 from downtown San Francisco. It's theCUBE, covering IBM Chief Data Officer Strategy Summit 2018, brought to you by IBM. Welcome back to San Francisco, everybody. We're at the Park 55 in Union Square. My name is Dave Vellante, and you're watching theCUBE, the leader in live tech coverage. And this is our exclusive coverage of IBM's Chief Data Officer Strategy Summit. They hold these both in San Francisco and in Boston. It's an intimate event, about 150 Chief Data Officers really absorbing what IBM has done internally and IBM transferring knowledge to its clients. Stephen E. Luke is here. He's one of those internal practitioners at IBM. He's the Vice President of Deep Learning in the Global Chief Data Office at IBM. We just heard from him and some of his strategies and use cases. He's joined by Summit Gupta, a CUBE alum who's the Vice President of Machine Learning and Deep Learning within IBM's Cognitive Systems Group. Summit. Thank you. Good to see you. Welcome back, Stephen. Let's get into it. So, I was paying close attention when Bob Picciano took over the Cognitive Systems Group. I said, hmm, that's interesting. Recently, software guy, of course I know he's got some hardware expertise, but bringing in somebody who's deep into software and machine learning and deep learning and AI, Cognitive Systems, into a systems organization. So, you guys specifically set out to develop solutions to solve problems like Stephen's trying to solve. Explain that. Yeah, so I think there's a revolution going on in the market, in the computing market, where we have all these new machine learning and deep learning technologies that are having meaningful impact or have a promise of having meaningful impact. But these new technologies are actually, significantly, I would say complex, and they require very complex and high performance computing systems. You know, I think Bob and I think, in particular, IBM saw the opportunity and realized that we really need to architect a new class of infrastructure, both software and hardware, to address what data scientists like Steve are trying to do in the space, right? The open-source software that's out there, TensorFlow, Cafe, Torch, these things are truly game-changing, but they also require GPU accelerators. They also require multiple systems, right? In fact, interestingly enough, you know, some of the supercomputers that we've been building for the scientific computing world, that those same technologies are now coming into the AI world in the enterprise. So, the infrastructure for AI, if I can use that term, it's got to be flexible, Steven, we were sort of talking about that elastic versus I'm even extending it to plastic. As Smith just said, it's got to have the tooling, got to have that modern tooling. You've got to accommodate alternative processor capabilities, and so that forms what you've used, Steven, to sort of create new capabilities, new business capabilities within IBM. I wanted to, we didn't touch upon this before, but we touched upon your data strategy before, but tie it back to the line of business. You essentially are presuming for a liaison between the line of business and the chief data officer office. So, how did that all work out and shake out? You're defining the business outcomes, the requirements, how did you go about that? Well, actually, surprisingly, we have very little new use cases that we're generating internally for my organization, because there's so many to pick from already throughout the organization, right? There's all these business units coming to us and saying, hey, now the data's in the data lake, and now we know there's more data, now we want to do this, how do we do it? So that's where we come in, that's where we start touching and massaging and enabling them, and that's the main efforts that we have. We do have some derivative works that have come out, have been new offerings that you'll see here, but mostly, we already have so many use cases that from those business units that we're really trying to heighten and bring extra value to those domains first. So, a lot of organizations, it sounds like IBM was similar, you created the data lake, things like Hadoop made it lower cost to just put stuff in the data lake, but then it's like, okay, now what? So is that right? You've got the data in this bog of data, you're trying to make more sense out of it, get more value out of it, and that's what they were pushing you to do. Yeah, absolutely, and with that, with more data, you need more computational power, and actually, Sumit and I go pretty far back, and I can tell you, from my previous roles, I heightened to him many years ago some of the deficiencies in the current architecture in x86, et cetera, and I said, if you hit these points, I will buy these products, and what they went back and they did is they addressed all of the issues that I had. Like, there's certain issues where I tried. That's when you were, sorry to interrupt, that's when you were a customer, right? That was when I was an external customer, you're still a customer. I'm still an internal customer, so I've always been a customer, I guess, in that role, right? But, you know, I need to get data to the computational device as quickly as possible, and with certain older gen technologies like PCI-Gen3 and certain issues around x86, I couldn't get that data there for high-fidelity imaging for autonomous vehicles, for high-fidelity image analysis, but with certain technologies in power, we have NVLink and directly to the CPU, and we also have PCI-Gen4, right? So, these are big enablers for me, so I can really keep the utilization of those very expensive compute devices higher because they're not starved for data. And you've also put a lot of emphasis on IO, right? I mean, that's... Yeah, you know, if I may break it down, right? There's actually, I would say, three different pieces to the puzzle here, right? The highest level, from a Steven Steem's perspective, the Steven Steem's perspective, or any data scientist's perspective is, they need to just do their data science and not worry about the infrastructure, right? They actually don't want to know that there's an infrastructure, they want to say launch job, right? That's the level of granularity they want, right? In the background, they want our schedulers, our software, our hardware to just seamlessly use either one system or scale to a hundred systems, right? To use one GPU or to use a thousand GPUs, right? So that's where our offerings come in, right? We went and built this offering called PowerAI, and PowerAI essentially is open source software like TensorFlow, like Cafe, like Torch, but with performance and capabilities, added to make it much easier to use. So for example, we have an extremely terrific scheduling software that manages jobs called Spectrum Conductor for Spark. So as the name suggests, it uses Apache Spark. But again, the data scientist doesn't know that they say launch job, and the software actually goes and scales that job across tens of servers or hundreds of servers. The IT team can determine how many servers they're going to allocate for data scientists. They can have all kinds of user management, data management, model management software. We take the open source software, we package it. You know, surprisingly, most people don't realize this. The open source software like TensorFlow has primarily been built on Ubuntu. And most of our enterprise clients, including Steven, are on Red Hat. So we re-engineered Red Hat to be able to manage TensorFlow. And you know, I chose those words carefully. There was a little bit of engineering both on Red Hat and on TensorFlow to make that whole thing work together. Sounds trivial, took several months, and huge value proposition to the enterprise clients. And then the last piece I think that Steven was referring to is, we're also trying to go and make AI more accessible for non-data scientists or I would say even data engineers. So we, for example, have a software called PowerAI Vision. This takes images and videos and automatically creates a trained deep learning model for them, right? So we analyze the images. You of course have to tell us, in these images, for these 100 images, here are the most important things. For example, you have to identify, here are people, here are cars, here are traffic signs. But if you give us some of that labeled data, we automatically do the work that a data scientist would have done and create this pre-trained AI model for you. This really enables very rapid prototyping for a lot of clients who either kind of want to have data scientists or don't want to have data scientists. So just to summarize that, so the three pieces, it's making it simpler for the data scientist. Just run the job. The backend piece, which is the schedulers, the hardware, the software doing its thing, and then it's making that data science capability more accessible. Those are really the three layers. Right, right, right. So all the resaid in my words, maybe if you don't mind. Ease of use, right? Hardware software optimized for performance and capability and point and click AI, right? AI for non-data scientists, right? It's like the three levels that I think of when I'm engaging with data scientists and clients. And essentially it's embedded AI, right? I haven't been making the point today that a lot of the AI is going to be purchased from companies like IBM, and I'm just going to apply it. I'm not going to try to go build my own AI, right? I mean, is that the right way to think about it as a practitioner? I think we talked about it a little bit on the panel earlier, but if we can leverage these pre-built models and just apply a little bit of training data, it makes it so much easier for the organizations and so much cheaper. They don't have to invest in a crazy amount of infrastructure. All the labeling of data, they don't need to do that. So I think it's definitely steering that way. It's going to take a little bit of time. We have some of them there, but as we iterate, we're going to get more and more of these types of commodity-type models that people can utilize. I'll give you an example. So we have a software called Intelligent Video Analytics at IBM. It's very good at taking any surveillance data, and for example, recognizing anomalies or if people aren't supposed to be in a zone. And we had a client who wanted to do worker safety compliance. So they want to make sure workers are wearing their safety jackets and their helmets when they're in a construction site. So we used surveillance data, created a new AI model using PowerAI vision and that we were then able to plug into this IVA Intelligent Video Analytics software. So they have the nice GUI-based software for, you know, with the dashboards and the alerts, yet we were able to do incremental training on their specific use case, but by the way, their specific, you know, equipment and jackets and stuff like that. And create a new AI model very quickly for them to be able to deploy and make sure their workers are actually complying to all the safety requirements they have on the construction sites. Interesting. So sometimes it's like a new form of captures is identify all the pictures with bridges, right? That's the kind of thing that you're capable to do with these video analytics, right? That's exactly right. Every, you know, clients will have all kinds of use case. I was talking to a client who is a major car manufacturer in the world. He was saying it would be great if I could identify the make and model of what cars people are driving into my dealership because I bet I can draw a correlation between what they drive into and what they're going to drive out of, right? Marketing insights, right? And, you know, so there's a lot of things that people want to do with which are very bespoke in their use cases and build on top of existing AI models that we have already. You mentioned x86 before and not to start a food fight, but... And we use both internally, too. So let's talk about that a little bit. I mean, where do you use x86? Where do you use IBM cognitive and power systems? I have a mix of both. Why? How do you decide? For certain types of workloads, I will delegate that over to power just because, you know, they're data-starved and we're noticing that computation is being impacted by it. But because we deal with so many different organizations, certain organizations optimize for x86 and some of them optimize for power. And I can't pick. I have to have everything. Just like I mentioned earlier, I also have to support cloud on-prem. It doesn't, like, I can't pick just to be on-prem, right? So... Well, imagine the big clouds of providers are in the same boat, which, you know, some of your customers. But you're betting on data. You're betting on digital. It is a good bet. Yeah, 100%. You're betting on data and AI, right? So I think data, you got to do something with the data, right? And analytics and AI is what people are doing with that data. We have an advantage both at the hardware level and at the software level. In these two, I would say, workloads or segments, which is data and AI, right? And we fundamentally have invested in the processor architecture to improve the performance and capabilities, right? You can operate on much larger AI models on a power system with GPUs than you can on an X86 system with GPUs. That's one advantage. You can train an AI model four times faster on a power system than you can on an Intel-based system. So the clients who have a lot of data who care about how fast their training runs are the ones who are committing to power systems today. Little latency requirements, things like that. Really, really big deal. So what that means for you as a practitioner is you can do more with less or is it, I mean... I could definitely do more with less, but the real value is that I'm able to get an outcome quicker. Everyone says, okay, you can just roll out more GPUs, more GPUs, run more experiments, run more experiments. No, no, that's not actually it. I want to reduce the time for a, an experiment. I get it done as quickly as possible so I get that insight. Because then what I can do is I can't possibly cancel out a bunch of those jobs that are already running because I already have the insight knowing that that model's not doing anything, right? So it's very important to get the time down. Jeff Dean said it a few years ago. He uses the same slide often. But, you know, when things are taking months, you know, that's what happened basically from the 80s up until, you know, 2010. You know, we didn't have the computation, didn't have the data. Once we were able to get that experimentation time down, we're able to iterate very, very quickly. And throwing GPUs at the problem doesn't solve it because it's just too much complexity? Well, it helps, it helps the problem. There's no question. But when my GPU utilization goes from 90s, 5% down to 60%, you know, I'm getting only a two thirds return on investment there. It's a really, really big deal. I mean, the key here, I think that Stephen, I'll draw it out again, is this time to insight. Because time to insight actually is time to dollars, right? People are using AI either to make more money, right? By providing better customer products, products to their customers, giving better recommendations. Or they're saving on their operational cost, right? They're improving their efficiencies. Maybe they're routing their trucks in the right way. They're routing their inventory in the right place. They're reducing the amount of inventory that they need. So in all cases, you can actually correlate AI to a revenue outcome or a dollar outcome. So the faster you can do that, you know, I tell most people that I engage with, is the hardware and software that they get from us pays for itself very quickly because they make that much more money or they save that much more money using power systems. We even see this internally. I've heard stories and all that, submit kind of comment on this, but there's actually sales people that take this software and hardware out and they're able to get an outcome sometimes in certain situations where they just take the client's data and they're sales people. They're not data scientists. They train it. It's so simple to use. And then they present the client with the outcomes the next day and the client's just like blown away. And this isn't just a one-time occurrence. Like sales people are actually using this, right? So it's getting to the area where it's so simple to use, you're able to get those outcomes that we're even seeing it, you know, deals close quicker. Yeah, that's powerful and submit to your point. The business case is actually really easy to make. You can say, okay, this initiative that you're driving, what's your forecast for how much revenue, you know, let's make an assumption as to how much faster we're going to be able to deliver it. And if I can show them a one day turnaround on a corpus of data, okay, let's say two months times whatever my time to break, you know, I can run the business case very easily and communicate to, you know, the CFO or whomever the line of business had. So that's right. I mean, just, I was at a retailer at a grocery store, local grocery store in the Bay Area recently. And he was telling me how in California we've passed legislation that does not allow plastic bags anymore. You have to pay for that. So people are bringing their own bags. But that's actually increased theft for them because people bring their own bag, put stuff in it and walk out, right? And he really wants to have a analytics system that can detect if someone put something in a bag and then did not buy it at purchase, right? So it's, in many ways they want to use the existing camera systems they have but automatically be able to detect fraudulent behavior or anomalies, right? And it's actually quite easy to do with a lot of the software we have around PowerAI vision, around intelligent video analytics from IBM, right? And that's what we were talking about, right? Take existing trained AI models on vision and enhance them for your specific use case and the scenarios you're looking for. Excellent. All right, guys, we've got to go. Thanks, Steven. Thanks, Samit, for coming back on and appreciate the insights. Thank you. You're welcome. All right, keep it right there, buddy. We're back with our next guest. You're watching the Cube at IBM's CDO Strategy Summit from San Francisco. We'll be right back.