[Music] [Applause] I'll Buena Burris and welcome to another cube conversation from our fantastic studios in beautiful palo alto california today we're going to be talking about what infrastructure can do to accelerate AI and specifically we're gonna use a relationship a burgeoning relationship between PDN and nvidia to describe what we can do to accelerate AI workloads by using higher performance smarter and more focused of infrastructure for computing now to have this conversation we've got two great guests here we've got Kurt ku kind who is the senior director of marketing at ddn and also Darren Johnson is a global director of technical marketing for enterprise and NVIDIA Kurt Gerron welcome to the cube thanks for thank you very much so let's get going on this because this is a very very important topic and I think it all starts with this notion of that there is a relationship that you guys have put forward Kurt once you describe it sure well so what we're announcing today is ddn's a3i architecture powered by Nvidia so it is a full rack level solution a reference architecture that's been fully integrated and fully tested to deliver an AI infrastructure very simply very completely so if we think about how this is gonna or why this is important AI workloads clearly have a special stress on underlying technology Darin talk to us a little bit about the nature of these workloads and why in particular things like GPUs and other technologies are so important to make them go fast absolutely and as you probably know AI is all about the data whether you're doing medical imaging whether you're doing natural language processing whatever it is it's all driven by the data the more data that you have the better results that you get but to drive that data into the GPUs you need great IO and that's why we're here today to talk about ddn and the partnership of how to bring that I owe to the GPUs on our dgx platforms so if we think about what you described a lot of small files off and randomly just riveted with nonetheless very high-profile jobs that just can't stop midstream and start over absolutely and if you think about the history of high-performance computing which is very similar to a I really I owe is just that lots of files you have to get it they're low latency high throughput and that's why ddn's probably nearly twenty years of experience working in that exact same domain is perfect because you get the parallel file system which gives you that throughput gives you that low latency just helps drive the GPU so we you'd mention HPC from 20 years of experience now it used to be that HPC you'd have scientists with a bunch of graduate students setting up some of these big honkin machines but now we're moving into the commercial domain you don't have graduate students running around you don't have very low cost high quality people you're you know a lot of administrators who nonetheless good people but a lot to learn so how does this relationship actually start making or bringing AI within reach of the commercial world exactly where this reference architecture comes in right so a customer doesn't need to start from scratch they have a design now that allows them to quickly implement AI it's something that's really easily deployable we've fully integrated this solution ddn has made changes to our parallel file system appliance to integrate directly within the DG x1 environment makes that even easier to deploy from there and extract the maximum performance out of this without having to run around and tune a bunch of knobs change a bunch of settings it's really gonna work out of the box and the you know nvidia has done more than just the DG x1 it's more than hardware you've done a lot of optimization of different of AI toolkits if Sarah I'm talking what about that Darin yeah so I mean talking about the example I use researchers in the past with HPC what we have today are data scientists data scientists understand pie tours they understand tensorflow they understand the frameworks they don't want to understand the underlying filesystem networking RDMA InfiniBand any of that they just want to be able to come in run their tensorflow get the data get the results and just turn that keep turning that whether it's a single GPU or 90 Jex's or as many dejection as you want so this solution helps bring that to customers much easier so those data scientists don't have to be system administrators so a reference architecture that makes things easier but that's more than just for some of these commercial things it's also the overall ecosystem new application providers application developers how is this going to impact the aggregate ecosystem it's growing up around the need to do AI related outcomes well I think one point that Darrin was getting to you there and one of the big effects is also as these ecosystems reach a point where they're going to need to scale right there's somewhere where ddn has tons of experience right so many customers are starting off with smaller data sets they still need the performance a parallel file system in that case is going to deliver that performance but then also as they grow right going from one GPU to 90 G X's is going to be an incredible amount of both performance scalability that they're going to need from their i/o as well as probably capacity scalability and that's another thing that we've made easy with a3i is being able to scale that environment seamlessly within a single namespace so that people don't have to deal with a lot of again tuning and turning of knobs to make this stuff work really well and drive those outcomes that they need as they're successful right so in the end it is the application that's most important to both of us right it's it's not the infrastructure it's making the discoveries faster it's processing information out in the field faster it's doing analysis of the MRI faster it's you know helping the doctors helping the anybody who's using this to really make faster decisions better decisions exactly and just to add to that I mean in automotive industry you have datasets that are from 50 to 500 petabytes and you need access to all that data all the time because you're constantly training and Retraining to create better models to create better autonomous vehicles and you need you need the performance to do that ddn helps bring that to bear and with this reference architecture simplifies it so you get the value add of nvidia gpus plus its ecosystem of software plus DD on its match made in heaven Darren Johnson Nvidia Curt Koo Kien ddn thanks very much for being on the cube thank you very much and I'm Peter burrs and once again I'd like to thank you for watching this cube conversation until next time [Music]

(laughing) >> This is V-3. >> Alec, you're going to open up, we're going to cut, come to you in a second. Good luck, buddy. Okay, here we go. Alright Peter, ready? >> Yup. >> And we're coming to you in. >> Hold on guys, sorry, I lied. (laughing) V-2, V-3, there it is. Okay, ready. >> Now you're ready? >> Yup. >> You're ready ready? Okay here we go, ready and, three, two. >> Hi, I'm Peter Burris, welcome to another Cube Conversation from our wonderful studios in beautiful Palo Alto, California. Great conversation today, we're going to be talking about the relationship between AI, business, and especially some of the new infrastructure technologies in the storage part of the stack. And to join me in this endeavor is Kurt Kuckein, who's a senior director of product marketing at DDN. Kurt Kuckein, welcome to The Cube. >> Thanks, Peter, happy to be here. >> So tell us a little bit about DDN to start. >> So DDN is a storage company that's been around for 20 years. We've got a legacy in high-performance computing, and that's what we see a lot of similarities with this new AI workload. DDN is well-known in that HPC community; if you look at the top 100 supercomputers in the world we're attached to 75-percent of them and so we have a fundamental understanding of that type of scalable need that's where we're focused, we're focused on performance requirements, we're focused on scalability requirements, which can mean multiple things, right, it can mean the scaling of performance, it can mean the scaling of capacity, and we're very flexible. >> Well let me stop you and say, so you've got a lot of customers in the high-performance world, and a lot of those customers are at the vanguard of moving to some of these new AI workloads. What are customers saying? With this significant engagement that you have with the best and the brightest out there, what are they saying about this transition to AI? >> Well I think it's fascinating that we kind of have a bifurcated customer base here, where we have those traditionalists who probably have been looking at AI for over 40 years, right, and they've been exploring this idea and they've gone through the peaks and troughs in the promise of AI, and then contraction because CPUs weren't powerful enough. Now we've got this emergence of GPUs in the supercomputing world, and if you look at how the supercomputing world has expanded in the last few years, it is through investment in GPUs. And then we've got an entirely different segment, which is a much more commercial segment, and they're maybe newly invested in this AI arena, right, they don't have the legacy of 30, 40 years of research behind them, and they are trying to figure out exactly, you know, what do I do here? A lot of companies are coming to us, hey, I have an AI initiative, well what's behind it? Well, we don't know yet, but we've got to have something and they don't understand where is this infrastructure going to come from. >> So the general availability of AI technologies, and obviously Flash has been a big part of that, very high-speed networks within data centers, virtualization certainly helps as well, now opens up the possibility for using these algorithms, some of which have been around for a long time, but have required very specialized bespoke configurations of hardware, to the enterprise. That still begs the question, there are some differences between high-performance computing workloads and AI workloads. Let's start with some of the, what are the similarities, and then let's explore some of the differences. >> So the biggest similarity, I think, is just it's an intractable, hard IO problem, right, at least from the storage perspective. It requires a lot of high throughput, depending on where those IO characteristics are from, it can be very small-file, high-op-intensive type workflows, but it needs the ability of the entire infrastructure to deliver all of that seamlessly from end to end. >> So really high-performance throughput so that you can get to the data you need and keep this computing element saturated. >> Keeping the GPU saturated is really the key, that's where the huge investment is. >> So how do AI and HPC workloads differ? >> So how they're fundamentally different is often AI workloads operate on a smaller scale in terms of the amount of capacity, at least today's AI workloads. As soon as a project encounters success, what our forecast is, is those things will take off and you'll want to apply those algorithms bigger and bigger data sets. But today, you know, we encounter things like 10-terabyte data sets, 50-terabyte data sets and a lot of customers are focused only on that. But what happens when you're successful, how do you scale your current infrastructure to petabytes and multi-petabytes when you'll need it in the future? >> So when I think of HPC, I think of often very, very big batch jobs, very, very large, complex data sets. When I think about AI, like image processing or voice processing, whatever else it might be, I think of a lot of small files, randomly accessed. >> Right. >> That require nonetheless some very complex processing, that you don't want to have to restart all the time. >> Right. >> And a degree of simplicity that's required to make sure that you have the people that can do it. Have I got that right? >> You've got it right. Now one, I think, misconception is, is on the HPC side, right, that whole random small file thing has come in in the last five, 10 years and it's something DDN's been working on quite a bit, right. Our legacy was in high-performance throughput workloads, but the workloads have evolved so much on the HPC side as well, and, as you posited at the beginning, so much of it has become AI and deep-learning research >> Right, so they look a lot more alike. >> They do look a lot more alike. >> So if we think about the revolving relationship now between some of these new data-first workloads, AI-oriented, change the way the business operates types of stuff, what do you anticipate is going to be the future of the relationship between AI and storage? >> Well, what we foresee really is that the explosion in AI needs and AI capabilities is going to mimic what we already see and really drive what we see on the storage side, right? We've been showing that graph for years and years and years of just everything going up and to the right, but as AI starts working on itself and improving itself, as the collection means keep getting better and more sophisticated and have increased resolutions, whether you're talking about cameras or in life sciences, acquisition capabilities just keep getting better and better and the resolutions get better and better, it's more and more data, right? And you want to be able to expose a wide variety of data to these algorithms; that's how they're going to learn faster. And so what we see is that the data-centric part of the infrastructure is going to need to scale, even if you're starting today with a smaller workload. >> Kurt Kuckein, DDN, thanks very much for being on The Cube. >> Thanks for having me. >> And once again, this is Peter Burris with another Cube Conversation, thank you very much for watching. Until next time. (electronic whooshing)