the need to write checkpoints periodically to save your progress in the event of a hardware failure or software issue. That's pure write activity to your disk drive. And the slower your storage, the longer that's gonna take. And while your checkpoint's being written, those expensive GPUs, all that great hardware is sitting idle and waiting for that to complete. And then, you know, God forbid you have an error

and you need to restore from the checkpoint, similar story on an even bigger scale, right? Then you're reading back and everything is idle while you go through that restore process. And so, you know, in a world where you want to minimize that kind of thing happening and sort of be well protected from failures, you're going to be doing a lot of checkpointing, right? And the storage devices are going to be a key part of your overall kind of training experience as you go through that.

And then on the other end, on a deployed model, particularly as we see more and more work moving to the edge, it becomes critical to have low latency storage there where you can read in your inputs, whether it's someone typing into a chat bot or a security camera somewhere, or whatever that source might be, to run that through the model in a FordPass is a storage intensive step of the process as well. So there's lots of places where storage plays an important role, particularly as it relates to, you know, the powerful and expensive

compute components, making sure those are maximally utilized, making sure that you're getting the most out of all the money you spent on the most important components in an AI server. Yeah, so when you look at AI workloads, I mean, they're far from being static, right? So they're extremely dynamic in the sense that a storage profile can

change when the data scientist gets more data, gets a different type of data, or even uses a different type of accelerator. What kind of drivers do you see for storage with regards to AI workloads? Well, as you say, it really depends. AI is not one monolithic thing, right? So we do see some trends in terms of the activities that storage is called upon to do in the course of a workflow. But it will vary widely from one application to the next. And then when you layer on things like concurrency, multiple data scientists working with the data set at the same time, or multi-tenancy where a single system resource might be used

on different projects entirely at the same time, it generates these really kind of more complicated mixed read-write IO profiles. But if we were to look at one kind of workflow in isolation, what we would see is up front, you're writing a bunch of data to the disk as you ingest it from whatever your source is, right? And then you take it through the ETL or the data prep stage where you're taking that raw data, you're cleaning it up, you're deduplicating, you're putting it into nice columns and rows and creating tokens that then are shown to the model in the training phase. And that's done in a deliberately random way, right, to prevent overbiasing and creating any kind of issues with the usefulness of the model itself.

So that's a lot of random read activity. The result of that is the model itself, which isn't huge in terms of absolute size. Right. It'll depend, again, on the model. But if you look at, you know, GPT-3, I think that's a 500 gigabyte or so model that is the output of the training phase. Right. And then those checkpoints I mentioned, those can be quite big. Those can be on the order of, you know, depending on model size, several terabytes at a time. And it's really up to, you know, the user to determine how often do you need to checkpoint. It may be every 30 minutes decision, but whatever that period is, you're doing heavy writing every so often. NVIDIA, I believe, has laid out guidelines for how long the duration of a checkpoint, you know, how long it should take to write a checkpoint, depending on the periodicity. And so you can use those guidelines to sort of calculate bandwidth requirements for a storage device. And then at the edge or whether inference is occurring there or in the core data center, we're seeing performance and capacity demands start to grow there as well. We're seeing an evolving regulatory landscape where, you know, beginning in the EU,

but here in the US as well, we're starting to see more sophisticated laws and requirements come online around data retention. You know, say you have some really sensitive use case and your model misidentifies somebody at an airport or something, right? You want to have a paper trail there and be able to determine what happened and improve the model. And so we're starting to see more and more stringent requirements for storing the inputs and outputs at the inference stage as well. Yeah. So you talked a little bit about the edge and certainly, you know, most of the training is happening in data centers. So I think it's easy to visualize that a lot of the storage ends up in a data center. When we talk about the edge, how should we see that?

I mean, I assume in most edge cases, you don't have access to the same power and cooling and space as you have in a data center. And so on top of that, the concerns are, how do we do this, right? So in some cases you might collect the data on the edge and then pass it along. Can you talk a little bit about the differences between a data center and edge storage solutions?

Data center, you're really talking about a single location that's really highly powered, you know, racks and racks of hardware. Whereas at the edge, it can be tremendously varied in terms of like how those devices show up. You know, it could be an edge device in a retail space or a restaurant. It could be, you know, we have a partner of ours, Supermicro, that has designed an edge server that hangs on a telephone pole. It's completely enclosed and ruggedized,

and it, you know, does a lot of work locally there, and it's built specifically to that environment. So that's really driven by, you know, these use cases where more and more intense work needs to be done at the edge, more real time inference, more lightweight training and reinforcement learning happening at the edge where you're not feeding your inputs all the way back to the core data center to initiate another round of training. Right. You might just be refining locally and making smaller improvements. So as demand rises for these kinds of use cases that demand, you know, more performance at the edge, an ecosystem has kind of grown around that, right? Solidigm's part of that in terms of, you know, devices that we offer that are a low latency, that are a good fit for some of these edge applications. But we fully expect in the near to midterm, we're moving to a world where you've got to think outside of the core data center.

When you're talking about an AI infrastructure, the architecture is going to be distributed it may be you know edge applicator excuse me edge devices feeding into you know what we call the near edge which is sort of a regional data center which then kind of feeds back to the core data center and we're going to start to rather than then the requirements of the architecture being dictated by, you know, the limitations of the hardware, as more hardware comes online and becomes more and more capable, those requirements are going to start to be dictated by what's ideal for the use case, right? So we're excited about that. And it's a very fast moving space, like everything in AI, but we think it opens up all kinds of really cool applications for AI.

Yeah. Can you talk a little bit about the different form factors? I mean, everybody who has an iPad or desktop computer, we're now having M.2s and U.2s and whatever form factor exists. Do you see a need for a new form factor, you know, being in a data center as well as on the edge? Because, you know, when you look at an SSD, it still has relatively the same form factor as a hard drive, right? I assume that with the electronics today, there is a way to make this a lot smaller and a lot more efficient. Yeah, it's a great question. And there's been a lot of development in that space over the last 10 to 15 years. You know, where we started was in a world where the form factor was entirely driven by what works for hard drives, right? So you had three and a half inch, you know, hard drives, and eventually for laptops and smaller client devices, two and

a half inch. And then that was essentially what was used in the early days of flash memory, you know, NAND devices as well. But as the industry has matured and as, you know, close collaboration with partners and customers has paid dividends, what you started to see in recent years is the emergence of a whole new class of form factors that never existed for hard drives, that frankly, a hard drive could not fit into anymore, right? And so some have called these the ruler form factors because some of them are long and skinny like a ruler. EDSFF is another acronym you might hear.

But what it essentially means is we have now form factors that are purpose-built for NAND that are great for optimizing density in, for example, a rack in a data center. So you can fit far more per one use space in a data center rack using one of these new purpose-built form factors than you ever could with one of the legacy form factors that we just kind of inherited from the hard drive world. So following on that, actually, one of the things I think that many of our listeners may not be aware of, in addition to the form factor question, is the capacity that we're talking about here. Because I think, well, normal people think of storage nowadays in the small terabyte range. They think, I'm going to buy a storage thing, and it's got eight terabytes of stuff, or maybe a little more, maybe a little

less. I don't think they understand what we're talking about when we're talking about the size of each of those modules, as well as the size of or the capacity that can fit in a server like the super micro servers that you showed at AI Field Day last month. So before I shift gears here, I do want to dive into that. Just level set people here. What are we talking about here in terms of capacity and performance? Yeah, it's a fast movingmoving space. I come from sort of the PC storage world in a prior role where two terabytes was a mind-blowing amount of storage to put in your laptop or your desktop PC. On the data center side, we're talking now where you see hard drives max out on the data center side currently is 24, maybe 28 terabytes, which is huge, right? But we actually have, there's a product in our portfolio now, the P5336, available in capacities up to 61 terabytes in a single device. And of course, as time goes, you know, we'll only go up from there.

But that's that's a particular product that we've had tremendous interest in from our customers. They love the density story there because not only is it, you know, fewer drives to store your data set or whatever you're working with, but that has all sorts of implications. You're drawing less power if you're using fewer drives, so you're saving a lot of money that way. Your cooling requirements are simplified. The amount of space you need in a rack to fit X terabytes or petabytes of storage is now far less, and so that can have implications in terms of your data center footprint, right? So it kind of has this cascading effect where density is really driving a lot of improvements around what we call TCO, total cost of ownership.

When a customer wants to optimize for that, the first thing we talk about is density. And hey, you know, these new high capacity drives are unlocking all kinds of new possibilities there. So, you know, you mentioned Supermicro, they have a petascale server that they sell. And I believe currently it offers like up to two petabytes in a single server unit that they sell. So pretty wild time for storage and a good showcase of kind of the relative advantages of NAND. It scales in a way that, you know, the hard drives are having a hard time keeping up. And on the performance side, too, I guess I should mention that each of these SSDs offers, you know, orders of magnitude more performance than a hard drive as well. And so it's not just the fact that they're big, it's the fact that they're fast. And a lot of people, I think, do understand that after having moved from hard disks in their laptop or desktop to an SSD, they realize just what a tremendous difference that is. And that's the same in the enterprise

space. Yeah, absolutely. That's a key difference as well. Typically, when we talk about performance, we talk about sequential numbers and we talk about random numbers. And that's, you know, we don't need to get in the weeds there. But depending on the nature of your workload, you know, how contiguous is the data? Are you moving big chunks or small chunks will determine whether that workload really relies on sequential or random performance. In the sequential world, it's a huge difference. You know, hard drives, you'll see numbers, two to 400 megabytes a second, something like that. On a PCIe Gen 4 SSD, you're talking about something in the neighborhood of 7,000 megabytes a second. And as Gen 5 comes online and begins to

take over, you're doubling that again. You're talking 13,000 to 14,000 megabytes per second. On the random side, the difference is far more drastic, if you can believe it. So random operations are not something that a hard drive does particularly well. There's a seek involved in moving around the plate with the little arm, the mechanical arm in the hard drive to do that kind of work. And a solid state drive, as the name suggests, has none of that, right? No moving pieces. And you can do upwards of a million operations per second on a solid state drive. And so those parts of the AI workflow that rely on really either are going gonna benefit, but in particular, the random stuff that we see in things like data prep and transformation

and things like inferencing are gonna show a tremendous leap in wall clock time, as we call it, by virtue of simply switching from a mechanical hard drive to a solid state drive. So I wanna get back to the core question that we started with, though. Does storage matter for AI? And it sounds almost like you're saying that storage may matter sometimes. And because it matters sometimes, you have to have great storage performance and capacity

and features all the time. Is that what you're saying? Or am I mischaracterizing you here? Are you saying that storage is critical? Because a lot of companies are out there and they're saying basically, you know what? AI doesn't have extreme storage requirements.

You just need a basic level of storage. You hear about hyperscalers and they're using plain old SSDs to do inferencing. Yeah, storage doesn't matter anymore. But yet some of the points you made, I think are well received and definitely true that there are points in this continuum where storage matters a great deal because otherwise things are sitting idle. And that's like the worst thing that can happen when you're talking about, you know, infrastructure that's literally millions of dollars sitting, just doing nothing. Or does it really matter always? Well, slow storage can certainly always be an impediment, right? So there is a threshold

under which you simply would not want to go for AI work. And that's why we're so bullish on the role of SSDs across the data pipeline. But when you start to get into more modern storage solutions, PCIe devices, compared to what was available even ten years ago, just, you know, phenomenal read and write speeds. The name of the game really becomes a GPU utilization. So, you know, that's you're not going to run into many situations where apart from checkpointing and restore, which we talked about, where where your storage is going to become

a bottleneck upon which everything else stops and waits. But you will see situations where your GPUs could be better used if data were fed to them faster. They could chew through that data faster. And so that's why you see, like, for example, one of the new AI benchmarks that's kind of gaining steam by ML Commons is called the ML Perf Storage Test. And what that actually returns, the output of that is like how many GPUs can a storage device feed at minimum 90 percent utilization? So that becomes, you know, sort of the benefit when you get into faster and faster storage is making better use of the compute in your device. And then, you know, aside from that,

capacity constraints, as we mentioned, are going to become more and more critical. These data sets are getting huge. The models are getting more and more sophisticated. So performance is a piece of it. Absolutely. But we fully expect as the ecosystem matures and the models get better and better at what they're designed to do data is going to be the lifeblood of that process right and having a great place to put it with high capacity and

performance is going to enable things that weren't possible you know five years ago and weren't possible five months ago that's how quickly we're moving now in this space. Yeah, I think if money is not an issue, you would always buy the fastest, you know, and the largest drives you could get. But I think the reality is, too, that it is more of an ROI conversation where if you don't want to spend all your money on storage, you would spend it on finding ways to optimize. And I think that there is definitely a difference between the training and the inference side.

Where we're training, it's where all your data exists. Where, in theory, you can only process a percentage of your data. So does it make sense to have all your storage on the fastest storage? Do you see that too, like a difference between the training and inference requirements regarding to storage? Yeah, we certainly do. So we've done our best working through, you know, good partners that we have. Vast Data is one that we've explored this significantly with, and there are others as well, to understand, you know, what are the unique requirements per stage of an AI workflow?

Because there's really two ways that someone could then tackle that problem and develop a solution, right? You could ask yourself, what's the optimal architecture, stage-specific architecture, right? What's the best hardware stack that I need to do training versus inference? Or you could say, hey, I want a single architecture. And so what should I put in my box that's gonna sort of give best performance in aggregate

across the different stages? And that'll vary from customer to customer. But absolutely, like the demands that the workload is placing on the storage device are quite different. For example, in training versus inference training, we see, as I mentioned, random reads, right, as you're exposing your prep data, your tokens to the model. And then we see sequential writes as the model itself is written as an output of that. Whereas in the inference phase, generally, depending on whether you're talking about

an LLM or sort of a rich media model that might deal with video or photo, those are going to induce different kinds of storage demands on the device as well. So we see one area where high performance storage matters a lot on the inference side, because if it's just an LLM, there's not a lot for, for example, the GPU to do there. It'll be slightly more involved if you're generating images or videos in response to a prompt. But for a straight large language model, storage tends to be the critical resource in terms of turning around that insight and responding to the user as quick as possible. I think a lot of us are also expecting that the industry is going to rapidly move past LLMs specifically and into more, well, I dare I say, multimedia applications of AI. And I necessarily, I mean, everybody can understand that video is bigger than written texts. I mean, it's just how it is. I think that that suggests that storage

might actually become critical, even if it doesn't seem like it now. Because right now, I mean, we're literally dealing in most cases with text. And it's a lot of text. It's the whole dang internet of text. But it's text. And yet, if you look at what's being announced here in the last month or two, there's a lot talking about images and video. And, you know, specifically, you know, I mean, we see Google talking about, you know, Gemini taking video streams and cutting that up into individual frames and processing each frame individually. That's a lot of data and a lot of IO, right?

I mean, it's not just an order of magnitude. It's multiple orders of magnitude larger than, you IO, right? I mean, it's not just an order of magnitude, it's multiple orders of magnitude larger than, you know, an LLM processing text. Yeah, absolutely. So, you know, these other so-called multimodal generative AI models, you know, they've been around for a while. Folks are probably familiar with, you know, DALI or stable diffusion or some other models that are commonly used to output you know images you know give me a uh um you know a wiener dog in the in the as an oil painting or something right like people will goof around on the internet all day long and do that stuff uh but gemini is a great example of uh you know where the future is heading and the possibilities are. So, as you develop models that are more and more capable,

and that frankly, the users are going to expect more and more of, yeah, it necessarily demands more data upfront, right, to show to those models as part of the development process. And as well, the workloads are going to become more and more intense on the inference side as well as the responses are generated. So it's an exciting space. And as quickly as hardware requirements have been, you know, rising over the past couple of years in the first couple of waves of innovation here, I fully expect it's only going to speed up from here. So it's a great time to be in the business,

great time to work at NVIDIA, obviously, but we're excited to partner with them as well and make sure that our customers get the best storage to help optimize their spend and get high-performance, efficient AI server infrastructure. Now, I don't mean any offense by this, and SolidIME's products are not cheap. But compared to the cost of GPUs, your products really are not a major line item for a lot of these big companies, right? I mean, we're talking about companies that are basically optimizing spend to get as much GPU horsepower as they possibly can. One of the things that I heard from one of the companies that's a storage platform partner of yours that is serving, well, according to them, a majority of AI training in, you know, they were

saying that although their product is fairly expensive, it doesn't even matter. It's completely insignificant when it comes, you know, when compared to the cost of GPUs. Is that really the picture? Is storage basically a small portion of this pie and is it growing? Yeah, it's a great question. And there's a couple ways that I think people look at that. You can look at the storage proportion of the bill of materials in an AI server. And then you'd be looking at it relative to things like GPU costs and it would look quite small, right? And you could say, well, I could put the best storage imaginable in there,

and it still wouldn't be a significant, you know, spend for me. But the way that we see people looking at it, you know, and frankly, part of where I'm spending my time is a lot of times the comparison is made relative to the cost of hard drives. And, you know, procurement departments, for example, look at it simply on the cost of hard drives. And procurement departments, for example, look at it simply on the cost of dollars per gigabyte. And they'll say, oh, per unit, per gigabyte, hard drives are a lot less expensive than SSDs. Therefore, that's the smarter purchase, right?

And it's our opportunity at Solidigm and among others, NAND vendors, to sort of try to illuminate, you know, the long-term story around cost of ownership, right? So, as I mentioned before, with an SSD relative to a hard drive, not only are you getting, you know, potentially much more storage in a single unit, that unit itself physically is also smaller because it's in one of these modern form factors. All that together means you're fitting a whole bunch more storage in fewer units into your rack. So you're saving on power, you're saving on cooling, you're saving on rack space, right? And all these things are huge drivers of long-term costs. We've seen, there was a study published by, I think, Meta and Stanford University,

where they indicated that 35% of their server's power footprint was consumed by storage in that use case, right? And so you can imagine those numbers over the course of five years, dwarf your upfront spend on the drives themselves. And so there's significant savings to be realized by looking beyond dollars per gigabyte. so there's significant savings to be realized, by looking beyond dollars per gigabyte, maybe there's a more useful metric,

effective dollars per gigabyte, or something that sort of accounts for, the costs of long-term ownership of these devices. But in those terms, we think that SSDs make a clear and compelling case for themselves as sort of the cost leader. And not just any SSDs too. I mean, again, the difference in cost between buying just some garbage and playing like the best, fastest SSDs is probably insignificant over the life of that infrastructure and is probably going to make enough of a difference that you might as well

just buy the good ones. I mean, that seems to me to be the kind of takeaway here. I mean, yes, you want to save on the costs, but it really isn't going to be a tremendous difference to buy the good stuff that's fastest and most reliable instead of buying the kind of mediocre stuff, especially if it means that you're going to keep your infrastructure up and running, you know, even for a short amount of time, you know, the difference of having it run versus having it not run, whether it's waiting for data, whether it's waiting for repairs, it just seems kind of like a no brainer to buy the good stuff. Yeah. I mean, I agree with you, Stephen. I think that's well said. And that is only going to become more stark as time goes on and the data requirements keep rising.

So you may buy something today that feels like, you know, sort of an optimal spend and you're trying to save a little money on the storage piece. And then, you know, we've seen how quickly this industry moves. So a year from now, 18 months from now, you may be playing a whole different ballgame, right? And so investing and future-proofing a little bit in high-performance storage is going to ensure that you can keep up with competition as time goes on. Yeah, my background is in storage before I got into all this. And I think that it's interesting how this small component has such a major ripple effect across literally everything we do. A lot of the things that we do are made possible by advances in storage, especially Flash. And so it's really amazing to see all that this little component makes possible. So thank you so much for this conversation.

Thank you also for presenting at AI Field Day. I think having Solidigm there, having Supermicro there brought a different perspective because, you know, a lot of the other folks are talking about models and software and, you know, here we are talking about something you can hold in your hand. And so if you're listening to this, check out the AI Field Day presentation, just Google or search engine, your favorite search engine, or ask maybe a chat bot for the AI Field Day 4 presentations by Solidigm. You'll be able to find them as well on the Tech Field Day website and YouTube and LinkedIn. Before we go, where can we continue this conversation with you, Ace,

and where can people learn more about Solidigm Storage? Yeah, so we've recently launched a new landing page on the website. You can go to solidigm.com slash AI, and that has a couple of neat things. It's got some customer stories and some kind of architecture stuff, but it's also got a great video conversation where we delve into kind of storage requirements by stage of the workload. So if folks are interested in that, I recommend you check that out. I will also be at NVIDIA GTC this week. And so if this is something you're interested in learning more about, please do reach out on LinkedIn or elsewhere. I'd be happy to meet up with folks and continue the conversation there.

Great. I think we might be at GTC this week too. Frederick, are you going? Yeah, I'm definitely going. It's the place to be for AI, definitely. Thanks for listening to Utilizing AI, part of the Utilizing Tech podcast series. You can find this podcast in your favorite podcast applications and on YouTube. If you enjoyed this discussion, please consider leaving us a rating or a nice review.

This podcast was brought to you by Tech Field Day, home of IT experts from across the enterprise, now part of the Futurum Group. For show notes and more episodes, head to our dedicated website, utilizingtech.com, or find us on X, Twitter, and Mastodon at Utilizing Tech. Thanks for listening, and weastodon at Utilizing Tech. Thanks for listening, and we will see you next week.