Unifying AI Enterprise Data into a Single Instantly Accessible Global Namespace with Hammerspace

So, my name's Kurt Kine. Uh, I'm with Hammer Space. I'm the senior director of AI Marketing, and with me today, Sam.

Hi, Sam Newham. I run our AI solutions practice here at Hammer Space. Thanks for joining us today.

And today we'll be talking about our AI data platform solution and how we integrate with, uh, NVIDIA's reference design. Thanks, Kurt. You know, part of this, I think it's always important to, to set a little bit at the stage, right?

And, and what we see in talking to customers every day, I assume the same way you experiencing your business work, whereas you talk to customers as well, but the, the data fragmentation problem's real, right? I think we've heard it on 17 billion podcasts at this point, is the data readiness is still a massive inhibitor to ai. The other challenge is the tools, right?

We saw an article this week that there's a major bank that kind of put up a fuss that, hey, this AI stuff isn't easy, right? These factories aren't as simple as just plugging in data and power, but there's a lot of endpoints, tooling, complexity in these workflows. And there's also a massive capital outlay that comes with this, right?

Most of our customers have fomo. They, they fear missing out on AI affecting their business. They have fo move as well, right?

A fear of messing up because we're, we're talking about massive spins that have to deliver on value. In short, you know, you've heard a lot now about Hammer space and in the, the things that make our platform unique when it comes to global namespace and Tier Zero, what we did is choose to package and extend this, right? Thinking about how do we really solve that data challenge for enterprises?

And so the idea being right is we take the hammer space software package that along with best of breed Nvidia software, right? As we look at their rag blueprints, the RTX 6,000 Pro, and as we mentioned, a new NCP server that we're developing that really helps stitch this thing together. We're allowing customers now to kind of buy an outcome, not, not a toolkit, right?

It, it doesn't need to be this recipe of things that have to come together. We're reducing all the tooling that happens so that a storage admin can aim this particular thing at existing data and get AI ready data within minutes of it scanning and seeing that particular file system. And it becomes extremely cost effective.

We see most of our customers not building, you know, thousands and thousands of GPUs, but they're doing project based ai. It, it may be a department, it may be a certain POV or use case. And so they're trying to find ways to, to basically start their AI factory with, with what they have today.

And so the idea behind this particular AI data platform, I is removing all that. They don't have to think about NV ingest and how do I create these NIMS and toolkits and how do I monitor and blah blah, all of this data. But we really want to give enterprises, uh, some freedom to take data where it lives today, move it through this AI process, and get to the point where applications and agents can use that data consistently and frequently.

The idea being is that we, we deploy this integrated appliance, right? We'll talk about how this is sitting on top of Cisco super micro bon novo. The great part about being software defined is I can run the same thing in the cloud as well.

The beauties. I'm not tied to a hardware platform or some bespoke hardware. I then connect it to my enterprise hybrid data.

As you heard Kurt talk about. We can do this through assimilation, that's even a full assimilation. Or we can do something called read only assimilation.

I can take snapshots of those other file systems and process, which is really important for Rack. We think about it, there's a portion of data that changes constantly, but as they suck in their first petabyte of data, that data's been sitting there for 2, 3, 4 years, right? We're scanning all the PDFs and product documentation support log, right?

And so a lot of this says why, why would I copy that to new shiny expensive flash when I could use that where it is and just accelerate it? And then we're able to automate these embeddings of pipelines. We're gonna talk about how we do this a little bit, but the idea is that you don't wanna have to understand re-ran what this index engine is and stuff, what we could package all that and obscure it.

It makes this much simpler and the time to value increases dramatically for our clients. So let's talk for just a minute about how we do this. What's the magic under covers?

And so the idea is that, you know, we can point at some other storage. It could be our storage, it could be object, it could be something. But as new files arrive, we discover those, right?

This is the beauty of metadata tagging and how we do things. And so we allow storage for these files to be economical. Really think there's two tiers of storage.

It needs to be as performant as possible when you use it and as economic as possible when not like, as we mentioned about the flash crunch, it's gonna be very expensive to keep stagnant data on very prime flash tiers as we think about how much new data is being created in the organization every day. So as those files are tagged, we move them to tier zero, right? This is a background copy motion, right?

That we can do that. We're all gonna do the heavy processing, right? As we, you know, go through the extraction process and here are all these PDFs apart and images and tables.

That's all gonna happen on our tier zero, right? So we're not burdening the existing systems that were designed for economic storage, our MCP servers, then the connective tissue, right? Between everything we can do as a platform, the NVIDIA stack and the agents that need to communicate with that information.

And so as soon as those files land, we instantly embed them, right? And it's not that we just run a script and hope it works. We do true validated output.

We can do test cases, we can make sure all this stuff, if it got, you know, if there was an error exposed, we retry those files automatically. Was it a locking problem? Was it a performance problem?

Those type of things. And that is now fully ready for our agents and things to consume. We then automatically clean up after ourselves, right?

We, we don't see data scientists do that very often. And then we have AI ready data. So again, simplistic diagram of a lot of complex things that happen underneath.

But this is the idea of being able to really reach into file systems. The concept of A IDP says why have to push all this stuff to a new location, to a new thing? It's already in storage, storage knows when a new file's created.

We should be able to do all this embedding and processing in the background, not as this other expensive human task on top of those platforms. Uh, so it's Ray ese is the tier zero actually running in the NVIDIA server? So, great question Flash.

So it's actually, so we are running on standard servers, right? This example is, is a Cisco blade solution to my point about starting small, I can start this in a single chassis. So Tier zero is using the NVME sitting on a C two 10 server.

We've got RTX pros in this particular scenario. And so we've stacked our entire software, right? The Hammer space software, the NVIDIA software, all into a single appliance that's fully combined, scalable and manageable in a single footprint.

So do you actually lay out a file system on the NVME flash or do you, do you use it as raw lock Data? No, we actually lay out a file system. So as Kurt talked about Tier zero earlier, which we can dive into, we can actually take that flash in all of those GPU servers, whether it's a cluster like this or a very wide large cluster, and create a full tier of file system at the NVME level inside the server, and then handle all the automatic tiering and positioning of that files as those particular jobs needed or as those files are called to use from, you know, a s LM scheduler or some other type of system.

What, what is the base file system that you lay down on the flash? So it, it's Linux, you know, basically XT file system, right? We, we are all standards based.

The, the beauty of the Hammer space platform, right? 2, those types of things. So we're using existing kernel resources to basically share those drives via NFS.

It all mounts through our Anil servers and gets represented to those servers is when ubiquitous file space, The, you had asked, you know, are we taking over the NVIDIA slash Yeah. And so in this case, if you're looking at a, um, AI factory, no, right? It's a self-contained solution here in the, um, Cisco server.

Um, in other cases, yes, we can take over the flash within the GPU um, cluster, yeah. And then serve up the, um, data directly from there. And there's even, you know, hooks that we've built in that allow the server to recognize whether that data is actually local to the GPU or on another server within that, um, within that cluster, and then be able to find the most lowest ency path.

The other question might be, um, NVIDIA offers, you know, RDME or NVS and other, other protocols to speed up data transfer. How does hammer space play in that environment? That's great.

We found that normally PNFS right is plenty fast to saturate the GPUs in these types of environments, especially for data processing. But even as you get into B two hundreds, B three hundreds, the, the new NVL systems, those types of things, PFS actually is incredibly performant for two reasons, right? One is because we're deeply embedded into the kernel, we can actually do things like NFS direct, right?

Where we can bypass the memory stack and be able to, to really feel like, plus local storage, right? To think about it to those GPUs, but we're also fully integrated into GDS and those types of things as well. So if we need to perform that type of technology, again, to us it's another protocol and method that we can serve data.

So, so I want a clarification on something. So is the, the tier zero MVME drive, is it then dedicated to hammer space or is it still shared With, in, in this it's dedicated to hammer space. We take over and own those drives, right?

So those aren't, you know, temporary space or scratch space or that sort of stuff within those systems anymore. We actually take over those drives and create a, think about almost a rated file system across those servers in those environments. Yeah.

And you already said something that was gonna be my follow up question. You said multiple drives. So tier zero, is it a minimum of one NVME drive?

Can it be more, is there like an ideal setup? Yeah, so, so again, this is the fun beauty of, of software defined storage, right? Is it we can take two drives, right?

That are normally populated in HGX systems, we can fully populate those systems with eight to 10 drives depending on the form factor of those things. So we're not limited, we can use as many drivers in that system. And this is what we find a lot of clients do, right?

Is they run out of space or they run out of power. And so while it's a speed play, in some cases it's actually economic play because now we're already consuming that Rackspace and that power footprint, we can continue just to add drives and create an entire fast layer of storage directly on those NVMB servers, right? With the durability, reliability, if there's nodes reboot or we take some sort of different maintenance window, there's some things we put in place from a policy perspective to perfect against that.

But yeah, it could be one drive or it could be 10 drives per server. We don't care. I guess my question would be what about boot and things of that nature where you don't actually have all your software deployed in the system, you must relegate some portion of those drives to, you know, OS or something like that?

Oh, a hundred percent. So most of these systems usually have a, a mirrored set of boot drives in the back, right? They've got four eighties, nine sixties.

And so that's kind of built into the platform itself, right? Is that boot in os flow lives on those mirrored drives, the drives in the front of the system are dedicated to data. So for the sake of time, I wanna push a little bit, and I I know we can continue to dive deep, right?

Any good seller wants the second meeting and so I, I think we're we're headed that direction. Um, what I wanna shift towards a little bit too is I know we're talking infrastructure, but the idea is that, you know, we're thinking about a data first strategy is that we've always thought about this from an infrastructure standpoint, right? It's what box does it live on, how do we manage it?

You know, permissions were almost tied to these silos. And then the data scientists in the rest of the world, the BI community right, comes from a a file context side. They're pushing from, Hey, I tag this and catalog it and half the data scientists I talk to don't, don't really understand their storage journal or they don't want to have to know about it.

Even in my career earlier, I did all SQL optimization, I learned raid backwards. I was like, what does this Dell think? What's a raid group?

Why is my tempdb slow? And so the idea of being is that there were always kind of two halves to this AI super brain. And so part of what we're doing in this data platform is not only taking the best of hammer space and the best of Nvidia, but you know, partnering with people like kui or Varonis or these other companies that are really good at, at data categorization and tagging, right?

We, we feel like this, the vision, right, that, that we're really trying to push out of there is that the security, the performance characteristics and stuff are tied to where the data lives, right? We've always dealt with where data is born and it's always been very heavy and very expensive to move. But you know, data gravity's always been a thing.

Now we feel like GPU gravity's a thing and it's really power gravity, right? It's like who's building the next nuclear reactor and who's gonna put the a hundred megawatt data center right beside it, right? It's a, so it's this idea of how do we eliminate the data gravity challenge.

There are times that we need to move it, we want to avoid that at all costs because it's expensive. There's ingress charges, egress charges, fiber's not cheap, right? But we think more about the high fidelity data pipelines, uh, how we evaluate quantity, right?

Over quality, the the proactive governance, right? And know Marian, you kind of asked these questions too. It's like, well what if that followed the data?

What if we could store that in metadata? And it wasn't that we relied on just the file system, but what if that was attached almost as an attach to that particular piece of data? And then this is really about time to value, right?

There, there is an AI race that's going on and so, you know, we've got some statistics to show how long it takes to cobble all this together in most enterprises. First buying a solution like this that really can get you up and running within weeks of hardware delivery being on site. Real quick, uh, I'll show you a little bit of the demo.

We talked about MCP, some of those types of things to, to me this is the glue, right? This is the Newent workflow. You know, we used to care about service accounts and all these different things and now we're talking about agent security and we've got father agents and mother agents and child agents and all sorts of things.

But what I wanna give you a preview of, right, is, is how we're trying to make this human interactable right, is we don't wanna have to think about NV ingest and pushing massive JSON files. But in this, what you see is we're actually asking this thing to build and execute a plan that uses the product launch update data bill of materials that's inside Kui, right? So this is a pre-validated data bill of materials that's been scanned for PII anonymized approved by the security team to be used that will validate ingest and clean up the data using A IDP cluster one because we realize that there may be more than one of these on-prem, in the cloud to a specific vector database and complete logging and everything, right?

And so, you know, on this interface, you see we're discovering files on the right hit side, but what this does real time is builds and executes a full plan. How many files do they need to be moved? Can I do it in place?

We're using metadata properties about performance and security to make intelligent decisions that usually would take a data scientist just saying, our sync this directory, let me go get coffee and I'll come back two hours later when that particular job's finished. So you can see we do everything from full testing and validation throughout this entire process. Everything we're doing, you know, I, I learned a new term, right?

We used to talk about eating our own dog food, uh, drinking our champagne sounds much more elegant. But this is what we're doing now is that we're tracking every microbe, millisecond move of data across the file system into these particular workflows. Fully diagrammed, because you know, what we found is most of our customers are using some script on one monitor, copying, pasting it on the other and hoping all this came together.

So for us it wasn't just about can we stitch it together, but as a platform, it's how do we agree? How do we integrate with third party security tools? How do we make sure we can diagnose and debug these workflows and credibly quickly?

And if we're gonna be that platform to own and manage everything, we have to expose all of this from a data standpoint. So I'll wrap here, we can talk through a few questions, but again, this idea is that we're bolting this onto existing systems. This isn't a buy new and migrate two, which we saw most of our competition pushing, right?

Copy it to my shiny new flash platform. And then we'll do the magic. We really will wanna help enterprises start where they are.

How do they build the AI factory with the existing capital investments they've made? Again, we talked about time to value, you know, below this Fortune 500 magic line. There's not all these AI centers of excellence and AI tiger teams and this sort of stuff.

It's John who's been running storage forever and two or three data scientists from marketing and product who are trying to stitch this together. And so we're trying to create a solution that we can add on to an existing Lex pod for existing, you know, Isilon stack or whatever we need to do to really help them take control of their data, reign in that chaos, and get AI ready data in weeks.