 So we're going to build upon what we just heard from George's panel and distinguished guests. And now we're going to talk about specifically how that turns or how that impacts the technologies of database, analytics, and visualization, sorry. So to do that, I'm going to, and look at this, by the way, I discovered a number of years ago that the best way to predict health is age and obesity. And you'll notice that this fine-looking fella is quite a bit younger and lighter than I am. But joining me today, we'll be, we'd like to introduce them and bring them up right now. Mark Brooks, Principal Systems Engineer at Connecticut. I thought I'd show you back here. Here you are, Mark, come on up. Bill Mamone, Bill, are you here? There you are, come on up, Bill. VP of Engineering at MAPD. And why don't we have you two guys sit next to each other. Mark Hammond, Founder and CEO of Bonsai. Big announcement this week, Mark. Scott Wiener or Scott Weiner. Board of Directors of Scream Tech, SQ-Reem Tech, and Claudio, are you coming back up? Come on up, I'll stand up. Claudio, it's okay, it's okay. No, we did enough damage to you, you don't have to reconfigure the stage for us. Okay, so as I said, we're gonna spend the next half hour or so talking about how these technologies are, the technologies of database, analytics, and visualization are coming together to impact with some of the GPU technologies and other things that are happening to really drive change in how we do things. And I think I'd like to start off with a relatively simple straightforward question to open to anybody. Let's start with the obvious ones. What are some of the, oh, actually, you know what? Let me introduce, let you guys, following George. Mark, why don't you introduce yourself, introduce your company, we'll go right down the line. Mark Brooks, I'm a principal system engineer for Connecticut. I'm a field guy, so I do installations of our GPU accelerated cluster database. Bill Mamon, the Vice President of Engineering for MAPT. Been doing work with databases for the last 30 years from early relational vector databases, multi-dimensional, and I'm totally excited to be working with GPU-based databases. Finally, our servers have come. I'm Mark Hammond, I'm the founder and CEO of Bonsai. We create a developer platform and tools to enable non-machine learning experts to take advantage of all of the artificial intelligence technology, and we leverage GPUs quite heavily in exploring and optimizing those models that get built. I'm Scott Wiener, I'm on the Board of Advisors of Scream Technology. We're a GPU accelerated database company with popular ambitions. We don't think GPUs are only suitable for cutting edge work. We think that as CPUs have topped out, GPUs represent the next general computing architecture. And so we're looking at the millions of DBAs and BI installations, new projects that require pretty straightforward analytics, and bringing them from the gigabyte to the terabyte to the petabyte range without requiring new skills, retraining, and new infrastructure. And so we think GPUs should be everywhere. And I'm Claudio Silva, I'm a professor at NYU. I teach computer science and data science, and I direct the Center for Data Science. All right, thank you. So I think the first question I wanna ask is, Scott, I'd like to build upon a comment that you made. GPUs should be everywhere. What is it about a GPU that specifically is so crucial to accelerating and improving the way that database analytics and visualization works? So it has to do with the way that our ability to push semiconductor technologies ahead has kind of hit some early limits. And so you get a certain number of transistors on a chip, and we used to be able to just make them go faster. Just crank a knob and there we get faster. Every time you'd buy a new computer, it would be twice as fast. That hasn't happened in quite a while. So everyone's been looking for ways. How do we make better use of these transistors to solve problems in new ways, in parallel rather than sequentially? And it's not just new kinds of problems that are amenable to these solutions, but we also looked back and said, what are these traditional problems that have stopped scaling? All of a sudden we need to deploy 100 servers to process reasonable-sized, a couple 100 terabytes of information. Why can't we do this on a single server? Why can't we use new technologies like the GPU? And that's in fact what we do. On a two-use server, we can process near-petabyte data sets in near real-time, very low latency on ingest, and with interactive response times for queries. So GPUs are, they're kind of mind-boggling in the transformation they bring even to traditional spaces. It's just not getting as much attention yet, I think. So as a technology, our ability to drive performance with technology by itself starts to slow down. By that, I mean, you know, different geometries and chip types and whatnot. We have to turn to architecture and organization. So when we think about that, there's some new things happening in the database world. So I'd like both the MAPD and the Connecticut to got the, sorry, the Connecticut guys to talk about it. So Mark, why don't you talk a little bit about what's going on in the database realm? Okay, so like others in the panels earlier and many at Connecticut, I spent four years at Cloudera and there's a lot of Hadoop context out there and people have gone through the trough of disillusionment trying to do fast analytics on big data. So against that backdrop, we look at people wanting to do very fast analytics on very deep and wide data sets. So they don't have to do calculation of aggregates on a Lambda architecture and then make them available to the speed layer. We just like people who just wanna have a speed layer that can go fast and deep at the same time. And Bill? Well, they're important component to add. So I think we agree on this Scott in terms of making GPU based database products easy to use with a SQL interface. I'd also point back to the GPU. So also important to be able to do something to get some insight out of your data. So another component that map the ads in this is to use those same GPUs for powerful visualization. So in the same GPUs that you're using for real-time analytics, you can paint a beautiful display. You could paint this on a virtual reality device, query your five billion rows in 20 milliseconds and refresh that 50 frames a second and swim in your data. So people don't just need to run queries under data. They need to be able to visualize it and understand it. And so from a visualization standpoint, we're starting to use visualization as a technique to explain or make data come alive for people. Now, as we think about that though, it still requires some new approaches to thinking about the relationship between knowledge, learning, et cetera. Mark, why don't you tell us a little bit about some of the things that Bonsai is discovering, is it envisions the role that machine learning and AI is gonna play in making it possible to do some of these other visualization and related types of things? Yeah, absolutely. So with databases, we gained a great deal of advancement from the abstraction that they brought. We didn't have to worry and think about the low-level mechanics of how the database itself was operating and managing storage, how it was managing balancing all the trees. We just got the benefits of all of that, right? So at the same time that we get the benefits from the compute, we were getting benefits from the abstractions that the database is brought to that. When it comes to AI, we are in the weeds right now. We're very much in the weeds where we're working at the equivalent of assembly language, right? We keep getting improvements there, but it's still very low-level. So what we're doing at Bonsai is we're enabling people to get that level of abstraction moved up so that you can focus on the concepts that you actually want the system to learn as opposed to the mechanics of how that learning takes place. It allows you to shift the focus to what you want to teach as opposed to how that can be learned. And that allows you to utilize AI in a novel way, allow you to bring it to bear in a lot of interesting context. From the last panel, there was the question about how do you do SQL queries across pixels? The answer with AI technology is that you end up in viewing the software and the databases and all these other technologies with enough intelligence that you don't ask it about pixels. You ask it about the concepts within the pixels that you care about, and the database can chew through that and get that back to you. So we need to be able to enable that, but not by requiring everyone to work down in the weeds. We need to enable people to impart their expertise and knowledge and intelligence into the systems at the level of how we teach it. Yeah, I will note that I was walking through the audience earlier and noted that there are more people coding at this session than texting, which says there's something about all of us assembled. That also says that there are some people with some really good questions, so I'm going to stop and take a second. We've heard that there are some GPUs, provide some significant advantages that lead to better visualization analytics in the database side, and that's having an impact on how we actually start thinking about building some of these applications and where we should be spending our time. Before I get to Claudio, anybody have any questions? Come on, guys, should I just point to somebody? This is your time, smart guys, yes, sir? What are some of the trends that people don't really appreciate in terms of GPU applications? What are the things that are sort of going maybe unseen in this first generation of application development? Please do. So GPUs, they're fantastic silicon, right, and they can do lots of work in parallel, but there are other architectures that make use of modern semiconductors to do things in parallel. They tend to be very esoteric in one-offs, and what GPUs have brought that I think is underappreciated isn't just the fact that we can manufacture at scale these kind of processors, but that there's finally a programming model to take advantage of massively parallel computation that's never been there before, and it's getting broadly adopted, and that lets you build ecosystems, and it lets each of us layer new capabilities on top of that core. That's architectural lock-in, and so that's why while there are lots of experiments with many, many CPUs in a single chip, the GPU as a mature technology really is way out in front of anybody else. I think an exciting frontier going forward, so pleased that we've done a good job at MAPTI of putting together analytics and visualization, but there's an opportunity as well to marry this together with the other types of problems that are amenable to GPU commuting, so deep learning I think is gonna be very interesting. A lot of the algorithms there feed very heavily off of data, and having all those things running together, you could potentially do this without actually moving all the data around, and that would be phenomenal opportunities. Maybe I could offer a concrete example. When you have, this is what you can't do on a traditional database in real time. When you have very high cardinality data sets, like you've got several years of transaction data, you wanna do a group buy based on account number, or you have a product SKU, and you have millions if not billions of SKUs, and you were trying to in real time do traditional database operations. Again, group buys aggregations, and you're trying to spot when an incoming piece of data exceeds by one standard deviation, the average of that. So many times that's been done in a batch process where you try to push it back up to the speed layer. Now you can do those in real time. And don't just do one. Try every combination, try them all. So I wanna turn it to Claudia just for a second because we just heard something and I'm sure your graduate students are spending a lot of time doing. Talk to us about what your graduate students are working on as it pertains to this question of some of the new use cases on the horizon. Sure, so I think that I just like to add one thing to what was said before here. The history of GPUs is also an important thing. It's called GPU because it's a graphical processing unit. So this thing was, it actually has architectural elements that make it different than just a general processor. For instance, there's this Rasterizer. And the Rasterizer, for those that don't, I mean, you have these polygons and you need to find what point in, actually you're doing like pointing polygon computations, right? This was a breakthrough. If you look back at SGI's in the mid 90s, and I remember talking to someone that actually worked at SGI before, right there, yeah. So what happens is that when the texture was actually able to, when they were able to do the texture mapping in real time, there was such a big advantage. And it basically caused a high bandwidth, high computation per pixel, right? So actually what we are doing now when you think about new applications, the applications that are best suited for GPUs are not just using GPU as a faster CPU, okay? So if you think about, I don't know if you remember what I had in my slides there where I was talking about a 6,000 speedup, you don't get that just by jumping to a GPU. You get there by using the GPU in a way that you couldn't do that on the CPU, okay? So some of the work that we are doing is trying to figure out what are operations, in particular, so much of the data we are handling is special temporal, right? So it actually has these temporal components and how do you handle that in the GPU so that it's really much, much faster than if you just make this whole thing work on the CPU, right? And I think that that's actually, maybe someone was mentioning this before, but this is the kind of work that the grad students and us in our group are trying to do. So everybody on stage is part of the process of turning these technologies into tools that eventually average developers will be able to utilize, let me not average, very good developers will be able to utilize to solve really, really hard problems. Mark, why don't you give us a sense of what kind of hard problems you think this combination of capabilities are gonna address over the course the next couple of years? Absolutely, so the scope of problems just getting bigger and bigger. And actually, I think we're gonna continue to be surprised by what people are starting to do with it, but we're starting to enter the realm where control, automation, optimization problems are readily accessible, and they hadn't been as much before, but now with the increased levels of GPU compute that we have, you can start to really tackle these problems. And it goes beyond pure automotive context, it goes into control systems for even urban planning scenarios, it goes into control systems for areas where you're doing industrial automation, where you're mining, where you're drilling wells for oil, all sorts of applications, they're all heavily data-driven. They also can leverage simulations pretty heavily to imply control and optimization. And so you start to see the marriage of GPU compute and CPU compute. And really for developers and for us to see this spread really broadly, we have to allow them to do that in such a way that they're using their domain expertise and subject matter expertise, and they're not becoming experts in SIMD instruction sets and CUDA and everything. CUDA and SIMD, they're wonderful technologies, and we make use of them every day, but at the same time, if you want to extend this to the masses, we need to give them tools that allow them to leverage all of that without having to think about it. We don't think about using the raw assembly instructions on CPUs anymore because the compilers take care of that for us. And in the same way, the database solutions that we're building, the streaming solutions that we're building, the programming tooling that we're building, it makes it possible for people to focus on their areas of subject matter and domain expertise and bring all this technology to bear without having to become experts at the lowest levels. So again, I'm going back to the audience. We've got a lot of very smart people on stage. We've got more smart people here in the audience. Is anybody starting to think about how these technologies are applied to a particular problem? Is anybody doing it right now that wants to talk about it? Or do you have a question for the folks on stage? I'll jump in. We have business today in oil and gas. I spent this afternoon with one of the New York's largest hedge funds doing, well, you know what hedge funds do. The faster they can load, visualize, and take insights on data, the better they do in the markets. Verizon, one of the largest wireless carriers, uses our product today for cell phone analytics to figure out how to give you all better service. So this is not just a, how are we going to get there? Where they're now and just accelerating. And Verizon's also trying to figure out how to deal with 500 million customers now being hacked. Claudia, and then I'll come right to you. So I think that applications where there is a human in the loop, and it actually causes that application to change from a batch computation to a real-time computation, that I think is where it makes a huge difference for people because they can act on their knowledge. So Claudia, I wanted to, just before, very quickly, your professor, you have some very smart students, are you discovering that they are learning these techniques faster because they can iterate on the problems they're trying to solve with the technology faster? Yes, I mean, that's the case, right? So the computers keep getting faster and faster, but now you have kind of like a two orders of magnitude speed up on, you know, so you can, if you can bring problems to a point where the person can actually interactively use a model, it makes a huge difference. Oh, hi, this is, I'm GreenCourse Air Tech Lab, Susan Krause. So this is a very exciting presentation, and thank you to NVIDIA. So my question is, in terms of health care, we have a huge scaling challenge. We want to scale to end users at hundreds of millions within the US. NVIDIA partners or collaborates in China. That's a 1.4 and over population. What do you see in terms of GPUs enabling us to scale to hundreds of millions of users in terms of end users in the US, and then over a billion in China? What do you see in terms of the architectural problems or can, at present level of architecture, can we scale to those levels? Or how many years do you see that in terms of the evolution of the technology? I mean, again, I don't think you're looking off to the long-distance future. We have a partner that's a third-party company that has benchmarked our product 40 billion rows in what, 200 milliseconds? That's not bad. I mean, I strongly agree with the statement that Scott made earlier. The first best thing to do is to put the best damn technology you can in one box. And that technology is GPUs. And then, if that can't solve your problem, then you start adding to it. But I do think some people sometimes wander into the forest when they start out with a thousand-node cluster of poorly written software and hope that it's somehow going to make them happy. I'll tell you our strategy in there from Connecticut. We designed our database from the ground up to be a clustered, horizontally scalable system. So we store data in main memory, and we're very happy about NVIDIA's NVLink and IBM Power Architecture, which lets us pull data out of main memory, feed the GPU very, very fast. And we just have a scaled-out architecture so we can do billions of rows in gestion per minute. And this can scale arbitrarily wide. Yeah, there's a use case that may help to paint a picture of what's very pragmatic today with this new generation of data management tools, where at one of the use cases that we handle is in advertising networks, exchanges, where in near real time, we're ingesting 24 terabytes of event data per day and responding in half a second or less when someone's about to view either a web page or an application page out of the inventory, out of the marketplaces, in real time, what's the best ad to place there to maximize revenue. So this used to be something that they did in batch mode. They did it historically maybe a week later and tried to build some rules by hand to figure out what to do better the next week. Now this is lights out. It's autonomous. And it's working at marketplace scale across the internet. These things have come a long way in a very short period of time. And I think it's surprising to people when you go in. I don't know if you guys have this experience, but you go into a new account or you go to a prospect and you say, we can do x, y, and z with this simplicity. And the initial reaction is, that's impossible. And it's always prove it. And you always go through some kind of proof point. And people, they're just blown away. So I don't think people, I don't think we have done a good job of communicating to the market just how capable these systems are right now. All right, so we're closing in on ending the time, but we have an hour of networking. Yes, Jim? We have one more question. Oh, Jim? Jim? Jim, who is the ultimate host of this, wants to ask a question? No, you can't. Come on, please. All right, so it's a really easy question. We'll finish up with everybody really quick. So starting with developing of the AI application with Bonsai, congratulations on the launch, by the way. Let's go across. How much faster are you than it was previously? Because this is an ecosystem of a rising tide. We're bringing the boats up. We're actually growing this market together. And I think it's really important for us to compare ourselves before there were accelerated databases, before there were rapid development tools. How long did it take? I'll give an example from a PG&E where we're doing a geospatial visualization for them. And what they've told us was it used to take their current geospatial vendor, who I won't name, 20 or 30 minutes to try to render close to a million points. And so what we've brought that down to is many millions of points in sub-second visualization. Another customer I visited this afternoon used to do this quarterly process of reporting is another advertising platform. Same thing like you described. Looking backward, trying to make some rules. It's interactive. A couple of tenths of a second. And they do it real time now. So we used to run simple models and have them take days. I would do problem sets when I was studying at university. And literally, I would set my homework and go do something else for days because you just would have to wait to see what happened. One of my good friends got in trouble because he couldn't wait and ran all this stuff on the JPL supercomputers. And they didn't take kindly to that. But as we've made all these things, the models bigger and tackling more and more complex problems, that didn't go away. So as recently as a few years ago, you would run models and they would take weeks or months to actually run through the computation. And now we measure that in hours and days for the same kinds of models. It makes it tractable to solve problems that before only a small handful could really try to tackle. And now we can start to put that in everybody's hands. So we've done internal benchmarking. There's a pretty traditional benchmark called TPCH, which measures the analytic power of data processing systems. And it's based in SQL. And we've run this. Typically you'll do it at some number of 100 megabytes or tens of gigabytes. So we've scaled this up to near petabyte ranges. And we get an almost constant factor of speedup of 100 times. This is not some made up thing to pump our own numbers. This is just industry standard benchmarks. So when we go into a customer, and this happens quite often, we'll go into a customer and they'll have some very expensive, very large iron database that has gotten them to 20 terabytes. And there's a team supporting it and huge budgets. And now they need to go from 20, not to 30, from 20 to 200, or from 20 to a petabyte. And there is actually no economical way to do that with these existing systems. And we'll come in and we'll do cook-offs with these entrenched, very well-known competitors. And we'll take queries that would take hours to run on this very expensive iron and do them in minutes or less. And queries that would take minutes to run, we do in sub-second. And it really, disbelief is the primary reaction. And you can read white papers. Until you see it happening and the interactivity that you can get from it, I don't think you really get how transformational it is. It's phenomenal. So I would say it's highly application dependent in our lab, but sometimes even almost no speed up to several orders of magnitude. Some problems are very, they lend themselves to very simple solutions in GPUs. Other problems seem to still be very hard. Maybe we have to find different versions of them to work on, but it varies. Okay, so that concludes our thank you, Jim. Thank you, Nvidia. Thank you, all the production team here who made this actually work with apologies to Claudio. But we have some time. We have about another hour left for networking. So this is an opportunity to continue to connect, ask panel members, some new things, talk to each other a little bit more. This is a journey that we're on. And the opportunity here is to find fellow travelers that you can take it with. So enjoy the rest of the time and hopefully nobody's gonna get too bummed up by the debates tonight. Thank you.