Warm Flash for AI Context Storage using the Open Flash Platform with Xsight Labs and Hammerspace

Hi, I'm Ted Weatherford. I'm VP of Business Development at Xite Labs, and joined with me is my partner. Kurt, please introduce yourself.

Hi, I'm Kurt Kine. I am senior director of AI marketing for Hammer Space. Awesome.

org. And, uh, this, uh, is a group of like-minded people that want to bring a new level of efficiency and performance into warm flash storage. And it couldn't be a better time to do that, uh, with the context, uh, type tiering that's coming up.

org is the kind of performance gains at a system level that you can get by adopting, uh, this approach, uh, of the org. And you're looking at 50% less total cost of ownership, whether it's coming from power in the OPEX side, or you're just weeding out all the X 86 servers that are traditionally part of a large, uh, more centrally managed storage system. Um, so many things we do, uh, there, uh, the, the heart of it is really the software and that's supplied by Hammer space.

And correct me if I say anything wrong, but my understanding is they've built a Linux-based NFS standard file system that they can manage in a distributed fashion, uh, such that you decentralize and you can pull all the X 86 out. And by doing that, you're going to reduce tons of costs, tons of power, tons of complexity, and even the, the state machine of the hops that you have to go. So you go from maybe five hops down to two.

Um, what it physically looks like is, is a cartridge like this that our partner Lumineer, uh, did, um, where in one end you have two by 400 gig ethernet connectivity for, you know, NVME type style fabric. And then our DPU, which we call the E one, the E-Series that goes here, it's 800 gig or two by 400 on the front, and then eight different flash drives. Uh, and it's its own contained endpoint server storage target.

Right. And then in the back you just supply power. And I think there's a control ethernet too.

So it's really a server on a cartridge we call it. And this is showing on the left with the red letters font, uh, the old way of doing it. Yeah.

And so I think the, the big thing, the big evolution that we see is that we can really take out the excess hops of a, you know, traditional storage system. So I'll pass this around here. Yeah, you guys have already looked at it, but if there's anything else in there you want to check out, um, you know, typically you've got a storage server that holds both metadata and data.

Um, well, with this new approach, we separate the metadata from the data path so that the data path is going straight, um, to a storage target, CA standard NFS, um, and you handle metadata, um, out of band in a parallel fashion so that you can get really high performance or really large clusters, um, while also getting the data on and off the media very, very quickly. Yeah. Well, well said.

And, um, I just wanted to show some of the renderings. We physically have the box we're passing around the cartridge, but this will just, you know, from a chip perspective, we were honored to partner, uh, with, with Hammer space. They designed with a partner, uh, lumineer Design house, uh, this wonderful system.

Uh, there's two versions. They're sort of the 23 inch six cartridges that's more of a open compute, meta centric, and then there's the 19 inch, you know, 17 inch usable that we really are driving, uh, everywhere. That's five cartridges.

Um, and this just gives you some of the wonderful density specs that, uh, that you get. You can get a terra, uh, uh, exabyte of, uh, memory inside a a seven foot rack. And, um, you know, I think, I know we are on infrastructure field day, but I do also wanna speak to just kind of the consumability Yeah.

Um, of the chip itself. Right. Um, we've gone from learning about the Xite ship to, you know, very close to production here in eight months.

And I think that speaks to just how easy it is for our engineers, um, with of course, XITE awesome support. Um, but right. We can run Linux full-blown regular Linux on this chip as essentially a storage server.

And for a long time we approached the open flash platform talking about it as nothing but Nick, right? Oh, you know, how great is it that you've got everything out of the way and all you really have is a nick connected to, um, flash storage on the backend, but actually it's way more powerful than that. And we're finding additional use cases where this is just, you know, essentially really a server that allows you to do all the things that a server does, but greatly simplifies all of the infrastructure that's in between your clients and your backend media.

And, um, a quick Question on, on these devices, do you actually have build any data protection into the individual blade? This would be a, a question for You. Yeah, so, um, no.

In fact, what we're looking to do is removing as many of the failure points as possible, right? So as we get rid of infrastructure, uh, you're getting rid of things that fail. Um, but what we're looking to espouse is really, again, it's at this scale, it's much more of a degrade in place, fail in place model where your consumable unit is essentially the blade.

And so you're using right, our erasure coating and things like that to build in the redundancy there. We don't necessarily see the redundancy needed to be built into, um, the blade itself. Are you worried about, um, are you worried about like, any kind of intrusion, like getting things BitLock or things like that?

I think, yeah, I mean, from a security standpoint, um, we're always looking at ways to better protect the customer's data. Um, I don't know if the, the form factor itself is necessarily, uh, Yeah. Traditionally just having the data someplace else, and it seems like you're, you're doing that with airspace, making the data go someplace else, so it just seems like everything's all together and being able to restore it quickly is that, 'cause it's kind of out of the range and that would have to be an add-on, Um, just from an, so, you know, if each individual, right, each individual blade is its own server and we are erasure coating across those blades as well as within the server itself, right.

We definitely have protection across multiple units. I think just within the unit having a protection scenario isn't necessarily right. Something we consider, Just to be clear, the device is not, uh, implemented to be able to, uh, field replace a particular SSD or a U2 unit in that sled.

So you've got softwares to, to sort of ignore the fact that it's there and, and recover from it and all that stuff. Yeah. But yeah.

Yeah. You can do it software, you Can replace sled Yep. But you're not gonna replace a specific SSD Yeah.

A specific SSD within that. Exactly. Right.

The SSDs are lasting a lot longer. Their time in life is actually quite high. Used to be the case that the server and the flash, the flash would fail earlier.

We're past that now. I mean, flash has come of age, so this is, yeah. This is efficient.

Mm-hmm. Yeah. Efficiency's the main trade off that we're going for.

Um, yeah. And I think, um, everything about replaceability and serviceability, it comes at a cost right. Of complexity.

Yeah. Yeah. And so as we strip those things out, um, I think it's just a mind shift that we have to make.

And you mentioned the data center warm flash, you wanna Classify that, you know, this is, this is a buzzword I'm hearing as I'm running around with all the customers and, and there's sort of a saying going around, which is AI doesn't have any cold storage. And, and so the, the warm, the warm flash comes from that, that same I know zeitgeist that, uh, that's going around. It just means that you're ready, you're on standby.

And in this case, a lot of time you're, you're reading, uh, depending on which application. So, um, it's just a kind of common buzz. Yeah.

And I also think, um, you know, despite some of the flash constraints that we're seeing these days, and maybe the price of flash isn't going down as quickly as we had anticipated even just a year ago, um, there are still customers who are talking about the fact that they don't wanna have disc drives in their environment and they are looking at archival storage that is on flash. Um, and so I think that's another piece of kind of the warm, um, aspect of this, right. Especially as you get denser and denser.

And that was another piece I wanted to talk about Yeah. Is that the E one chip is such a perfect balance for us in terms of performance. Even when we get to what we have here is our advanced form factor, right?

Where we actually have, um, these blades of flash within the, um, sled and, you know, then we're getting to really, really large flash sizes for each blade. There's eight of them in the chassis 800 gig. Having that a hundred gig per blade is really balanced.

You don't wanna overprovision right. The performance for any drive, but you also wanna make sure you can extract all of the data on that drive as quickly as possible when needed. So the E one chip is such just a great balance for the throughput that we need where we're not over provisioning, but we're giving them all the access to that device that they need.

Yeah. Thank you so much for partnering with us and, um, just appreciate it. It's been a, a wonderful, and there's been a very fabulous uptake.

We're really getting an amazing amount of interest, and we'll have the product soon, um, production by the end of the year, we hope. And, uh, actually some, the first waver coming in a couple weeks. Yeah.

It, it'll be a lot of software, uh, like it always Is. Yep. And, uh, it's been really fun and really rewarding working with you guys.

Thanks for, yeah. We're excited to play. Yeah.

Thank you. Thanks so much. Thanks.

Yeah.