00:05
Greetings, and welcome to The Pure Report. I'm your host. Rob Ludeman, and it is time to bring orange, I'm super excited for our guest today, Hari Kannan, who has been with Pure for 11 years. Yes, to be on. No, it is great! We are long overdue to have you. We work every year around Accelerate and other things.
00:27
And you finally made it in the studio, and I'm super thrilled to have you here. What? Tell everyone what you do at Pure now. Over 11 years, you've probably done a bunch of different things. Yeah, I have. So, yeah, 11 years, going on 12. Um, I've spent all my time in the engineering product development land.
00:46
Um, so I started at Pure when we were deciding to start a new project. SoPure was founded with FlasherRay as a primary product. And then. A few years in, Cause and Dietz, who was the CEO at the time, and the board had an idea of kicking off a second program. Which has since evolved into FlashBlade.
01:06
Um, so as part of that, they decided to go start effectively what was then a startup. Within a startup. Yeah, it wasn't quite the same company. I mean, it was related or somewhat related. It was sort of an incubation. Um, a spin in some parlance, um. But aside from one of the founders, the founders,
01:26
So Cause and John Hayes, who's the other founder of Pure, There was very little other interaction with the parent company. So, we shared resources, as if on Pure's premises. Um, Pure's original investor Mike Spicer was our CEO then. Uh, some of us came on. But yeah, the intention was for this to be a
01:45
Separate group within Pure; not to distract. The other crack engineers we had on Flash Ray, sort of working towards building that Into a multi-billion-dollar business, but to be Able to incubate a new project. And that's where I was hired on. So I've been here ever since. It was like a startup within a startup, because we were still,
02:04
I think I joined just after FlashBlade actually became a thing. You know, was birthed and went to market And we're figuring out the different Positioning, because around that time, I was coming in and doing database marketing. It was like, OK, we have these cool blocks for Oracle and SQL Server storage things, But hey, does data warehousing, when you, know, throw that off,
02:25
Is there a play for that, and Flashblade? So it was really fun, Exciting times. What was it like in the early days? I mean, how many people were there? That you guys were initially working on FlashBlade, was it about 10 seconds? Was it hundreds? Like, what was it? It was in the low tens,
02:41
10s, yeah, so when I came on, I think there were maybe ten-ish people. Uh, and then over the next 6 to 9 months, we hired another 10–20 people. So we were; it's a fairly small team, let's say 20 to 30 core people who are incubating the Project, trying to figure out exactly what it is. So, when I came on, we were.
03:03
We had sort of just figured out we were going to build a scale-out product, and we're Still throwing ideas against the whiteboard. So, the following couple of years we’re super busy just trying to, We're running as a startup, right? So you're trying to find, Uh, incubate ideas, figure out the stakes, find product-market fit once you get in the
03:21
Process of building betas and all that stuff, uh, but under the auspices of the larger parent Company, right? So that afforded us a bunch of flexibilities. Terms of. Uh, latitude to, um, talk to others Customers get in relationships that grow, larger, pure, and had. So, it was really the best of both worlds. You get to run as a startup,
03:40
But you also enjoy some of the advantages of the bigger company. When it was really exciting, I recall too, because we are expanding into many use cases, Which is one of the reasons I'm having you on today, because one of those has evolved. Quite extensively. But I remember looking at one time we had This concept of a data hub, right? If you remember those days, it was like,
03:59
Well, this is a really great product for doing backup and restore. But there are also people doing many things with log analytics now. But also, there's this thing called AI that we were kind of scratching our heads about and weren't Really sure. What it was, or what it would be. Although, interestingly, we were doing a lot of stuff internally with all the telemetry data
04:19
That we're pulling back in. We were doing machine learning and cool things, Which I think we also, you know, eventually ran on FlashBlade. Which is, which is super cool. But it's, how did that expand out when you Guys looked at the different use cases and the market fit. Like, where was FlashBlade initially gaining traction at that time?
04:40
Yeah, so, FlashBlade was an unstructured data store. Right? And we had this, um, unstructured data store. The time was primarily file-based. Like, there was obviously object-oriented, But object at the time, and until recently, it was really thought of as the slow, Cheap and deep type of storage.
04:57
And one of our insights was that Performance Object made a ton of sense. So we were first in, in that regard of thinking of file as an object, sort of um. Being similar in terms of performance and capacity capabilities they introduce and Not necessarily segmenting them into files Equals fast, and object equals slow, equal Citizens are equal, right? So, the architecture was constructed from the
05:23
Ground up, in order to accommodate that. Now, if you look at most other competing offerings, people generally layer objects on top Of file, um. Right, because that's sort of a bolt-on, it's sort of, uh, an add-on mechanism. Which largely contributes to performance and other capabilities.
05:40
But since we thought of this as we pondered the future, and we, We saw the future of unstructured data grow, so we definitely made that bet. Right, and we bet on Object continuing to grow. So we architected from the ground up for files. And object to be co-citizens over there. Yeah, which was very different from some of the bolt-on things or maybe where it was just
05:58
Accepted that. This is just going to run more slowly. Or is, is going to, you know, be a tiered type of type of operation. And I think that’s part of the advantage we had in designing from the start with Flash. OK, we know what we can do with Flash on the block side. Well, what happens if we apply that to this unstructured space?
06:16
Aha, right? Same, same capabilities. And from the system design, I'd say this: Flash. Sort of concurrently at the time on FlashArray 2, we started going down the path of what has Since becoming Direct Flash. So, FlashBlade was the first product on which we started shipping it commercially.
06:33
In fact, fun nugget: back in the day, we had two different physical implementations. Direct flash. The logical construct was still the same. The idea that array-level software manages all of the flash bits and that if we were to hoist Those, uh, the flash management to the global operating system. We could make much better use of the flash.
06:54
The concept that ethos was the same across our products, But, In order to architect for maximum velocity, We actually had two different approaches. Also, at the time of FlashBlade, we were shooting for extreme density compared to that Existed on the market. That's right.
07:09
So, if you remember back at that time, we were probably shipping. One terabyte was a really large drive, and when FlashBlade came out, We had 8 terabyte blades and 52 terabytes on a single blade, times 15 within a chassis. So, we were really aiming for extreme density and architecting so that our Software must take advantage of all those bits.
07:32
So that was the genesis of Direct Flash way back then. It's a great callback to remember that, and interesting to hear about the desire. And that's what engineers do, right? It's the difference. The different choices you could have taken longer and had some sort of unification, But I was like, we’ve got to get this thing to market.
07:48
But then also being able to hit petabyte scale, right? Which we did, but did it with, with, with really fast performance. I do remember coming in, you know, having people test this out, and, And benchmark it, do POCs, and say, oh, this stuff's really fast. Yep, exactly. And, to the point of, uh, to the other
08:07
Point you made: we were architecting for a world where, as NA prices dropped, We wanted to ensure that the non-NA components in the system were as efficient as possible. Right? Because 10 years ago, NAN was significant More expensive than it is today. I mean, we're, But thanks to all the advantages that we brought in terms of extreme efficiency.
08:28
Design: getting rid of unnecessary parts. Remember back then, Back in those days, NVMe had barely settled, right? Like NVMe wasn't really a thing yet. So we were looking at extreme efficiency. Cutting down a bunch of components, like no SAS, no cables, Like, how do we remove the things that exist between software and the naan chips and make
08:49
This is as simple, efficient, and easy to scale as possible with Flash Array and Flash Blade has sort of benefited from that design experience that we've had. DirectFlash is what it It is today, because we gained significant experience building this on Flasher flash Blade: at the time, we were able to converge on a single architecture that then moved from sort of 2 bits per cell, uh, 2 bits per cell type technology we were dealing with initially,
09:15
MLC to TLC, which is 3 bits, then finally QLC: 4 bits per cell. We were able to leverage all of that into building extreme performance on QLC, which was previously thought to be just fit for slow, cheap, and deep types of use Archive-type applications, right? So, it’s really a process, and having lived Through that, it sort of informed us on how challenging it is to build for
09:42
NAND from the ground up and what sort of challenges the legacy storage vendors face Overcome and how hard it is to do so, because if you're just bolting on again, you're bolting on To top non-hard drive primitives, there's only so much efficiency you can Gain, and we're really seeing those, um, those benefits shine through as we look at some of The newer products that we’re building as we look through them.
10:04
Some of the hyperscaler announcements that we've recently been able to make, finally culminating in FlashBlade XA, where all of these advantages come together, like the fast Metadata we've built through FlashBlade and all our advantages in Direct Flash. Yeah, it's really interesting to me because we have gone through the 15 Architectural decisions, and I look at what you're describing.
10:25
The way that you design FlashBlade. It adhered to the same types of principles, right? You can see the DNA of cause and effect, the founders, and the whole philosophy. Not only applying to flash array types of things, but also when we did this incubation Project with FlashBlade—it was built the same, It's fascinating because we learned so much.
10:44
And the team had learned so much from Flash array that we were able to leverage a lot of The best practices, lots of even data structures from FlashRay. In fact, the original FlashBlade code was a fork off of a lot of the FlashArray stuff. And then, of course, there are significant modifications that had to be made because of you We're going from a scale-up architecture to scale-out.
11:02
Uh, but since then, we've been able to leverage the larger team, then reconvene in places where it makes sense to. Yeah, super exciting, and the portfolio has been built out as well. Right? It was just FlashBlade, and then, you know, There's AI made advances and added new capabilities into the software layer.
11:18
And then really, why—why I dragged you in here today was to, To kind of jump into FlashBlade X. And maybe not focus directly on that, but talk about why behind XA and hit some of The myths. I do want to hit a stat for you and just Get your perspective on it.
11:35
I was looking back on this this morning, because I was searching for something interesting We could put it in. I had no idea that in 2024, There were 251 million. GPUs that were shipped, right? Like, that's it, That blows my mind, and I'm an old, I'm an old CPU guy.
11:51
Right? Like I do, I go back to CPUs for servers. Obviously, much of it is dominated by Nvidia, but that number is, It's mind-blowing. Oh, and the other fact that I found. Was it that, for the last couple of years, more GPUs have been shipped than CPUs? Right? And you think about all the demand for PCs and
12:11
Servers and all that. It's mind-blowing. Like, is that something you were aware of? And then, what is, What is it? How does that make you feel? Yeah, it's crazy. Um, it's interesting. So before I came to Pure, my background was actually as a chip CPU architect.
12:26
So I used to build CPUs, uh, and it's, it's kind of, It's crazy to see the resurgence that chip design has had over the last few years. I mean, who would have ever thought that a chip vendor would be the most valued company The NASDAQ, right? Yeah, it's crazy. I mean, years ago, they said it would be commoditized, or that we wouldn't get it
12:45
The line widths and the fab continue to shrink, and that trend has continued down. And now you have these GPUs that just are dominating shipments, largely driven by AI. And I pulled this stat in because I think it's going to inform some of what we will get Relative to FlashBlade. X around this consumption. So, what we're going to do,
13:05
You know, and for viewers, we're going to hit like three myths. Right? And they’re in and around. AI and things that we believe we're answering with XA, because part of our job is to Listen to our customers and examine the market and understand what’s going on there. Before we head into that, maybe just a quick primer on XA.
13:24
Right? Because we brought this out earlier this year. In 2025, at GTC back in the spring, just make sure people have a level set on. Uh, you know, if you could do, like, a totally quick, high-level. Yeah. So, FlashBlade X really is an extension of The FlashBlade portfolio. It's extending FlashBlade into these areas,
13:43
Really large environments, ones that house many of the 250 million GPUs you need Just alluded to, many of those GPUs. Um, obviously, the growth in these clusters over the last couple of years has been Mind-boggling. Um, FlashBlade from the ground up was Architected for really good performance, uh, really good multi-dimensional performance where
14:07
We really focused on metadata; we focused on the ability to handle large and small accesses Concurrently reads and writes, um, because it's easier in general to be able to handle large reads or homogeneous types of applications, but where storage becomes more More involved and interesting is when you have. High degrees of concurrency and many of these operations occurring at the same point in time.
14:31
So FlashBid was really designed to handle that, but handle it with extreme care Availability and resiliency. And was built for what I'd call a high enterprise scale. So you'd scale into, say, tens or hundreds of petabytes and to terabytes of low Terabytes per second of performance.
14:52
Over the last couple of years, the build-outs we've seen have really pushed on the upper Barriers where you see 100,000s to millions of GPUs within single clusters, And they're looking for a unified storage experience. So our thesis was that we could leverage the advantages that FlashBlade already had. Namely, around our superlative metadata performance and really extending to these types
15:14
Of ecosystems. So the fundamental modification we made was to use the FlashBlade system as is. But to handle metadata, then to separate the processing of metadata and of data. So thereby allowing direct access. High-bandwidth paths to what I’ll call data nodes.
15:33
These are, well, think of these as servers with drives inside them. Now, these could be SSDs; these could be our DFMs, our Direct Flash Modules, Um, but in effect, you're creating these really wide, really large pipes to directly transmit Access data from, um, those data stores. And the critical applications, the ones that manage metadata are all handled by us
15:55
Flash Blades. So, by virtue of this disaggregation, We've built a product that effectively knows no bounds. Like, we can scale into tens of terabytes per second. And honestly, that's only going to be limited by how much gear you're able to throw at it. It performs really well on all types of metadata operations.
16:11
And it preserves many advantages that FlashRid brings to bear. Namely, high-performance file, high-performance object, and all of the other capabilities that exist within FlashBlade. But it sounds like you’re really kind of stripping out all the noise, Right? I mean, it's taking advantage of what
16:27
We've always done really well managing and manipulating metadata, Right? Creating basically data. Database for metadata, that’s what we do super well. But then removing or offloading some of the things that can become noisy and disruptive, So that you're really just focused on the metadata alone.
16:44
And it is tuning for applications that are less enterprise-heavy, Right? And tuning for places that need extreme scale. And extreme density. Uh, and it’s also tuning for places that candidly are fairly opinionated, Like some of these large build-outs, large vendors are somewhat opinionated on types
17:02
Of servers that they want to be able to work with, or that they'll house in their data centers. So with FlashBlade XA, we're also going down The path to enabling third-party data nodes hardware, which departs from the typical Enterprise motion that we have with FlashAir and FlashBlade. We pride ourselves on the experience that we provide to our customers.
17:21
Which is why That's why we build our own hardware and Flasher and Flashblade, but at an extreme scale, we, We recognize that some customers will be fairly opinionated about what they want To, uh, act as a server or even provide them choices around, um, the types of media they Wish to accommodate. Um, the other thing we've seen in this
17:40
Market is that. You need extreme flexibility, and I'm sure we'll get into that as we go through this. Right? We'll get to that, but I like the whole message. Around like a BYOS: Bring your own server if you want. You can bring your own server, kind of. If you like, but also designed for the level of flexibility because these,
17:57
No one customer is the same at this scale, and they have different, differing needs and Oftentimes, they don't even know because if you are offering a service to other customers, Sit on top of you, then you as a service provider—you’re just architecting for max Flexibility, and that's what we set out to build via Exa. So it's an extension of all the things that you know and the metadata performance that is
18:19
Superlative on FlashBlade, but extending it to extreme scales that the appliance can't Typically, scale. Got it. And I think it's good and important to have that as a foundation as well. Are taking what we built out in FlashBlade, just. Doing it a little differently, but still creating that extension.
18:35
It's not like we went out and built like an entirely new animal of Flashblade. It is effectively the same thing. We bring 10+ years of building, and supporting this, And all the learnings we've taken are enough Sort of rock solid. Um, metadata performance and all of our um, advantages around.
18:55
Availability and resiliency — everything that we've baked into our DirectFlash, Which leads me to myth one, and you’ve hit on a few of these things preemptively, Which is great, because it never hurts to repeat things. But myth one, or, you know, what we're trying to counter is that traditional architectures Are well suited.
19:12
For modern AI, we kind of hit that upfront where we're talking about even the development of FlashBlade, but take that on. I think the reality is we're We're doing something like FlashBlade XA because we believe in traditional architectures Cannot. Totally. Um, so we can hit this from multiple angles. Right? We hit the metadata separation, the need for
19:31
Really high-performance metadata and the ability to disaggregate it from yours, Um, data stores so they can grow independently. Um, so that's definitely a big reason for building FlashBit X and something we see Of extreme value at these large scales. Um, the other piece that I've hit on is, and we kind of,
19:52
We spoke about this earlier, where handling one type of operation is not the hardest. Like you can do large reads, and that’s generally what some legacy architectures do are good at. But unless you’ve purpose-designed for Flash And you've purpose-designed from the ground up, for your operating system to be flexible. Take advantage of all the inherent benefits of Flash; it is somewhat impossible to.
20:14
To achieve the level of metadata performance we're able to obtain, Right? And that has to do with designing from the Ground up and able to handle this high degree of concurrency and a high degree of Variability in the types of applications that you're running. Um, so again, you need to have a really good small file access.
20:32
You must have really good large-file access, but be able to handle reads and writes. Concurrently, what we've learned over the last years is It's hard to do that with an approach that just takes a legacy system and bolts that On top of SSDs. We see different vendors have differing, um, strategies around this.
20:49
Some people use a tier to manage their metadata. Um, now then. That works because you don’t have to do the hard work of pushing it to flash. You could use a more performant tier, perhaps a different type of memory. Now the challenge is that it becomes a boat anchor because as your system scales,
21:09
Your tier must grow in concert with the capacity that you’re scaling. Right? So, our thesis has always been that you want, You want to enable direct access to Flash because we're trying to optimize it, For minimizing the cost of all non-NaN components to deliver you the best TCO. And the way you do that is by doing all of the
21:29
Hard engineering works to have the right data structures so you can ship your data all The way down to Flash and not have to tear it out somewhere else. Yeah, worry about the complexity or classic challenge with converged infrastructures. Right, where you could, you had to, you had to scale differently at different times and things. Got kind of wonky or out of balance, right?
21:47
The multiple legs of the stool sound like a similar approach. Um, it is; it's all about building the right primitives to enable us to scale in the right Dimensions as needed, dictated by customers. Yeah, versatility, Right? That’s all; that’s all baked in. Um, great one. Myth number two that I want to hit.
22:05
And maybe this is perception, not reality, but high-performance AI workloads these days Are really only file-based. And I think the counter to that is, We see all these objects, we see all the unstructured; there's just a lot that's going On in there. Totally, yeah, um, yeah, that's definitely true. I'd say that if you look at where most of unstructured data lies
22:30
Data is stored. Well, it happens to be an object, Right? Now, the object hitherto was always this slow layer. That was kind of between disk and tape, perhaps, and, like, Like I might have mentioned this previously, it was always thought of as object equals slow Equals disk; file equals fast, equals flash.
22:50
Right? But if you think about this from AI's view Perspective: well, AI needs data. Where does most of your data reside? Well, a lot of it happens to sit on Object. So, how do you get maximum business insights? From your data if you're not able to? Quickly process all of the data that's sitting on an object,
23:06
Right? So, um, yes, it’s definitely a myth. We definitely see a rise in the, um, involvement of objects in, In AI use cases. Um, and that's what we set out to do with FlashBlade and FlashBlade XA: perform, is to provide extremely high performance to file an object, and we spoke earlier about how objects
23:24
Wasn't constructed. As a bolt-on, but was actually constructed. As a first-class citizen where file and object both have equal weight and equal access to an Underlying KV store. Like that's the fundamental reason for Flash Blade is able to scale as well as it does. So with XA, you could hit our extremely high performance levels.
23:42
That's about 3.5 terabytes per second per rack. You can do that on File and Object. So, other things that contribute towards this are advances, improvements we've made on. Advances we've introduced on objects such as the capability of doing RDMA through objects, Um, the capability of running a pure, pure key-value accelerator,
24:02
And that's something we recently integrated with NVIDIA's Dynamo. Um, we've also recently published a blog about how with a single chassis of FlashBlade, We were able to hit more than 3 trillion objects or so on a single FlashBlade. So that's a level of scale that's just mind boggling, right? And if you think about these AI use cases, again, you think about them,
24:24
Well, I have 100,000 GPUs and need loads of data that sit behind this. All of these scale factors become super important. Um, the other thing pushing in this direction is the evolution of AI use cases Going from single dimensional, perhaps like they were just doing text or just, Just images to multimodal.
24:44
That effectively means AI workloads are now processing text, video, and audio. And increasingly, that pushes toward more capacity being dedicated to these Workloads, and many of these videos and images, etc. Are all objects native? They all come as objects, so it makes more sense to be able to process them as such
25:04
Native formats rather than build an architecture that relies on objects getting Converted into file, and then File feeds the GPUs, then returns Object back to what we had earlier on first-class citizens, Right? Exactly. So, first-class data type, we're just treating Them all seem similar, right? So that extends AI’s reach across data
25:21
Center, so you're no longer bottlenecked in, Hey, I just have this filer sitting over Here. I have to move all of my object stuff on the Filter to be able to access the GPUs. No, you want to access it where it is. Sits, but Across the huge swath of data that exists, Is that kind of why we hear sometimes that this, this AI data that needs to be accessed is
25:38
Sitting in silos because of these different treatments and approaches. OK, that makes sense. Yeah, because it's not only garnering Access to GPUs, which are clearly in demand These days and networking is quite a hot commodity Around them, but it's also the storage system That sits behind. And oftentimes, that could just be a filer; it could just be running on file.
25:57
Um, we're seeing that change, and we're sort of, we're at the forefront of that. Um, and As we, as we work with our more sophisticated Customers, the larger ones, they’re all definitely object first. They're all looking to go object-native when they work with AI. And that's come up on a couple other pods I've done as well.
26:14
That's definitely a trend that's going on. I think today I'm publishing one with one of our unstructured data specialists, And that was one of his takeaways when we did our hot takes. You know, things that you think IT folks should be paying more attention to, He said. This whole object-first trend is really,
26:29
It is really going down right now. It's really happening. Yeah, super cool. Um, OK, the final one, and maybe we we gave away the, We gave away the end of the story. No, I always, I'm always to blame for leaving too early, but the myth that AI workloads need only GPUs
26:48
Horsepower, right? And that's why I pulled that stat in. Look, we know that hundreds of millions of GPUs are shipped annually. And that is the hype. That's what tends to drive the market. Drive the stock, the market; drive the excitement, make people excited about Nvidia and other GPU manufacturers. That's all great.
27:07
But make sure that you have the right story. Yeah, that extremely fast engine still needs a Steering wheel, right? So you still need to ensure you have a Way of feeding all of those 250 million GPUs. Um, I mean, I think it's been, um, it's been mind-boggling as to how, Um, the tenor of these discussions has changed.
27:28
I feel that, from a storage perspective, up to about a couple of years ago, When you get involved and try to size something with customers, We would generally go in and then get, get a sense from the customer around how large the Environment is how much performance they seek to push from storage. And try to find the right envelope that would fit, fit them.
27:48
Uh, that has evolved into a power-first conversation, of course. Right? Like the first thing that customers talk about About are, um, how many megawatts they have, right? Like the data center is thought of in terms first of megawatts or gigawatts. Second, how many GPUs are you getting access to?
28:05
Can we cool it? Can you cool it, please? Like that, that's primarily the data center, near water. That's the currency we're dealing with, right? So, um. Yes, people aren't yet conditioned to Think about networking and storage as those first-class citizens,
28:21
But again, the more sophisticated the customer, the more advanced, Uh, their needs are clear now. We're finding that once they set this up Stuff, they go, "Oh well, I really need a large storage solution." Or I need to right-size my storage solution to hit whatever I'm looking to hit out of my GPUs. Yeah, and certainly having us go to market this
28:38
Year with something like FlashBlade X starts to get them to ask that question or include Those considerations in decisions around power, space, and cooling. Because now they're looking at storage solutions in concert with GPUs and the Networking and all three legs of the stool, Right? And that was a huge consideration in our architecture for XA.
28:59
I mentioned earlier that we have a mode where you can bring your own server, Um, but we also have a mode where we can use our existing. Appliances as data nodes, along with RDFMs. So what we've definitely seen from customers is this, And I mentioned this earlier: no, no two customers are the same.
29:17
Some are pushing for extreme performance density, while others have more Moderate performance density needs but extreme capacity needs. So it's important to be able to right-size the solution to their needs. And this is where we see the benefits of our DFMs, or Direct Flash Modules, really shining. It's a super powerful message for us to be able to go into one of these customers and then talk
29:37
About how we can shrink their data center into a couple of racks. And the most important metric to them is well, Look at all this power that we’re saving you. You can actually go through more GPUs against your power budget. Or you have more space for the GPU side, right? The side as well. I mean, and that applies even on the block side.
29:54
Of the equation. Precisely, we're looking at a flash array. XL 190 and a 20 or 22x, you know, smaller ratio compared to the other. Solutions out there that free up space for GPUs in them. I mean, that's the same idea behind Excel, right? It's like how we jam as much performance Density, and how do we jam as much power density into our footprint as possible, freeing up space
30:16
and power and cooling for the GPUs, so it's sort of, Uh, um. It's a different way of talking about Storage, which is interesting, but again, if you come back, the primary currency is Keeping those GPUs fed and able to cool them and fit them. Then it's important to make sure that you construct the system with that in mind.
30:38
Storage design should not be your weak point. So that's where Flash Blu Exit shines. And at the same time, We're still talking about those different vectors. We're just doing it at scale, right? The capacity versus performance, versus like, And, like you said, the customers have different trade-offs on what they care about.
30:54
We're just now doing it at a crazy extreme scale, and that's the part that gets really Mind-boggling as well. And that's the importance behind being Flexible and being disaggregated, right? Like the, The reason to separate metadata. Data again is maybe you don’t need more. Metadata performance, but you just need to add data performance, so you have to be able to grow
31:11
One tier—or I shouldn't say tier one part of the system, independent of the other. You must have the ability to increase capacity without necessarily changing it Performance profile the whole time. Um, again, when you talk about your customer, being a service provider who is then dealing With one or two layers of customers, jamming in as much of this flexibility and versatility as
31:30
Possible, it helps make for a much better, um, experience altogether. It's a great solution. Alright, we're gonna hit our hot takes. Are you ready? Cool! So, we got hot takes. OK, we have three different questions. I tuned them a little bit for us. I tend to make them the same for many
31:42
Guests, but I tuned them a little, just because I know your background and what we're covering. And some of it may relate to what we’ve already hit on. But if there’s something different, the first thing I want you to do is ask, what’s a trend? Around AI data management, or maybe a blind spot that you see organizations having with making Their AI initiatives succeed.
32:04
I would say that, to me, it's the notion that, Um. AI should actually be thought of as horizontal. It extends across all you do in the data within your workflows. It's kind of like security, right? There are applications that like security which used to be an afterthought,
32:23
And then, increasingly, we're building it into everything that we do. Um. I feel like AI is going to be the same thing. I love that. It's not about applying AI in this one use. Case or this other use case. Everything is just like you need security everywhere.
32:35
Everything's going to have to be thought of with AI from the ground up. So, it's one of those horizontal ones. You had a good think on that, then you came up with a really provocative answer. I like that! That was really good. No, but it makes sense, right? Because if you return to security,
32:47
It was kind of, well, there’s the security part. Exactly. There used to be a security silo, and then we'll Just toss stuff over. Oh, here you go; make these. Go put these patches on, or you need to. You know, run this antivirus thing. And now it's pervasive in everything we do, and AI probably goes the same way.
33:01
AI goes the same way, from a developer standpoint, you have to think of security When you're building a product, similarly, you have to think about how it Intersects with AI when you build the product. Really provocative! You may win the award. For the most interesting ones of those that I've done so far.
33:16
Uh, I want to pitch on, so normally we ask folks about an oops. If they were working in, you know, I could ask you about when you were doing CPU design. If you, you know, threw something there, but instead, I'd rather focus on a positive, which is in the early days of FlashBlade, You and the engineering and product team management team.
33:34
We're bringing that thing to life. What was, What was a breakthrough story you had with a customer that stands out in your mind? We don't have to mention who the customer is obviously, unless they're a public reference. But, uh, one of those where you said, “Oh man, we've been doing years of work on this.” We went and solved this complex problem, and look at what it's doing now.
33:55
For the customer. You have one of those? I do. I have a couple of those. Um, one that we were public about was our working with Facebook AI Research. So FlashBlade and Flasher were integral parts of making that happen. Uh, that was a huge breakthrough moment for us in terms of,
34:17
Hey, we built this product to hit a bunch of caps; we hit a bunch of checkboxes, and And sort of capabilities, rather, uh, around architecting for maximum. Performance density, maximum power efficiency, and excellent performance metadata Performance and all of that were validated in the use cases we worked on with Facebook At the time, no Meta—so that was really cool. Uh, on a more personal basis,
34:40
Um, we were actually deployed, um, at. One of the places, I'm actually not sure if they're a public reference. You might remember this, but that was, um, at the forefront of COVID research. Oh, yeah, I do remember. Yes. So, that was... that was, uh.
34:57
That gave me the warm fuzzies. Yeah, it's good warm fuzzies to know we were Part of them trying to figure that out. And exactly, Like we were at the forefront. I remember us engaging with those who the customer and meaningfully moving stuff Along, you know, as they were looking to sequence genomes and everything that came
35:15
Around with COVID when the pandemic hit. Yeah, and I won’t mention them because I don’t I think they're a public reference, but aren't. I think I saw. More investment from them recently, showing that they value the relationship, The relationship continues after four or five years, which is, Which is super cool. OK.
35:31
Final one: take yourself back a couple of decades; go back 20 years, Maybe back when you're doing chip design, when we stopped the recording. I'm going to ask you where you were and how you were doing that. Because I'm curious; I come from that space. Uh, go back 10 or 20 years. What would you have told yourself?
35:46
Your younger self about data management in the future or the industry? What would you have liked to know 20 years ago? That would have informed what you're doing today? Well, OK, so if I go back that far, I was definitely building chips at the time, Like I said, right? Um, it's funny; I was still trying to
36:05
Eke out. So, I guess it's the contrary take, Right, not many changes; just the scale that things change at. Um, I was trying to eke out milliwatts on operations, which is huge for a chip. Um, and we just spoke about how important power is, and it's those milliwatts that just stack up, Stack up, stack up, and here we have a super, super-efficient power solution for
36:30
Storage that is enabling these bigger chips and AI, so, Um. I guess I would tell myself that it's important. To sweat the details, because I can tell you, it was a really painful process at the time. You're always questioning, well, do I really need to figure out what's below milliwatts? Kilowatts, and do I go save them? But the details matter, and you stack them up, and that's,
36:52
That's how we build products as efficiently as it has been worthwhile and still is worthwhile to Extract as much power out of the silicon as possible. Yeah, I get it. I like it. Um, this was great! Did you have fun? This is awesome. It's my first time here.
37:05
I can't believe we haven't done this before. No, but you'll be back because this is great. Super easy. Well, anybody who comes on has the, The history, and, you know, it's kind of how, How you’ve evolved is always, always interesting too. It's like when I have Chad Kenny on; it's the same kind of thing, as you guys go far
37:22
Back and know, you know, but then you're able to relate it to what we're doing currently Times too, which is awesome. So, we'll do that. We'll figure it out. I'll work with Bish; we'll figure out some Other things that we can do, but I'm glad we finally got you on here. FlashBlade X is super exciting! Congrats to the team.
37:37
I know it's been about six months or so, but just a great announcement. People are paying attention to what we're doing in the AI space. Which I love, and there's more goodness to come. I'm actually going to have one of the AI specialists on to talk about KVA soon, too. Because I think that's an interesting enough thing that we could do a full episode on
37:56
Around the Key Valley Accelerator. It's a super fun topic. Thanks for that. I think you're going to have a lot of AI-related Topics. Oh, yeah, yeah, I have to kind of fish through. And make sure there are plenty of other things. But no, This was super cool for anyone out there who wants to learn more about what Pure is doing.
38:11
This space, you can go to Pure.AI. That is the custom landing page we have. For all of our AI projects, including, um, FlashBlade Xaa, Uh, super fun. People want to find you, and you're a bit Low profile, but you're findable on LinkedIn. LinkedIn. OK, if somebody wants to ask you a question,
38:29
Something as a follow-up is never a bad thing. Um, thanks again for coming on. This was great. Thank you all out there for watching this episode. Thank you very much for listening. If you're on the audio program, as always, tell a friend, Tell a colleague, and we'll keep having the great guests like Hari on the program.
38:47
With that, we will wrap up for Pure Storage. Hari Cannon, this is Rob Ludeman saying, don't look back. Something might be gaining on you.
