Industry First SSDs Tackle AI Challenges with Solidigm

I'm Scott Shaley, director of uh, leadership narrative. And right now I'm the evangelist. And as part of this, you know, we have to reintroduce ourselves.

I figured I was overdressed the first time we had ACE dressed down a little bit, so I thought I'd find the middle ground. And I also noticed that on 90% of the tech field days, at least one of your presenters is in a branded Hawaiian shirt. So there you go.

Um, so we've got roughly 10 minutes left to go through this next little section, but we're gonna take a little bit deeper dive now into what ACE was talking about, about the storage itself and how we build products around that. Like I gave you a whole litany of why we do what we do type of things. So right now we're gonna focus on the network attached.

This is the solid Im P 53 36, the 1 22 terabyte that's represented in your Lego kits. Um, I made a comment earlier that it's something about we don't always have to be first yet. My intro slide said industry first.

So yes, we can do it. It is okay for us to say that we are the first to do something, but one of the key parts of it is to make sure you're doing it with someone that knows how to take advantage of that product. So this is an example of our direct customer in the case of Dell, talking about the excitement they have for us to potentially introduce this product.

This was a conversation we had with them when we made the technology announcement about the availability of a product. We launched the 1 22 product in New York in Times Square, late November, had some fun, bought some billboards and all that kind of good stuff. But you always have to follow it up with something.

And so we'll get to where we're going with it from there and to get to the point of true customers if you will. Um, so one of the things about this is this concept of TCO that we were talking about. So we spent a whole bunch of time talking about how to use GPU clusters.

Uh, this is taking a look at, in this case, uh, 12 DGX platforms. And if you're using traditional storage and what we call the hybrid approach, which is hard drives plus SSDs for cash, the modeling shows that you need nine racks and 54 kilowatts to satisfy the hard drive and SSD combination for the capacity requirements for this particular 50 petabytes of effective storage. 7 kilowatts, same 50 petabytes.

Now it's not realistic to assume that you're only gonna do the shrink. So the idea here is you're also now freeing up eight racks of physical footprint to do whatever else you need to do with it. And it's kind of mind boggling to think we've talked gigabytes, terabytes, feet, inches, all that kind of fun stuff.

Um, this is a nine to one reduction, 90% network reduction. So I know I'm standing in our great friends Juniper networking. Again, we're not trying to belittle the value of their technology, we're trying to augment the support of it with future scalability.

Uh, and then you can basically, you could use 50% more servers as you develop this footprint out. So if I think about it, I could just literally cut paste drop and I'd still have leftover space. So, but at the end of the day, what does 1 22 terabytes look like?

So one of my favorite things to do growing up as a a nineties kid was to be, uh, babysat by my best friend, the television set and the movie theater. So how much data is 122 terabytes? We have to have some fun slides.

I changed shirts a little more lively. 6 times on that little drive that's sitting in front of you if it were real. Um, and so this is a just a snapshot of some of the, the wonderful movies we all all got to enjoy in the 1990s as part of that.

And just thinking about, you know, all those little servers we used to have of our not properly licensed streaming platforms we could put all of this information on, right? So when you talk about it, there's always the question how does it compare? So I I do have a one-to-one slide.

Here you go. Drives R two scale. So the physical size is relative as appropriate from a three and a half inch hard drive to a two and a half inch.

Um, SSD this is the point we were making earlier. At 122 terabytes with a gen four interface in a full rack server, we can guarantee that it has 24 7 reliability for five years. You cannot wear this drive out.

And if you wanna take the time gen four, how fast that bus is, the performance you can get at 100% rights and the amount of time it takes to fill a drive, it aggregate across the system that you're normally striped and all that good stuff, you will not be able to touch every cell enough times to wear it out in five years. So, and we've done that on purpose, we've architected the firmware to make sure that you can do that. And you also get the energy efficiency, the better.

Afr the things like the 2 million R-M-T-B-F, we've got all that quality data. But this is part going back to my what's next, and I know I'm not supposed to point at the slides 'cause it's all live stream, but the idea of, um, the footprint and the performance are so drastically different. We have opportunities there to fill that gap, right?

Because this represents the next best thing to take down, if you will, a hard drive. And I don't want to replace them. I have a desire to displace them in some places, right?

And so that's a, that's a good kind of lead into what could be next and what, what we have an opportunity to do. And when I showed you the, uh, drives earlier, the E three, this is an example of the 1 22 U2 physical board. So top bottom side side.

So there's two boards and they kind of, so I'm showing you all four potential sides of the drive and that gives you an idea. The, the brains is the controller in pink, the purple is representing all of the NAN devices. We have innovation in there.

So if you think about it, the dye is in terabits, the drive is in terabytes, there's a eight to one and all that good stuff in there. So we've got stacking going on, we've got all kinds of packaging innovation that takes place there. The drive design innovation, which is kind of represented by the the blue stripes.

There's signal integrity issues we have to worry about when we get this big. Just think about the idea of how you keep signal boosting in networking. We have an amazing network going on inside of this drive.

And then of course, uh, the firmware on the way of being able to maintain it, right? It's 122 terabytes, it's non-volatile. There is a cache, right?

We've got a whole bunch of DRAM on here 'cause we need more DRAM to satisfy the content. So we've optimized the firmware power loss protection. It is everything you want in a data center product with high reliability, high quality and that capacity and also that low price point dollar per gig price point.

Physical price point is 122. You can do the math. So we built it.

Dell was interested in it. Well just last week Dell announced it. So yes, we do have end customers, end end customers that are very much interested in this.

Our first partner to actually qualify this drive in their platform to resell to those customers formally announce their uh, power scale platform to be enabled by this 1 22 terabyte drive. So yes, it's an industry first we have a partner. We talked to the partner, we found a customer all from November to April is what it took to get this product from when we let the world know about it to when people can actually start consuming the product.

So it'll be officially shipping out of their doors end of May. And so this is a quote directly from their blog. The blog is linked in the side deck and uh, we can certainly make that available to you.

So I've got three minutes left for the fun part, which is the direct attached storage. So I'll do my best. Um, this is cooling down.

So this we took to GTC. This is a mockup. This is the idea of this little footprint right here is shown here in this non-video video, because we don't have non audio video, we were given a footprint of eight 15 millimeter E one Ss, which is what is in the brains of the direct attached storage of the Nvidia reference design.

And they said find a way to take the fans out of it. So the traditional way we do these is it fits eight of these guys with the fan with the air fins on it so they can cool it with fans. How do we translate this into a liquid cool design so we can take the fans out of the box.

So you'll see here we've got liquid, uh, in and out ports for each of the drives. Now when we look at it from the other side of it, we have a single cold plate touching the drive. So the idea of the fins as you put the fins on the hot side of the drive cool it off.

Unique part about this is in order to do that you have to have full contact generally on both sides for bold plate, all that kind of stuff. We designed the frame, the actual enclosure of the drive to allow single-sided cold plate attachment. And it's a spring loaded design so that it's full contact when you put it in, but still fully removable.

That's why there's the hook on the drive in production. It won't be a hook, it'll be as traditional latch. But for purposes of the demo we have spring loaded snap in drives and then now that cold plate is only touching one side of the drive.

But conductive active conductive cooling by way of the material and the way the frame was designed allows us to cool the top side the product as well. And then we also designed the cold plate itself to make sure that it wraps within the cold plate, the liquid flowing at the hotspot on the drive. How many terabytes in that?

These are eight each today moving to 16 in each drive. 'cause it's the E one and you can do eight of them. This is an example of the eight drives per enclosure.

TLC or The TLC or QLC options do exist. Okay. But this is direct attached.

So generally hitting the higher performance products. Yeah. Which are the TLC drives.

Okay. So, but I'll send that around. You guys are welcome to take a look at it.

I know we're running short on time, but the idea here is we're able to get a single-sided cold plate attached to a, uh, double-sided heat generating drive. And actually there you go. Cool.

Uh, it's, it's 3D printed, right? It's a demo. Mm-hmm.

Um, so the, the concept here, this is a reference architecture. So very much like the metro platform where we're making it open source, everybody can grab a hold of it. We are co releasing this and the IP associated to it from a licensing perspective through Nvidia for the new GB 300 platforms to be able to support those products with a fully liquid cooled system.

'cause traditionally today we say fully liquid cooled and they really do have fans on the storage 'cause we need hot swap ability. This solves the hot swap ability and goes fully liquid cooled. So hot swappable from the perspective of each one of those.

Yes. Could be swapped or the whole, not the whole unit. Obviously Each individual drive embedded in the system is still hot swappable, like a traditional air cooled system.

Wow. Yes. Okay, cool.

Because everybody still doesn't trust storage, right? We've got the AFR so low, we've done all this stuff, but drives fail, right? We know it's not really SSDs, it's hard drives, but this gives that flexibility and continued evolution.

So if we make something better, stronger, higher capacity, it allows you to still do all of those things you do with your traditional storage. Hot swap on the front of a box but be fully liquid cold. Alrighty.

So I made it at the 10, I made it within a couple of seconds, so mm-hmm. For the day, lots of opportunity, multiple ways AI can see storage used. We have a first with a customer.

Uh, we partnered with Nvidia. It will be part of a reference design for the next solution. And hopefully you've seen technology innovation never stops at solid.

I, we continue to find new ways to solve our customer's problems.