The Open Flash Platform Initiative with Hammerspace

Hi, my name's Kurt Kine with Hammer Space, and, um, I'll be talking to you about the open Flash platform initiative. So, um, what Molly and Floyd talked about, right, was really core product capabilities of Hammer space, um, and operating on tier zero, right? The performance, um, of the product, uh, the performance of the AI workloads really highlighted there.

Um, but beyond performance, right? We all know there's this really impending existential problem around data and data centers and all of that in terms of power Sure. In terms of pooling and in terms of the fact that we're just trying to cram so much data into what is a, a, a finite space, right?

Um, and so real estate is a challenge, and we're starting to generate more and more data out at the edge, right? All of our phones are getting more powerful. All of the devices that are out there, all of the sensors are getting more and more powerful, more and more granular.

So all this additional information that we're trying to ingest, um, and a lot of the focus of AI so far has been compute centric, right? We're focused on which flavor of GPUs is most recent and how powerful those are. Um, but what of the data estate, the challenge that we've got with getting all of the data into the data center and then being able to store it there efficiently.

Um, and so with that, right? We saw, we see this need for really, really large capacity of storage that essentially disappears into the data center. And I'll tell you how we do that.

So, you know, there's a few components that make the open flash platform that we're proposing, um, possible today, right? Well, we've seen developments in, so the flat, large capacity flash, right? With the adoption of and wide availability of QLC.

We've got flash that's really, really dense, um, and can supply that cost optimized tier in a flash platform. Um, DPU have essentially gotten powerful enough that they can serve as a replacement for those CPU based storage servers out there, right? And when you look at storage servers out there in the environment today, right?

That's how essentially every storage platform out there operates, right? You put in a storage server and put a bunch of disks behind it. Those storage servers are very expensive, both from a procurement standpoint, right?

You've got CPUs, you've got memory, you've got all the infrastructure around it. And then also from an operating standpoint, you've gotta power those things. You've gotta cool those.

So these dpu coming very, becoming very powerful, are allowing us to essentially replace all of that server infrastructure. And then we've already gone over, right? Floyd and Molly, um, a lot of the enhancements that have been done to the Linux fer to supply, um, data services.

So to, are you Talking like of a, of a jab b That's, that's actually a Linux server with DPU and lots of QLC flash. I'm, I'm trying to figure out what right. Called Flash.

Yep. So I'm gonna jump ahead just because, uh, my friends have given me, uh, limited amount of time here and go through some, you know, jump over some of the laying the groundwork and really just jump into, okay, what is this thing, right? So number one, what is this thing?

The first thing is this is an industry initiative. It's not a hammer space only thing, right? Mm-hmm.

We wanna make sure that people understand. We want to continue with this. We talked a lot about the standard based, um, approach.

Um, and so this is a standard based approach. We've got other industry players I'll go over, um, who's already in the fold, um, and a lot, there's a lot of momentum and interest as well. But essentially what we want to create is a platform that allows us, uh, our goal is basically an be able to deploy an exabyte in Iraq.

We wouldn't necessarily deploy it that way, right? You might just deploy some smaller, um, amount within each server rack or something like that. But these, this is, here's the basic proposal, right?

Is we design a sled that then fits into, um, a tray That then can be deployed within Iraq. So how does this change the game fundamental? Um, it's from a density and power standpoint, right?

So this image here is essentially the entire server, right? This replaces an entire storage server. You've got a DPU, you've got your networking and you've got your storage really, really lightweight in terms of power and cooling requirements.

A really lightweight server in there on top of the DPU. Um, and then you've packaged your NVME in there very, very tightly because you don't have same pooling requirements that you have with the full server infrastructure. And so essentially this is the little one use sled that fits into these Larger trays, right?

And you can deploy them, you know, modularly start with two and grow as needed, but this provides your capacity layer. Now, what we're about to show to the market within about a month or so is a really initial concept, right? Which is this using U2 drives within this platform.

Um, we can get to a fairly reasonably dense position, um, but it's not getting us to our complete goal. What we'd like to do is actually, um, extend the E two standard that would fully allow us to pack even more capacity into these, these, um, pledges. I was gonna ask why you were using U2 drives, but yeah.

Mm-hmm. Okay. What's available today, um, and allows us to really show a good proof of concept to the market so that we can get everyone interested in where we want to go.

Okay? And so, um, this is really what we see as the ultimate result of the open flash platform initiative, right? Why Does the, is adoption something?

Why does the power not go up? You quadrupled the density, uh, and the power stayed the same, Just the different form factors. Yeah.

In Terms essentially different power. Yeah. Uh, power distribution and form factor.

Um, so, you know, the idle requirement mm-hmm. Right? Of each drive is pretty nominal.

And so you're not driving up, you know, that, uh, maybe if you're hitting them really heavily, you might get an increase, um, in our requirements, but generally, um, it's not going to be, um, much there, right? So, so what is Hammer Space's role? You're leading the consortium.

We're Really advocating for, For this, but it's gonna actually build this. You're, We have, uh, contract manufacturers that are engaged with us today. So you'll then sell them yourselves.

Uh, We will, um, sell them ourselves, but we're also saying that this is not, yeah. We don't, again, want to another proprietor. Other people have packaged flash in interesting ways, right?

Pure is out there, um, doing it themselves. But we want something that is non-proprietary, that's widely available, um, to the industry. We've also talked to, you know, folks like the hyperscalers.

Um, we see this potentially as an OCP adopted. We haven't gone through all of that process that needs to be there. Um, but we would eventually like to see it as part of, um, the OCP initiatives as well.

Um, and Our founder was the founder of Fusion io. Um, lots of, Yeah. Our founder, David Flynn was the founder of Fusion io.

So he's, he does have a lot of experience in building this kind of technology. What we're bringing is the initial funding of the contract manufacturer to pull it all together. Mm-hmm.

Um, of course all the Linux contributions, so you can run Linux and take advantage of the DPU. And then we're pulling the group together, the DPU partners, the SSD partners. Um, if you look at the website, you can see who they are.

Right now there's a bunch more jumping in. Um, so we're kind of bringing folks together, but with ultimate goal that most likely OCP will take it over. And then this will be a system which we're particularly, um, optimized for because we're used to using Linux agents that are built into Linux.

If another storage company wanted to use it, they'd still need to have their storage servers in front of it so they don't get the same efficiency of power usage and density. 'cause they have to add the additional storage servers, which we don't. So we're, that's the motivation behind it.

Okay. And ultimately that these can run in hot data centers, low power environments, which is where we see AI infrastructure needing to go in the next year or two. Great.

Uh, does it have the high availability, I mean, dual server processing and stuff like that? I mean, I dual ported SSDs, I mean all that stuff, is that all part of this pack? Oh, What we're seeing is building kind of redundancy in much more at the higher levels.

At at, at the next level. Yeah. At the software levels.

Um, and that way you keep these systems as simple and as cheap as possible. Um, and in fact, right, if you get it to such a place, what, what's the reason why we need all of this, uh, these replaceable components within a server? Uh, and you're constantly servicing your servers because they're very expensive, right?

And essentially, if you get to this level of cost, you can even just fail in place, right? At the, um, at the N level, right? Just leave it in there and run with just a little bit less, uh, n capacity within each unit.

It's not gonna be something you're gonna want to be constantly servicing. And so that drives down overhead. What advantage do you see this over on ES E-D-S-F-F, uh, form factor for, um, SSDs?

Um, I, I think the, the biggest thing that, um, you know, we see is again, just how tightly you can pack these things in. Okay. But I mean, that was the point of the EEDS, that's about the ESFF form factors is that you could, you could like, I think they can fit 16 of them into one e chassis At this point.

Mm-hmm. So yeah. But now we're talking about, I mean, many factors larger than 16, right?

This is a, Yeah, this is a capacity play and this is all QLC, right? So it's re heavy. So, well no, I'm sorry, you're saying pet, you're, so 24 petabyte capacity assumes QLC, it Assumes QLC.

So you can have different, have to be An interesting enough, while we think QLC is the right form factor or right. For this form factor, right. For capacity workloads.

Actually some of our first customers that are talking to us about it are looking at putting QLC. Yeah. Okay.

It's also usage. I mean QLC is much more for read heavy. Exactly.

Applications. Exactly. So when you're looking at these kinds of capacities, right?

You're talking about probably archival Yeah. But online archive again. Yeah.

Right. Um, in the AI workloads, um, your Streaming also much as, yeah, Right. Your streaming thing.

So the performance out of the system too, right? While each individual, um, of these sleds is fairly good, right? Um, performance, the aggregate performance that you're getting out of the system, and I'll do a pretty quick comparison here, right?

Um, of, you know, so lemme go back to, right? So here's what we think of a, you know, a modern actually using flash capacities, um, that'll be available in early 2026, right? As the comparison here.

Um, we can take, you know, a really nice all flash scale out NAS platform, right? And reduce that footprint from a full rack into essentially two U of one use phrase, right? And so this is really an order of magnitude decrease, uh, in the amount of space power that's required for this massive.

So, and, and the dus you're using are just, you know, industry standard GPUs, your depending on your, um, members and, and the Exactly. So what again, right? We don't wanna look at just one gpu.

Yes, of course. For our first POC, we're working with a partner who's, um, really eager to get this to market as well. Um, but we've looked at other GPUs, um, as well, and they've been just as functional and interesting.

Um, so again, we think this form factor is something we would like to see adopted much broader than just hammer space. Um, right. And, and talking to some of the hyperscalers that are interested, right?

That's where we can also continue to drive down costs by getting volume right. We want to be able to get volume, uh, through this form factor as well to drive down the cost. Mm-hmm.

The comparison you have has got metadata servers and storage servers in this configuration, correct. Versus just storage servers. So if we're Being, if we're being totally honest, uh, honest, yes, we would require some, uh, hammer space servers within that rack and we should put those in there.

Um, two U and that's gonna take up another two U. Can the anvil be a, uh, one of these units? I mean, is that We, uh, so the, um, DPU that we are using today, it does look possible.

Um, we're doing some engineering to see if that would work and that would be a panacea for us for sure. Um, right now we're probably taking a little bit too much resources up, um, to be able to fit it in that, but I think ultimately that is a goal as well. So Amva wouldn't need that much capacity.

No, it wouldn't need that much capacity. That's correct. Um, but you know, if you can bring it down to just be a component within that form factor, it makes sense.

Um, it's just something that we're not doing initially that really there for the capacity. So What all do you offer in the way of hardware? Um, so actually Floyd, do you want to talk a little bit about, um, kind of our hardware offering today?

And again, right. This is an initiative. We're still in early stages, um, and we're getting more and more partners on board.

We're expecting at the, um, next OCP show to show our Working prototype, Our working prototype. Um, and we actually plan to ship early systems to customers within this year. Um, but I would say wider availability systems, right?

For folks who are interested in this as in 2026, And again, you know, hammer Space is still a software company and our primary motion is software, but as Molly mentioned earlier, with the AI ready nodes, um, and Hammer space, what we call Hammer space Ready nodes, you know, we are working with hardware partners. This would be part of that. So we will not have, and maybe not initially, um, a hardware sku, you know, that is, you know, that is, we are a manufacturer of, although we may have branded Hammer space Ready nodes, both for on the OFP side for a capacity layer, um, and on the, uh, on the compute side for Anvil and DSX.

And to your question earlier, it wouldn't really be the anvil that would be embedded in this, it would be a containerized DSX, right? So the DSX would like today, I understand that that would be the storage server side of things, but you still need to have Anvil and a Hammer space solution at some point. Sure, Sure.

And by containerizing both the DSX and the anvil, and I'm getting into roadmap that we're not really supposed to be talking about, but by containerizing these things, it just makes them much more easily deployable. Mm-hmm. Think especially in the cloud, you know, as you have scale out anvils containerized so that you can go as wide as you need and get that parallel io So you're, you're seeing a theme here, build into a standard on the storage side, enhance it with our investment, we're, we're, we're funding the prototyping, the initial, uh, cms, all of this, just like we have funded all of the work that we've pushed into the Linux community, so that standards on the storage side, standards on the client side, and that, and then building that same architecture, that Hammer space is uniquely able to activate into Hammer space to do it, but do it in a way that is non-proprietary.

Um, here, just here are the early members, right, of, um, the initiative. Um, we are building our early prototype with the, um, xite PPU. Um, it's a really interesting and really powerful, um, PPU and so suggest looking into them, um, you know, for, for more information around that, um, along with a number of, um, of the Flash players.

And we have many more coming on board, uh, here in the, not just future. Um, so I think really to summarize, you know, a lot of what, um, we were talking about today, um, is, you know, one of the things that Hammer Space really wants to emphasize, um, is that, you know, we've, we're following this open highways versus walled garden model where, you know, it doesn't make sense to us to say to the customers, Hey, get rid of all of your data silos by moving all of your data into a data silo. Mm-hmm.

Right? Um, what makes more sense is leave your data where it is today, but figure out how to manage it centrally. Make sure you can modernize your storage systems, right?

Without having to necessarily buy additional storage systems. You can utilize that flash that's available in Tier Zero. You can utilize your infrastructure that you have today.

And when you need to expand, then we've got these optimized platforms for capacity, right? Um, but you're not having to invest in a completely new infrastructure environment to take advantage of Hammer space and to get the data acceleration that we provide, as well as the span of control over your data, where it lives, where it needs to get to, um, whether that's on-prem or in a multi-cloud environment. Um, and that everything should be based on this standards based approach.

Um, rather than having to put your stuff into a proprietary file system that then also introduces this need for these really, really power and, um, pooling hungry source server. And so, we'll, we, we, yeah, definitely think this is the right approach, right? The modern AI environment.