UT08x01: Accelerating and Protecting Storage for AI with Graid Technology

Modern AI servers are loaded with GPUs, but if you spend too much time, uh, waiting for data, then you're not getting much out of them. This episode of utilizing tech focused on AI at the edge with soy features Kelly Osborne of grade technology, discussing the latest in data protection and acceleration. Welcome to Utilizing Tech, the podcast about emerging technology from Tech Field Day, part of the Futurum Group.

This season is presented by soy and focuses on AI at the edge and related technologies. I'm your host, Steven Foskett, organizer of the Tech Field Day event series. Joining me today from SOY as my co-host is Mr.

Scott Shaley. Welcome to the show. How's it going, Steven?

It's great to be, uh, have an opportunity to join you this season. So I'm excited about what we're gonna be talking about today and for the rest of the season as well. Absolutely.

And, um, you know, what we're talking about is basically all the various components that are required to build AI applications at the Edge, AI applications in the cloud, AI applications everywhere. But really we're gonna focus in on the importance of basically the full stack under ai. Yeah, it, it's a great opportunity to kind of talk to that point.

'cause, you know, AI is our shiny object today, and we know that it's here to stay, uh, for the foreseeable future, but we still have to start looking at, you know, 2025 is a year we start to optimize our infrastructure around all the AI stuff that's been booming for the last year and a half, two years. So, uh, taking this opportunity to meet with folks like our, our guests today on ways to manage, manipulate, protect, and take care of all that data that we're dealing with around all these AI workloads is very important for us. Yeah, as a storage nerd, uh, it always gets me when people don't take storage seriously, because the truth is, you know, you can't just assume that it's gonna work.

You can't assume that it's gonna be reliable and redundant and performant and so on. That's why we've invited, uh, Kelly Osborne, senior director of OEM and channel business at, uh, grade Technology Inc. As our guest today.

Kelly, welcome to the show. Thanks, Steven and Scott. Appreciate the opportunity.

Um, as, uh, Steven mentioned, my name is Kelly Osborne. I'm with Grade Technology and I'm in business development. You can find me on LinkedIn at Kelly Osburn.

Not hard to find. Um, just make sure you put the E and the URN and you can find me. Kelly, tell us a little bit to kick things off.

Um, what is grade technology? I mean, I think most people have heard of RAID technology. What is grade technology?

Um, so we were founded about three, three and a half years ago to solve, uh, a problem in the space around NVME, high performance, uh, flash storage. And essentially the problem is when you develop a server or install a number of these drives in a server, um, once you impose some sort of raid data protection on them, you create bottlenecks that don't allow you to achieve the full performance of those drives. And when you spend a lot of money on drives like that, you want to get to get all that performance.

So we identified that as a problem in the market. So G Grade is actually kind of a twist on GPU rate, and that's exactly what we're doing. So we have a product called Supreme Rate, and that is a software rate stack that deploys on Nvidia GPUs to accelerate the rate operations and allow you to achieve very close to a hundred percent of the performance of those expensive drives.

It's very interesting, you know, to that point 'cause uh, you know, soy being the solid paradigm on storage and being a, a self per lam storage geek, much like Steven, uh, one of the unique things about NVME drives when they first came out, and you know, with I being one of the pioneers in that front with our long history was NVMe stripped out a lot of what you could do with some of the other interface technologies. And so having to find a new way to work at a, you know, formerly existing problem is nice to see that, uh, g rates coming in and doing some of that work. Yeah, we, um, I think identified two problem areas with traditional hardware r where you plug your drives into a controller that's sticking in a slot, you artificially create lane limitations.

Think of a toll booth on a super highway. Um, when you think of these solid dyne drives, each of them can maximum performance. You need four PCIE lanes.

So if you connect four of those to a card that has 16 lanes, you've already hit a hundred percent of the performance of that card. So how do you go past that? Um, we allow that because the drives in our world are connected directly, directly to the motherboard.

And if you have 10 drives, you need 40 lanes. We deliver that 40 lanes of performance because we have a patented, uh, technology that is out of path we call peer-to-peer DMA that allows the data from those solid nine drives to make it to the CPU or the GPU directly across some mother board without going through a gate, if you will. The other side of that is software rate, where as you have more and more of these drives that are really fast, they're, they're real, uh, um, really want a lot of attention.

So you get a lot of interrupts and the CPU gets really, really busy CPUs are, are also are not very good at mathematical calculations compared to A GPU. So you start to choke on the parody calculations for rate five and six in those kind of environments. And so the hardware rate scenario doesn't scale and the software rate scenario doesn't scale.

We really shine when you get beyond four or five drives in a server and go all the way up to 32 in a single chassis. Yeah, this is really a, a big trend in the HPC and AI sector overall is, uh, basically having direct, as you said, DMA direct memory, direct access between, uh, peripherals and ships generally. And I say chips as rather than CPUs because that's also what's going on in many cases with, um, a lot of the, the GPUs and other, uh, AI acceleration sys, you know, engines in, in modern systems instead of being, you know, sort of a, a star topology with the CPU at the middle and everything going through the CPU, it's almost a mesh or fabric technology where a topology where things are going direct.

It's actually more like a web where the CPU is still at the center, but there's links that go sort of around the CPU, is that right? Very much so. So the, the way NVME works, it plugs directly into the motherboard of the server chassis.

So the CPU, the memory, the PCIE slots, the drives, everything can see each other. We call that a, a peer-to-peer, uh, direct memory access, if you will. So what we do is act more like a traffic cop.

When a read comes to us, we know where the data is and we know where it needs to go. So we just tell that drive, send the data over here, we don't have to read it in and forward it, that would introduce latency and create hops and other things in your data path. So, um, in the GPU world that's called, uh, Nvidia created something called magnum IO or GPU Direct storage, and that allows, uh, drives to send data directly to, um, A GPU and bypass that host, which eliminates that extra hop for those high performance workloads.

The problem is when you impose raid and data protection into that, you, you still have to force everything to go through something else. And we now have that built into our new version. Yeah, that brings up some interesting ideas.

I mean, you, you mentioned earlier, you know, kind of the idea of how many slots, right? And we look at new reference designs and things like that, and all the fun stuff that's getting put out there around, you know, how to fit so many of these chips to, uh, Steven's point into a system. There's still a limited number of slots that are associated to that, whether it be, you know, PCIE slots in the hole where, where the, uh, product you guys are developing is at, or even the number of slots where a number of drives we can put in those are in place.

So we get into those conundrums of, okay, how do best stuff the box and make sure that we're getting the maximum use out of that from both the storage technology, the partner technologies, like the, the raid, uh, Supreme Raid product and things like that. And you guys have recently worked on something new too that, uh, has recently made the press, so I'd love to hear a little bit more about that as you kind of started down that path just a moment ago. Sure.

Uh, we created a product, um, called Supreme Raid. Once again, that's our product. Uh, the new edition that we have is called Supreme Rate ae, which is AI edition, and it implements several new features like GPU direct storage, uh, intelligent data offload, um, and that peer-to-peer.

We also are incorporating a lot of NVMe over fabric and incorporating our product into, um, parallel file system environments like SAF and BGFS and Luster, and providing that local data protection for those storage nodes. The other thing, uh, that we've done, uh, our traditional product is required to dedicate A GPU, um, and these large GPU servers like a DGX or like you see some of these, uh, servers from big manufacturers that have, you know, eight Hopper or Blackwell, um, SXM chips that are, you know, envy linked together. There's not a lot of real estate to put another card in, but you have so much GPU performance, we now have a version of our software, the AE version that can run on one of those GPUs you already have, so you don't have to have yet another GPU in there powering our software.

So, um, you know, that AE version, uh, we just, uh, demonstrated and showed it off at GTC, um, in San Jose. So, um, that's getting a lot of attention because when you buy servers like that, that have so many GPUs, if you are not feeding that data to that beast fast enough, you're not getting the full utilization out of that. And those are typically very expensive and you want to get very close to a hundred percent utilization if you can.

Yeah, we were recently, uh, at AI Field Day, uh, we had a couple of presentations focused on that at very topic, and one of the things that was pointed out was a study that showed that, um, the majority of, uh, enterprise, uh, AI deployments, and this is not hyperscaler ai, this is not, um, you know, sort of wannabe ai, this is the actual enterprises running AI workloads was, uh, the, the GPUs were utilized a very low percentage. Uh, if you had asked me, I would've said, you know, during actual work, you know, a production GPU cluster, you know, yeah, you're probably, you know, getting 50, 60, 70% utilization outta those things. No, 35% of the respondents or the not respondents.

'cause this is actually a me a a metric based survey too. So it was actually measuring it. 35% of the respondents were getting less than 15% utilization outta their GPUs.

Can you imagine if a factory bought a $30 million, you know, stamping machine or something, and they let it idle, all, you know, 85% of the time somebody would be fired. And, and, and, and, and yet that's what we're looking at when it comes to these GPUs. And there are a lot of reasons for that.

But again, this was during active use, so the main reason was those GPUs were not being fed with data. Exactly. And so what we, what we see in those environments, a lot of times, uh, those GPU servers will have local NVME eight or 16 drives.

They'll run that strap space in rate zero, and that'll give them the best performance they can possibly get because they don't have any RAID bottlenecks. Um, but then the issue is they have to start a whole job over again if there's a dry failure, and it's not a negligible situation that you could have nonrecoverable errors or A-P-C-I-E error or an actual drive failure. And so then you have to roll back to a known checkpoint.

And so it may look upfront like you're getting better utilization, but in the long haul you may not. So, um, our, our theory is let's provide raid five or six or 10 or one or even erasure coating protection of those drives in that machine, but let's not impose a bottleneck to do it. So you get close to raid zero performance, and yet you have that protection being able to lose the drive, continue running, rebuild the drive, um, the kinds of things that we're used to with raid.

Um, and, and so that's really our focus. So it's an interesting step too. 'cause I mean, as the drive guy, right?

I, I'm gonna sit here and tell you I have a, I have a high quality product, but there's always those situations where you do need that protection. We know you can't get rid of it. And so you have to do something to manage it and, and eliminate the overhead to your point of what you're trying to do with those particular products.

And as the data sets grow, there's always this term of blast radius. And one of the focuses of what RAID has always been about, uh, the, the idea of raid, not your product, but just raid in general, is to prevent those things like that blast radius and other issues. But even with these high quality drives, to your point, there's fewer of them in the system, and you still got the legacy infrastructure that's around them.

NVMe has definitely taken over the world a little bit, which is nice to see, but it, it's not, you know, all in everywhere and systems still aren't optimized for it. So having solutions like yours tied together with the products and things like that, you know, we recently showed off a fun new toy at GTC two, so it's exciting to see how, you know, our advancements in technology combined with your advancements in technologies aren't just there for fun, but to actually help our end customers solve problems. Yeah.

And that's one of the interesting things, isn't it, Scott? Um, you know, you and I go way back in this, in, in history storage has been a bottleneck for a very long time. And in many ways, you know, I'm sorry to say this, but storage is still the bottleneck, but it's not the storage's fault.

It's the way the storage is being used. That's the bottleneck. You know, I mean, these, these systems have incredible amounts of bandwidth, especially when you take, I mean, these modern SSDs.

And as, as Kelly was saying, you know, the, the, you know, NVME drives with, you know, four, you know, PCIE lanes, you know, that's a lot of bandwidth, especially when you have multiple ones. But the problem is most people don't have the capability to actually use these things efficiently. They don't have the capability to use all that bandwidth, or as Kelly was pointing out, they're strangling it behind a controller that kind of, you know, acts as a bottleneck on it.

Um, it's a little frustrating it, isn't it? 'cause we finally have all this incredible performance and yet people are still not using it. Yeah.

And then that's another aspect of kind of as we evolve this ecosystem and these reference architectures and, and putting companies together like we are with, with grade and soy and even, you know, the GPU guys, you know, if you want, you know, owners of GCC nvidia, um, as we evolve these ecosystems, the need to isolate the bottlenecks and solve them is one thing that we're actually starting to see with things like the new E-D-S-F-F form factors that Soddy has been a, a kind of a big leader in, because adding this new form factor, which makes it flash only you can start to realize that now I know I'm not killing the hard drive. It'll still be there. It has its reasons, but in the performance bottleneck situations where you really do need to really make those systems work better, truly DeVol developing that system to work with this new interface and the subsequent technologies that are tied to it, because we wanna make everything work well together, is, it's nice to see.

I mean, to your point, Steven, we go back far enough that I remember putting one SSD in a box was a huge success, and now we're talking about let's a little, you know, put 32, 64, however many they can start shoving in these boxes. So it's, it's a great evolution. I'm excited about what's next.

And I, I wanna call attention to another thing that Kelly said and make sure that people heard it. Um, one of the things about these modern, uh, AI servers and, you know, and, and frankly, AI servers are gonna be everywhere, not just in the cloud or in training or anything. I mean, especially at the edge.

One of the interesting things about these is that they have a ton of compute power on the GPU or, you know, AI accelerator side way more than they do on the CPU side, and yet many, many supporting systems, I'm just gonna say that generally. And that would include storage, uh, use the CPU to do the work. Kelly, you said that you're actually doing the RAID calculations in the GPU.

Tell me a little bit more about that. Are you really doing them in the GPU in the right, right there in the course? Yes.

So we, we have a technology that's intelligent, um, basically intelligent parody production and what we're, what we're doing when a, when a right comes in, we actually allow that right to go straight to the drive across PCIE. It doesn't flow through our cart. We do have to get a piece of that data into the GPU, the NVIDIA GPU, and that's where we calculate the parody.

So then we lay the parody down, and then so only a tiny piece of data is being written that comes from, uh, the GPU itself. And then we acknowledge back. So that's how we achieve very high rate performance and very high iOS per second.

On the re side, um, I mentioned when RES come in, we just do a redirect and tell the drive, send the data straight across PCIE, um, to improve that performance. Coda cores, um, on GPUs are far faster at mathematical calculations. RAID parody is nothing new.

It's based on Reed Solomon, uh, depends on which level you're using, but those are just intensive mathematical calculations to generate this information that allows you to create data protection around these drives and around your data. Um, and we're not doing anything different in the algorithm, we're just implementing it in a way to maximize performance of these MVME drives. When you, when you look at hard drives, we mentioned hard drives, hard drives benefit greatly from things, things like right cache.

So, uh, traditional hardware rate controllers, um, that you see out there, you'll have batteries on 'em, you know, or capacitors. And what that is, is for right cache. So the data comes in, we tell the application the data's down on the drive, and then then it gets written to the drive on the side.

And then if a power goes out or something, you need to, you, you need to be able to have, you know, that cache backed up because the application thinks that's all written to disk. Ironically, in our world, we identified that, that kind of thing, caching, right? Caching actually introduces latency into the data path.

That's why we allow the data to go right from the CPU or right from the GPU straight to the drive. All we're doing is writing the parity. And so that, that improved performance greatly.

Um, and, and so that's, that's why we do it the way we do it. And it's, it's a very simple, very simple concept once you start to understand what we're doing. Yeah, those of us who've actually done some, some work on AI at the edge, um, I know that we've experienced the, the joy of tensor processors and GPUs and offloading this stuff.

I mean, you know, you go from trying to inference on A CPU to adding in a moderately powered GPU or a, you know, TPU module, and suddenly you're able to do not just a little bit more, but an order of magnitude more inferencing using basically the same hardware. In fact, many of us are using pretty low powered systems at the edge with pretty high powered GPUs there. And so it's analogous to e exactly that.

So you're doing, uh, object detection or, you know, facial recognition or whatever it is, using the GPU cores. Well now you're using those same cores to do the raid. It's the same thing and it's offloading that CPU and it's allowing you to use a smaller, lower powered, you know, less cooling, et cetera, um, in the CPU side.

And you're able to leverage that as well. And as we said, in many cases, the GPUs are not maximally loaded in many cases. There is a little bit of bandwidth there, there's a little bit of, of slack space that you can use to do these calculations.

So that's, that's a big, a big benefit. Um, and, you know, from all the way, all the way out to the edge, do you have many people using, um, this sort of technology there in, in, in edge and inferencing use cases? Oh yeah.

So I, I, I'll even mention some real customer scenarios. Um, we have one, uh, this is a company that does, they've kind of started to blend AI machine learning with CFD, computational fluid dynamics and, and simulations. And they specked out a server from a well-known server manufacturer, and it had 24 NVME drives Gen four, so that should be seven times, uh, three and a half times 24.

So, um, somewhere around 80 gigabytes of a second at rate zero from for right performance the customer needed for, to get their application to run properly. They, they needed around 40 gigabytes a second to write to those drives. So they bought the server, set it up, tested it with read zero, it was great.

They set up, uh, the software raid that's built into Linux, Linux, MD raid, not picking on any one Linux because it's a ubiquitous freeware. Um, and once they set up raid five on these drives, they got one gig a second. So the drives were theoretically capable that writing at 80 in, in total.

Um, and they were only getting one. We demonstrated 68 out of 80 with our technology. And so the customer has solved that problem.

We have another similar situation in, uh, a customer that's got a large high F frequency trading model database. They actually have, um, 22 servers that happen to have 32 solid dime QLC drives in each one. They're writing huge amounts of data really, really quickly.

And they have to keep this for analytics and forensics and, and historical, you know, logging and similar problem. They were only getting one or two gigabytes a second with, uh, the traditional software array. So in that kind of environment, we've come in and, you know, helped out the customer, the customer sat, retrofitted, whatever you wanna say.

The challenge now is how do we, how do we start to condition server manufacturers like this to understand where, um, the performance bottlenecks are and how we can solve that with great technology. And we're making a lot of inroads with these server manufacturers. That's, that's great to hear.

And, and to that point, like we, I've talked a couple times now about the ecosystem. It's good to hear where that's, you know, seeing some traction from your perspective as far as those efforts and things like that. When we, when we kinda shift back to kind of what Steven was mentioning about the edge environments, there's also the concept to keep in mind that the throughput capabilities, he mentioned underpowered CPUs, for example, in the systems, being able to do anything to relieve that stress is definitely valuable to these customers.

I love the, you know, between the two of you, you had the, the super highway with the, you know, the pinch points or the manufacturing company that spends way too much money on a stamp maker that makes one stamp an hour instead of a million stamps an hour. It, it all shows to what we're trying to do here in this, uh, amazing new ecosystem we're developing and excited to see how the AI train continues to move forward and gets more properly utilized, I guess is the best way to put it. We've thrown a billions that, you know, the companies have thrown billions and billions at it and recents, uh, you know, snippets between, you know, the, the fat, the fight of G chat, GPT and deep seek and all that kinda stuff have thrown us all into a little bit of a loop showing that, you know, what we're working on, the stuff that we're doing under the scenes, behind the scenes, under the hood, whatever you wanna call it, really does make a difference in what's going on in the world.

And it's, it's almost thankless in some ways because we don't necessarily expect the recognition, but real, we can see what we're doing as having a major impact. You bet. I, we're working with some medical device companies that build CT scan machines and MRI machines and things like that.

And one of their gating factors is how quickly can they take the information that's generated from these, uh, scanners? How do they get that down onto, you know, media where it's stored securely? How do they get it done quickly?

If we can double the speed that machine can handle twice as many patients, you can help more people, the clinic can pay for that machine more quickly 'cause they cost millions of dollars. And the real thing that we keep seeing over and over in, in our market, and it's not just storage. Anytime you put something faster, any time you improve the performance of one component in a system, you rarely get the full performance of that component because it just exposes a bottleneck somewhere else.

Um, I kind of think of it as, you know, you put a bigger, faster engine in your car or supercharger on your car engine and then you find out your brakes are really bad, so now you gotta go upgrade your brakes so you can stop the thing, or you need a better transmission to handle the torque. You can't just make one thing faster and expect the whole system to be faster. And so, you know, we're specifically focused on protecting data.

That's our a number one. You know, your data's more important than anything, but if we can do it and get outta the way and allow that data to flow and allow you to achieve higher, you know, performance, then that's even better. Yeah, that's actually a, a key point there too.

We've spent so much time talking about performance protection is critical as well, and in, in, unfortunately what I've seen is many, um, use cases, especially in edge use, unprotected storage instead of, and, and, and that's a big risk. You know, that's a big challenge, you know, there's constraints. Sure.

Uh, you know, you, you may not have as many drive bays, you may not have want to invest in raid cards. You may have been, uh, burned by software rate or whatever, you know, uh, that didn't work and didn't perform as well as you had hoped because it was, you know, reliant on old technology and CPU cores and things like that. Uh, but with a lot of those problems are being, are being solved, you know, we're kind of in a new, in a new world, you know, flash and SSD based computers are kind of like electric cars in a way.

You know, it's, it's sort of a, a completely different paradigm of, um, of the same, the same thing, but you have to treat it somewhat differently. And, and when it comes to data protection and, uh, and storage, you know, as Scott mentioned, there's new form factors that will allow you to pack multiple drives into a compact form. There's, uh, you know, new technologies that would allow you to, uh, very, very efficiently, uh, serve data and, and also protect data in those environments.

So I could see this as really a transformative technology while maintaining that sort of compact and less expensive form factor. Right? Yeah, absolutely.

Yeah, I mean, to your point, we did kind of gloss over to some aspect. I mean, raid by definition is of course, uh, a protection scheme, but for those that may not be as familiar with it or the implementation challenges, hopefully through the course of the conversation, Kelly's done a great job of explaining a lot of that to us too as well. But, uh, data protection is absolutely paramount.

There's things that you have to do at all the different levels to, you know, preclude all the wonderful things we're hearing about cybersecurity problems and whatnot, let alone just data integrity problems for the base hardware. So it, it is gonna be an interesting ride for the, the foreseeable future, and we're excited about the partnerships that exist with, uh, with grade and, and a and the whole solution stack that we're working together on. So Yeah, you know, we've, we've talked about a lot of the high performance features we're adding like GP direct stores, but there is a slew of data protection features that we've added, like, um, right, journaling.

So, um, in a right, right in a rate degradation mode where you're rebuilding a drive, we now journal every right. That way if you have another failure, you eliminate, uh, a corner case known as right hole, uh, which call comes from a double fault where it's unrecoverable. Uh, we now have, we can always go back to the journal and say, this is where we can pick up where we left off.

'cause we know that was written. Um, there are transient errors that can happen across PCIE and even in these drives, you could have something called an unrecoverable error and A URE happens when something's changed in the data and the data's now no longer what you wrote. Uh, maybe a gamma ray hit that cell in the flash and it's flipped a bit, you know, okay, if you are running raid zero, you are having a bad day now because you have a URE and it can't be recovered from, well, we can go reach across the drives with the striped parody that we have and recreate that on the fly.

So we have now built a unrecoverable, uh, URE recovery on the fly as the application doesn't even know what happened. Um, and we just rewrite it somewhere else and then pass it on to the application. We have customers that they believe they can run read zero because they're looking at just MTBF, um, and drive right per day specs from a drive manufacturer.

And those are really, really good. But when you have 20 of those, you increase your failure potential by 20. But a lot of things can happen that aren't the drive's fault.

Um, many of my customers that are doing edge type things, we're talking about military battlefield installations, we're talking about harsh environments like, um, ships, you know, mapping the sea floor like data collection in an airplane. And drives can fail in that scenario through no fault of their own. But it could be because of, um, EMI interference.

It could be because of heat, it could be because of vibration, it could be because of corrosion and assault environment. So you've gotta take that into account as well, because, um, you know, the drive specs are fantastic if you're in a hermetically sealed data center that's at the right temperature all the time. But that's not always the case.

So, uh, if you don't protect your data, you're just asking for, you know, asking for a problem at some point. Yeah. That, that's a great way to kind of segue us right into, you know, what's coming, coming soon in the, in the rest of the series as well around as we start focusing on AI and Edge and getting further away from those, you know, perfectly cleansed environments that a lot of these systems sit into.

So I really want to thank you a lot, Kelly, for taking the time to join us, and I'll hand it back over to Steven here to kinda wrap us up. Yeah, thank you so much. Um, and that is exactly what we're gonna be talking about.

That's actually, that's why I love Edge Field Day and all these edge companies. You know, it, it's, in a way, it's kinda like the real world. Uh, you know, it's when things get real, you know, it, it, it is, it is all well and good to, to build these systems in an environment that will never face challenges.

But, you know, edge by definition is systems that are not in the conventional spaces. They're not in data centers, they're not in cloud, in the cloud, they are everywhere. And they could be, uh, under the friars, they could be on a battleship, they could be in the back of a self-driving car, they could be on a, on a oil, uh, exploration, uh, site out somewhere out in the middle of nowhere.

And in, in all of those cases, um, you know, we have to think about all the things that can go wrong. And in all of those places, AI is making a tremendous impact on how people are, you know, what people expect and how people are using these, this data, these sensors are, are, are pulling in way more information than ever before. They're processing more information than ever before.

And that demands better and better capabilities in terms of, of GPU Sure. But in terms of storage especially, and that's really what we're gonna be talking about all this season on, on utilizing tech. Thank you Kelly, so much for joining us today here on, on Utilizing Tech.

As we wrap up the episode, where can people connect with you and continue this conversation? Absolutely. com is our website.

We have a number of, uh, parts of the website where you can look at different use cases, uh, customer case studies, et cetera. Um, we also go to lots of the big shows. I we mentioned GTC, we're gonna be at Cloud Fest, uh, over in Germany.

We're gonna be at the Super Compute show over there called ISC. We're gonna be at SC 25, which is in St. Louis this year.

I'll personally be at NAB, um, I'll be at Dell Tech world. So we're, uh, we're gonna be at a number of different, uh, events like that. And, uh, would love to re you know, connect with anybody who, uh, is gonna be there and would like to speak with us.

Just reach out to me on my LinkedIn, uh, page and I'd love to talk. Well, it's great to have you. How about you, Scott?

Uh, where can people catch up with soy? com/ai, uh, you'll find all kinds of cool stuff there related to what we're doing across the entire AI pipeline. Myself, I'm SM shaley at on the different social platforms.

And Scott Shaley, of course on LinkedIn. So happy to chat with anyone about what's going on in the world as uh, we start to move this stuff forward. And so I will also be kind of tramps in the world with Cloud Fest and, uh, data Center World and Dell World and sc all the, all the fun shows where we can kind of showcase some of the really new innovative stuff we're doing together with all these great partners.

And as for me, uh, you'll catch me, uh, most Tuesdays on Techstrong Gang. Most Wednesdays on the Gestalt it rundown. And of course, on the Tech Field day and utilizing tech podcasts, uh, look for Stephen Foskett.

Thanks for listening to this episode of the Utilizing Tech podcast. You can find this podcast in your favorite application as well as on YouTube if you wanna see what we look like. If you enjoyed this discussion, please consider leaving us a rating or a review.

We would love to hear from you. This podcast is brought to you by Soy and by Tech Field Day, part of the RUM Group. com, where you'll also find the previous season, uh, focusing on storage, uh, with soy.

You can also find us on x Twitter, uh, mastodon and blue sky at utilizing Tech. Thanks for listening, and we will see you next week. Alright.