Thank you to all of you that are taking some time with us today to talk about what is arguably the hottest and maybe the most important topic in technology in possibly decades. So, I'd like to thank all of the audience, but I also like to really thank my peers andsome of our joint presenters that you're going to be hearing from today. Let me just make sure I give them their credence and their credibility. Kevin Levit, thank you so much from Nvidia director of financial services for joining us. Shayen Patel who is a senior cloud solution architect actually working to implement these solutions with financial services customers here at NetApp and Max Amendy I'm so excited about that a real world live data scientist that is putting these technologies and practices in place every day and is going to share some of the benefit of his expertise and uh and background what he's seeing. So what we're going to talk about today is basically what we are seeing. You know NetApp or you should Fortune 500 company according to uh the CEO Jensen Wong ofNvidia almost half the world's files live on NetApp. We are working with some of the biggest and most demanding customers around the globe in implementing AI and Gen AI projects models and getting results. And we want to share with you what we're seeing. We also then want to make sure that you understand what NVIDIA and NetApp not only have done but are doing right now to enable you to put the power of AI and Genai at your fingertips more easily. And then we're going to get some perspective specifically on the financial services sector use cases, technologies, workflows. And then again, as I said before, I'm so excited that Max is going to give us a littledemo of what he's working on, things he's seeing, and then we'll wrap up with your questions and answers. But I started out by saying I want to share with you what we are seeing. So what you see on the screen is uh you're in the financial services industry, most of you on this call. So you are either proud or paranoidbecause you and your peers meaning your competitors are leading the way. We just did a study released yesterday. It was the NetApp cloud complexity reportand loud and clear throughout that report. You'll see a link to it at the end is the emergence of AI as thedominant workflow and use case that customers are struggling with right now. We sliced and diced how different sections of the world, different industries,different size customers are approaching AI. Congratulations to banking and finance for being a leading category as opposed to some others. For example, media and entertainment. 55% of the banking and finance customers are using AI projects right now. They're up and running or they're close. Like I said, either you're going to be proud or you're going to be paranoid because your competitors are doing it if you're not. And the bigger they are, the more likely they are to have an a funded project underway to see results. And results is what they're seeing. Andspeaking results, I wanted to touch on something that has emerged in about the last six months. It was defined in 2020, but it's really coming strong in the last six months. If you're working in AI and specifically in generative AI, the kind of things where you have natural language proc uhprogramming and you'redealing with customers through chat bots or you're summarizing and analyzing reports or you're doing what if analysis and trying to collate lots and lots of data. Genai is one of the first stops on the AI journey that most of our customers are having andthey have really begun to adopt a mode of operations called rag or retrieval augmented generation. You see here the definition by our partners at Nvidia how they define rag there's they're not the only ones. Oracle, who's arguably a very substantial player in the data space, also says, "Rag provides a way to optimize the output of your LLM by using targeted information, even if the underlying model and data to train it didn't contemplate that information. Microsoft gives you direction when they say the rag architecture means you can also constrain the generative AI model and learning to your enterprise content." Really what this means for those of us atNetApp and the customers we're working with is that you can take an existing large language model so you don't have to buy or build your own. You can take a pre-trained model that is in the same context area and not have to train it on your data which you can do. Fine-tuning is really important but rag allows you to immediately inject your data into a prior existing model and get relevant results for you. And in many cases it's greater relevancy than you would have otherwise. Consider three financial institutions that are going to launch a financial services product in Asia Pacific in Q4. If they all ask the same model, the same question with the same publicly available data, no matter how hard you work to find it, and the model's been trained, they will all likely get the same answer. But if one institution can say, "What has our rollout been for specific maybe this is an investment banking product and what has the adoption rate been from our customers and what have we learned and where we can price relative to the market?" Whatif we also include to that some of our customer surveys and even our support calls and chatbot calls as to the things that customers struggle with most that institution is likely to get a far more targeted answer than the competitors and that means velocity means moving faster with greater force and that's why 82% according to IDC of enterprises surveyed are really looking for gena models that leverage their own business data in an easy secure and replicable manner. That's the promise of ragg. So think about what you're dealing withAI and Gen AI. You know that your data is everywhere. You might have a Tokyo data center, a Dallas data center, a London data center, a Singapore data center. You may have cloud accounts in Azure East West, all over the place. You've got data everywhere. And by the way, it's unfortunately for us, it's not always on NetApp. According to Jensen, it's about half the time on NetApp. But you might have on other footprints. It might be in Azure Blog. It might be in SharePoint. What if you could use all of that data to answer your Genai queries in a way that's simple and secure and easy? That's what we're working to bring together. What if you could have that data across all those environments automatically classified, categorized, and tagged both according to metadata and data levels, owner location, category, content, reax expressions, certain key phrases, business, departmental, and active directory group level. you. What if you could get more relevant answers than your competition? Matter of fact, what if you could do this as a matter of course? What if you could programmatically have Genai workflows, rag workflows that have this built in to categorize that data and pull in just what I want at that time from that query and then get a little crazy with it. Be able to look at I'm going to ask it a question from different points of time from today, from yesterday, from last week, from last month and see what the variance is. That's AB testing. But what we find with our customers is they're doing AB CDE E FGA. They're doing ABN testing or what if they could look at it from different angles andprompt it differently. You you've all seen what's happened with prompt engineers and the explosion of that uh that particular skill set. So what if I could ask the question different ways to the same data. But if I have 20 different ways I want to ask that question, I don't want 20 different data sets. That that's incredibly expensive. What if one data set could be queried simultaneously in these different ways and do that in such a way that I don't have massively redundant data sets that I don't kill my staff that I don't crush my bandwidth moving data back and forth and back and forth. Those are some of the problems that we're addressing here at NetApp and with Nvidia. Just to get to give you a perspective, we're working with our customers on Gen AI and rag workflow operations that allowed them to have a common data footprint everywhere. Different data centers, different cloud regions, all with a single control plane and a single way of looking at and managing that data. And then having the ability to have that data automatically scanned, classified, and tagged again according to metadata, context, even data content. Once you have that, you can have just the right information pulled into a volume which you mount into your inferencing stages. And then you can play some games with timebased queries of that with different flows, what I call ABN, at different granular time intervals. And of course the big thing is if there's an issue with the production or the infrastructure or the network, you don't disrupt or lose data because you're never touching the production data. And then if you could in parallel ask different queries to the same data in different ways. You do that if you can have the ability to have space efficient clones. I had 50 pabytes of data. Only one pabyte or only500 terabytes were actually germanine. So I pulled that down and then instead of making 20 copies of 500 terabytes, I have one, but I have space efficient clones that take almost no space and I can have 20 queries going at the same time and then blow them away. And even better, I don't train, I don't retrain or fine-tune a model. So my data and its context stays relevant only to me. Finally, if I have Nvidia NC or NV uh V uh CPUs that I want to use up in Azure, but my data is somewhere else. What if I could simply and easily take that data and extend it into my Azure NVIDIA enabled environment to be able to be processed without having to move so much data because I still want to under centralized control, especially if I'm dealing with um secure domain and sovereignty regions. Those are the kind of things that we're working on with our customers. Iknow I hit you with a lot. So my first tip for you today, I would make sure your teams are up to speed on retrieval, augmented generation, or rack, how it could enhance your Genai efforts, what the kind of workflows are that would benefit and how you would put that into place for your Gen AI workloads. If you do that combined with the power of what our partners at NVIDIA are about to walk through with you, I think you're going to see that there's an amazing world at your fingertips and you don't have to be paranoid anymore because you're stepping into the proud leadership position. With that, Shayen, you have control. So, I'd really like to invite myfriend from Nvidia, Kevin Levit, to come share his perspective and background with you set up. My role today is going to be take everybody through a little bit more of a focused lens from a financial services perspective on what we're seeing uh within the industry as we work with leaders across uh banking, trading, and payments. I love that uh Chuck not only focused certainly onrag and its importance, but also the critical element that data plays here. And so when we did our survey late last year around, you know, what are the biggest challenges to financial services uh companies and achieving your AI goals number one was data issues related to privacy, sovereignty and disparate locations. Uh so a great topic and it's also paramount uh for why our partnership with NetApp is so important on the next slideshow and if you don't mind and that is you know obviously data is what fuels all of our capabilities when it comes to AI and ML uh you know so many tenants between our partnership resonate culturally you know very focused uh on solving customer problems uh very focused on enabling an ecosystem and having a suite of uh solutions ions that meet you where you need to be, how you need to be met when it comes to your data, storage and AI and ML capabilities. Uh so let's get into the next slide and just do a quick background. When we think of the role that Nvidia plays particularly within financial services, it is very much about powering AI factories. Uh in financial services, the way that companies differentiate and compete is truly on data and insights garnered from that data. And every one of your firms is actively generating intelligence 24 by7. how you integrate that intelligence into chat bots using rag into other customer service experiences into underwriting into uh doing a better job of risk management uh cross-ell etc will all be powered by AI enabled applications and you need a partner that has the ability to surface an AI factory that supports enterprise transformation in AI across every line of business and that's where we focus in working with our financial services clients. On the next slide, what it means to be an AI factory is very much having a full stack platform. And hopefully you've heard this from us time and time again. It is not simply about making amazing hardware and networking capabilities at the foundational hardware layer. is very much about the software that sits on top of that hardware which is able to maximize the performance and utilization of that hardware from a operating system software perspective all the way up through the application frameworks that enable your developers and data scientists to get to market faster with AI enabled applications because our leading customers in financial services today are operating against hundreds of use cases where they're going to be leveraging AI and ML in the future it will be thousands. So if we turn to the next slide and take a little bit deeper dive into the use cases as it pertains specifically to financial services and generative AI. You can see it's a range of use cases. everything from AML and KYC and transaction fraud for payments uh to high performance computing related to pricing and risk and summarizing news feeds and everything that's coming inreal time to organizations to help balance uh risk portfolios and then uh creating new strategies as it summarizes that data for algorithmic trading and uh identifying alpha from a portfolio management and investment standpoint to intelligent data processing and document management uh to customer service chat bots certainly creating tailored marketing uh collateral and imagery to help with cross sales and acquiring new customers to certainly optimize portfolio strategies generated uh through generative AI forwealth management. So that's a quick overview of kind of wherewe see the priority use cases today for generative AI in the context of the financial services ecosystem. Uh on the next slide, just wanted to talk about our large announcement coming out of uh our GTC conference last month, which is making it easier to deploy these generative AI capabilities through inference microser microervices that are custom created for specifically generative AI workloads. What's there a couple of things that are really important about these uh NVIDIA inference microservices? Number one, they reduce your time to deployment from weeks to literally minutes. Uh, and they are available everywhere across the entire ecosystem. They can operate uh on Nvidia's platform regardless if your infrastructure is onrem, in the cloud, multicloud, hybrid, etc. Uh, so I encourage all of you to go to ai.invidia.com nvidia.com where you can actually interact and experience uh these NVIDIA NIMS or NVIDIA inference microservicesuh for yourself. On the next slide, I did want to expand the lens a little bit beyond pure generative AI as a workload and the use cases associated with it to talk in a little bit more detail about the wide variety of use cases where we're seeing AI and ML being utilized across every function really within a financial services company whether it's on the financial revenue front in terms of quant finance underwriting and analytics and utilizing alternative data or from an operations standpoint where AI and ML enabled applications are removing significant costs, eliminating redundancies, helping fight fraud, improve the claims processing for insurance companies, and again uh enabling intelligent automation particularly for document management and then certainly within customer experience, you're familiar with conversational AI and uh natural language processing forchat bots and certainly improving the personalization and one to one communication with Europe end customers to improve customer loyalty, drive cross sales, etc. So this again is just a sampling of kind of those hundreds of use cases where AI and ML is being leveraged today across financial services. I did want to touch on the workflows on the next slide associated with some of these use cases. And for the next couple slides, we'll just focus on transaction fraud, which is really one of the easiest places to get started uh within uh banking andpayments organizations is leveraging these new techniques to improve fraud detection accuracy. And what we're demonstrating here is through the transaction fraud AI workflow. And this really pertains to any AI workflow. There are three critical steps. There's the data factory or the data preparation. There's the model building. And then there's deployment. And across all three phases, you're going to find software that enables your data scientists, your data engineers to improve the productivity of their AI and ML pipelines such that you're removing costs, getting greater productivity out of those teams, delivering more accurate models into production faster, and delivering a better experience when those models are at the point of inference such that they meet latency thresholds. s and deliver the customer experience that you'reexpecting to enable. So that's a highle overview of the workflow that we see in AI and how NVIDIA's software is enabling acceleration across all three of those phases. Wanted to give you some real life customer examples on the next couple of slides. So if we go to the next one where we're highlighting a blog that PayPal recently released about their use utilization. If we can go back one slide, their utilization of NVIDIA Rapids for data preparation. That's where they're actually saving 70% on their data preparation costs. And then for American Express, we're able to highlight across all three phases in terms of training and inference and data prep a significant improvement in the training time. Uh an improvement in the actual performance of the fraud detection accuracy where the model uh reduced uh false positives by 6%. And then we were able to hit those latency thresholds of literally running these more sophisticated deep learning models at less than 2 milliseconds. Uh so another proof point with American Express here. And then the last slide uh that I was going to cover just brings us back to the concept of an AI factory and that is regardless of where your financial services company sits within the ecosystem whether it's trading, banking, payments, insurance, etc. You're going to have a range of business drivers from improving fraud detection to building virtual assistance, improving cross-ell to your existing customers and there's going to be a consistent set of AI infrastructure challenges. It is complex to build this infrastructure particularly to make it productive at scale and of course we're all operating within a given budget and a total cost of ownership threshold. Well, when you operate and partner with companies like Nvidia and NetApp, you know, we're able to deliver this full stack platform that again runs all the way from the hardware layer up through the financial application layer meeting the demands of your variety of internal stakeholders uh from certainly the MLOps folks up through the data scientists and researchers ultimately resulting in faster time to insight, greater productivity at scale and the right ROI for your organization. Uh so thank you for uh supporting and enabling uh Nvidia to be a part of this conversation. With that I'll pass it over to Shayen uh who will take us through the next set of slides. Thank you Kevin. Thank you Kevin and thank you Chuck. As you heard, as you have heard from Chuck and Kevin from that NetApp, Nvidia and Microsoft together, they're building great solutions uh to help overcome some of the current challenges uh in a IML workload and I have some of the examples uhexamples of some of these solution and we we'll get right into it. Um so in this particular slide I wanted to focus first on the at a high level on the high level architecture of the of this particular uh slide. Uh so typically when customers are deploying a IML workload in Azure they're either uh using native Azure AI services uh Nvidia power GPUs or any other thirdparty marketplace a IML platforms. Uh most of these IML workloads have similar profile um meaning high performance requirement, large capacity requirement, often times it turns into high file counts which can result into high metadata and that becomes extremely difficult to move, migrate, replicate uh these data sets and this is when the hyperscalers like Azure comes into the picture when who can provide the infrastructure as on a needed be as needed bas phases uh with the help of netup storage this becomes even better. So if we focus on from a data in integration perspective u with netup storage now you can integrate all your unstructured data from various sources into a centralized dep repository. Um this helps supports the analysis and model training and as Chuck mentioned in his uh overview that we do need to know what our data is what are where our data sits which data is and so financials have that challenge and wekind of help you u get an idea of where the data sits u from the scalability perspective as these a IML applications scale it becomes challenging to manage these large data sets efficiently uh with high performance as well as uh need with the need of the cost optimization and netup storage services canhelp you with that as well. Um other important point that I want to point out about the GPUs Nvidia power GPUs availability as uh youguys are already familiar with it every organization and as Chuck mentioned 55% of the financial services are actually already deployed or inprocess of completing the deployment in um uh art artificial intelligence. So uh and applications associated with that. So with that high demand Nvidia power GPUs are very high in demand especi andpeople are whether it's onrem or whether it's in hyperscaler customer can run into the shortage of GPUs in the Azure regions where they want uh and this is an important factor that using netup technology we can move the datasets where the GPUs are available so for example H100 which is the topof line and media power GPUs. They are only available in certain Azure region and extremely high in demand. So you may not have GPUs available in your primary Azure region and you may wanted to move your data sets or replicate your data sets to the Azure region where you have those NVIDIA powered GPUs available. So with the netup technology youthat helps withthis challenge replicate these data cents efficiently to the Rajure region where you have these GPUs available from the when wecan also use collaboration tools like Azure machine learning studio uh netup dataops toolkit or any other Azure automation to facilitate these collaboration between the data engineers data scientists and other stakeholders of the organization uh to come together with this. Um on top of this if you look at on the right side of the chart or the bottomportion of the architecture slide um for services which needs persistent storage like Azure Kubernetes service Azure data factory Postgress machine learning uh you can use Azure net file which is a firstparty Azure Microsoft service which provides all these benefits ofall these benefits of the storage uh that the net of storage thatyou are you have seen. Uh so it's extremely important to manage the data sets as well as the metadata along with these data sets for this workload. So simplified metadata management it is the biggest advantage of using NetApp storage and uh for the a IML workloads. Um the way we do this is if you look at some of the areas where we can help uh we can help implement metadata management practices to track the data lineage versioning we can help with the versioning which I think Chuck mentioned earlier which is very important as well withthese workload. Um another important thing is the cost optimization as your as these datasets are getting larger and larger u automated tiering of the cold data sets which you don't no longer need it's also important and that helps you with the cost optimization. So knowing the right data sets whichyou really need for your uh training and other uh other purposes you it's important with our automated triing you can help it helps achieve you uh get you that uh what volume cloning ability to make the replica of existing data sets almost instantaneously. This also goes by goes with ability to make the replica of a previous version of a data set. It's also extremely important and we could do that allthese it's almost instantaneously. We also have a support for snapshot based backup which is instantaneous no impact to application performance and the it'sreally backended by snapshot technology that NetApp has. Uh so basically uh you are you're getting the backup instantaneously. There is no application performance need impact and we you have an ability to instantly restore the full volume file system or directory or at the individual file level. Uh and one of the other important part withthe financial services is the business continuity or disaster recovery inthat manner. uh so replicate you can replicate these data sets to another region for disaster recovery purpose as well asI mentioned earlier where the GPUs are available. So these are some of the great benefits of the data management capability that net brings to the table to help you uh get more out of your existing data source or data sets. the inthis next use case I want to discuss the integrations with integration with Azure machine learning studio. So data scientists face several challenges uh today uh they need to have the access to high performance persistent data volumes to train their machine learning models while also needing to protect these data sets. They work with large data sets as I mentioned. They also need to be able to create the exact replica or previous version replica uh versions of existing volume anf coupled with Azure machine learning studio can also provide you can provide you the functionality that's required by today's data scientist u the integrations of Azure machine learning studio with anf is possible in several ways um few the scenarios one leveraging the high performance uh storage for AI model training task another one is provision provisioning ANF volumes with Azure machine learning notebooks for the data persistency and protection. So point is you don't really need to be a storage expert to use and utilize the NetApp storage services. Um, with that I wanted to u hand it over to my colleague Max uh Maxe. Uh, please take it over. So, good afternoon. Can you all see my screen?>> Yes, >> Yes, >> Yes, >> I think so.yeah.my name is Maximandi. I'm a data scientist and technical solution specialist here at NetUp. And before I came to NetUP, I worked in data science for a major European bank. And I think as we heard from Nvidia before, one of the lowest hanging fruits in AI for financial customers is fraud detection. Why? Because usually it's easy to obtain a high quality data set internally since all of the banks are already logging their transaction anyways. The data is usually in a nice to process tabular format. So that it's relatively easy to process the data and work with the data without crazy amounts of cleaning required. But nonetheless um yeah fraud detection comes with quite some smaller but also bigger challenges. One for me as a data scientist was always the challenge of the imbalance in the data set. What I mean with that is that in with those transactions you have waymore valid transactions than invalid transactions. If a data scientist would simply predict um or simply build a simple model which just predicts whether a model is or whether a transaction is true or not. The easiest way would be in most scenarios just to predict everything is valid and you would have amazing scores then but that would be far away from the goal of a task. Of course the task is really to find out thefraudland um the fraudulent transactions but don't worry I will not talk too much data science today instead I will focus on the following three points in this talk today. First, I want to demonstrate why snapshots are really valuable tool for data scientists and how they can make the data set versioning just so much more easy and efficient. Next, I want to talk about how we can really integrate um Azure netup files and workflow in Azure machine learning studio. And last but not least, I want to show you the performance and like really the speed of Azure files when it's used as a base for training jobs. So what further should we do? Let's give it demo time or let's it's demo time now. And I recorded the demo just a couple of minutes before the presentation today. If you look on the bottom right of the screen, you actually see I'm I recorded the demo for most of you in the future since I'm located in Europe. Um, but we just did that as a safety reason. And what you currently have there in front of you and what I'm going to show first is simply that nearly all the steps we're going to do are easy to do over the user interface of Azure. Because Azure net of files is a Azure native service, you do not need to go over the marketplace to provision it. You can simply access um access the sources right from the user interface. And therefore, the first thing we're actually going to do today is let's have a look into Azure Netup files. And if it loads, we're going to see that I already created uh yeah, a volume or no, not a volume, uh registry. an account for it. We created a resource group and a capacity pool for this demo in the US region, west US2 region. And if we have a look into it, we see that we provision a capacity of overall 4 terabytes of data, which is quite a lot for what we're going to do. But more importantly, what you're going to see is that we provisioned the highest speed class since speed is key for training AI models. We see I provisioned the ultra glass the highest we have with NetUP. I configured the protocol of NSF and ELP NS 4.1 and we see that we have a throughput of half a gigabyte per second which really is a lot but what I personally as a data scientist really like is how easy it is to do all those steps. For example, a net up files or um gives you an easy mount command so that evennot sometimes the best infrastructure people like me can automatically mount the stuff or easily mount the stuff. But what for me really is the amazing feature here in this step is snapshots. Since one of the key challenges for me as a data scientist was something called data set versioning because in my task in my job back then I worked about a year on um on a big project in the bank and in that year the data and the data coming in changed quite a lot of time during the process. Meaning if I would have worked the whole year with the same data set, I would have had huge data drift or model drift at the end of a year. Meaning my model which had been optimized very good for all data would have been basically worthless at the end of a year. Most what we data scientists tend to do is we take every couple of weeks, every couple of months a new complete copy of a data set and then only continue working with a new data set. But nonetheless, we have to keep the old data set in place since um we have to often back test um with which data set which code comes to which regions to which results. But complete copies take a lot of time and resulted in quite some nice coffee breaks for me as a data scientist. As we just saw on the top, instead of using a complete copy of a data set, what you can do is hey, simply take a snapshot out of it. a snapshot no matter how big the data set is takes couple of seconds no matter like I said if it's 14 15 GB like we're having now or 14 15 terabytes couple of seconds and with a snapshot in place I can at every point in time jump back and kind of um and kind of see with which yeah with which results with which values with which data which code of me resulted in which yeah results in AOC values or will help AOC scores or F1 scores as we're going to see later on. And with a snapshot I can and especially with a snapshot in Azure measure in Azure I can restore the files every second I am very easy from the graphical user interface I can also create a new volume with the values or with the data from a um yeah from interface or from the snapshot veryeasy and very efficiently. But efficiently. But also take in mind I'm just doing the snapshot via the GUI. You of course can do the snapshot also via the command line which would be most likely the better choice for the data scientists but for a demo just looks nicer. So what we're going to do next is let me jump a couple of seconds back. We're going to go into that's what I think most of you all are here today. Let's see or what the most of you are interested in here today and that's if we have a look into um the machine learning and how really NetUP can speed up that process how NetUP how if you use Azure Netup files if you um you can leverage your GPUs even to a further extent and this what we're going to do next is we're going to open Azure machine learning we're going to see of course I also prepared all those steps ahead of time to make sure that everything works in this recording. What is no opening up is Azure machine learning and like the user interface the data scientists will use most of the times and which is most likely provided to the data scientists themselves. We see all those awesome functionalities here for the data scientists like the opportunities regarding rag which Jack talked before about and which makes all the steps very easy. But for this demo today I want to focus on something else. And first we have a look on the bottom left here in the commute in the into the compute field and have a look into what compute I already provisioned. We um I provisioned basically two different types of instances on the one hand side compute instances and compute clusters.But here what is here the big differences is that basically the computer cluster that's kind of a share medium that's where we have the big Nvidia GPUs where we have in our example A100 GPUs with crazy literally crazy amounts of power available in there but data scientists should not work directly on the GPUs since a lot what they do is data wrangling data preparation feature engineering with all those steps you don't really need gigantic amounts of GPU. What you usually do instead isthat you give the data scientists, the data engineers instance to separate compute instances with low amounts of compute. Sometimes just some CPU cores and memory inbest case scenario of course also probably a smaller Nvidia GPU and a T4 GPU for example which cost significantly less than um a A100 or H100. But even with a T4, even with CPU own or with a T4, you can leverage the amazing functionalities of QDF and Rapids like Nvidia was presenting or telling us before that. And if we have a look into the working environ or we're going to now have a look into the working environment for the data scientist where most likely he will look spend the m most of his day either usually a Jupyter lab or a Jupyter notebook. If it opens up we have to give it a couple of seconds. Um we're going to see here a Jupyter lab. It's just the working environment for the data scientists. Yes, I know with Visual Studio and etc. It's good to work but me in the field usually till this day I see Jupyter notebooks and Jupyter labs due to the simplicity for a data scientist to work of work with and I think to be honest it's just an amazing working experiment since it allows rapid prototyping and I think we all know AI data science at the end most of the time it's a combination of prototyping and experiences combined into the best possible Okay, I will quickly stop here because I think in the recording I went over that too quickly because if we have a look I just jumped a bitback in time on the left we have on one hand inside our script our Jupyter notebook script and we have a folder called data. The data folder is m is um is hosted on an ANF. In preparation I did run through an automated script to mount that um data folder directly to my working instance in um yeah in Azure machine learning studio notebooks. You can also fully automate that for your data scientist so that every data scientist can easily access the data on that share. Also I did run some additional performance improvements in there for mounting command. Not to get too technical at that point, but features like endconnect just really allow you to get the last bit of juice out of um out of Asia files. So really leverage it to maximum and therefore getting amazing amounts of performance in here. If you have a look quick look on the data set, we're seeing it's about 14.3 GB large, which is not crazy big, but still not the easiest thing to process. I think we can all agree on that since if you work with the standard environments usually you have to load all the stuff into thememory or VRAM of your GPU or memory of your system. So 14 gigs is not too easy to process that in that steps but let's have a quick look on the script which I wrote. Don't worry I will not get too technical. At the beginning, we're going to import all the libraries we're going to use and we're going to start um and we're going to create a start timer which at the end we're going to use to calculate how long the job really took, how long the machine learning part took. Then we're going to read in the data set um and going to do some simple data selection like select what rows are useful, what we really need. We're going to normalize um the data and scale the data basically to a standard set of values and make the columns the categorical columns more easy to work with just data science stuff. Next, assumably the stuff we data scientists love the most, feature engineering. Since here it really is a combination out of all the skills and knowledge we have and experience to generate um features which really bring forward the algorithm. For example, we might know here that um during the daytime um or during business hours where it's happening less frauds than during the night was based on the times of a transaction we have. We're creating now a feature. We now creating a feature which is basically is a daytime and was helping the algorithm to better understand what we are actually doing here and helping the algorithm to produce in here better results. Um but afterwards we just continue. We're just going to drop somefeatures which we know from experience are not valuable but are in the data set. And next we're going to feed in all the data into our algorithm. We're going to use XG Boost. Personally I'm a huge fanboy of GXG Boost since it's a simple and quite fast algorithm. And for me the best part of it since you can imagine it like a treebased algorithm. You can really understand what the algorithm is doing. While with deep learning it sometime is a bit difficult. At the end here, what we're going to do after the model has trained, we're going to calculate some numbers like I said or some metrics like I said, not only of accuracy since accuracy alone for throw detection is not really valid, but also that font score, which really can help you understand what results we're getting out of here. And at the end, we're going to print the time to see how long now everything really took. With that done, basically the data scientist if he has finished the script um he would most likely not train the algorithm directly but instead get back to the draw not to the drawing board but back to the console and schedule the finished code into the training cluster into the big guys um GPU in the big guys GPUs um into the A100s. And I kind of simplified the demo here and don't go too much into detail but what I did is um at that point in time is like prepare a working environment so that it's easy to sk to schedule the um the job also I included NVD reps and QDF andmade the um XG boost run on the GPUs to get the best to [clears throat] get the best possible performance. And now I accidentally jumped back a bit. One second. Ah, that was a spoiler. No idea what just happened. [laughter] Um,so basically here we just [clears throat] [clears throat] [clears throat] basically what we're going to do here is we're going to schedule thejob at that point in time. We simple command line. Um, of course the copy pasting in went a bit wrong so I had to correct it by a bit. Um but basically one line of code we data scientists that's even for us easy to schedule the job using that way and what is kind of cool of our integration here in Azure machine learning studios that as soon as the job scheduled we can jump back into the GU into the graphical user interface and yeah I'm already see here but it automatically says the job has started. We can jump in to the jobs menu and see here all the um that the new job got scheduled for execution and now it's running on our cluster. That might take some time. So I will quickly show you something else. What we're going to see here is basically where it's running the status out of it on which cluster it is running. Here we see it's running on the A100 cluster with crazy amounts of compute resources. I started data science 5 years ago. It was it's unimaginable how fast those A100 and especially H100 are. It's crazy. By the way, all the automation I did in the background is available on the GitHub repo so that you can easily clone it and do it yourself. Also noteworthy I decided to do all the stuff today using Azure machine learning studio. Of course you can use your own MLOps tool or use AKS for it. We can integrate in all those stuff easily. Just for demo I think me a machine learning studio is the iskind of the most beautiful way to demonstrate all those steps. Let's jump a bit forward. And um I think I jumped here a bit too far, but basically what we did get informed is that the training job hasfinished. has finished. And we're going to see here we have accuracy of 99.8%. We have F1 score of 99% which is really good. But what for me at that point in time really is the mind-blowing thing is we trained a model with nearly 15 GB of data in less than one minute back when I was at a bank training something close to that would have taken at least an hour probably two hours but leveraging Nvidia Rapids leveraging um the GPU version of um of XG boost is just crazy fast. But that is also like so fast that another bottleneck comes inmany of the customers I'm talking to and that's the data loading since if you have those monster GPUs which are processing so much data um like previously Nvidia said that um that PayPal was able to I think reduce the data wrangling time by factor 7 a new bottleneck often appears and is that's loading in the data to the GPUs loading in the data to working environment and that's where we as netup really came in with ANF since we as NetUP are four times faster than Azure files. So Asia Netup files is four times faster than Azure files. Meaning if now due to all the speed of Nvidia the data loading at some point immediately um takes up 50% of the time we can reduce that 50% of time by huge margin simply due to the sheer speed of ANF due to basically we can remove that bottleneck and allow you to leverage your GPUs to complete other level which I think can save you a lot of money in with um with your application and can really help you in your daily lives. But with that, I'm over with my demo today and I hand it over back to Chuck. >> Thank you so much, Mac. Iappreciate it. Number one, having a real world data scientist who understands and loves this stuff. I think you did a great job. We're getting short on time, so if you have any questions that you would like us to answer, put them in the chat. We have a few of them so far or in the Q&A. And look, let's just wrap up with this. Um, right now what's really important is what Andy Grove, former chairman of Intel, said when he said, "Only the paranoids survive." I I'mvery familiar with companies from 10,000 employees down to 10 employees that have made it a strategic imperative to employ AI in the next one, two, or three quarters in order to keep up, much less get ahead of their peers. Finding a way to do that, your industry is leading the charge. Number two, pull out the big guns. Work with the companies that are defined stated leaders in this space. And by the way, I I'm proud thatincludes NetApp as well as Nvidia and Microsoft. We've harnessed the power of the most powerful data systems, plus data management, the ability, as we talked about here, to look at all of your data, have it automatically classified using AI ops, allow you to put that into a volume, which you can then move or extend to another part of the world and put at the fingertips. the most powerful AI GPU processing environmentsand get your results ahead of the competition. And number three, move fast. Genai, if for those of you that are early in the cycle, Genai is actually a really good starting point. And if you understand rag operations, you can take advantage of it without having to build your own model, without even having to train your model. actually without even having to fine-tune your model. All of which takes some degree of time, money, and resource. Andwe're going to find that AI, Genai, and Rag operations inserted into that are going to become a great equalizer. Those of you that are further ahead, as I show you showed you earlier as we got into this, the large customers with large budgets and large resources, they're already putting those stakes in the ground. But Nvidia, Microsoft and the tools that we have provided here at NetApp are going to offer a great equalizer. So look into some of these use cases. We are happy to talk to you about them after this. Emma, did we have any questions come in that we need to address for our audience? >> Wedo have a few that I'd like to address. Um and the first one goes to you, Chuck, and to Kevin. And it sounds as follows. um say our company isn't ready for Gen AI. Is there an easier way to get started with AI? >> So I I'll give one perspective on that because uh not only amI here with netappa fortune 500 company with 12,000 incredibly talented people, but we work with in some cases startups with 10 12 15 people." Um there's a number of areas in which you can begin the AI journey. Many customers find Genai to be the easiest on-ramp due to some of the existing models that are available, existing tools that are available, and it's often lower risk use cases. Getting comfortable and familiar with the AI paradigm by using Gen AI to do things like document creation, presentation, summarization, and analysis, and customer-f facing chat bots. They bring a huge return with only a little bit ofinput into that. And most people are finding out that they can have um a small amount of their resources in terms of people or infrastructure dedicated to that. And youmay find that people are raising their hands to be your in-house expert. Um I would also say that you may also want to find some of the Azurebased training to get uh up to speed on this such as there's one that is the Microsoft certified Azure AI fundamentals course. That's also a great place to start. Kevin, did I miss anything? The only thing I would add is as you get started, two things. One, make sure you're clear on your business objective and your KPIs for measuring the success of whatever initial pilot you might start. And then when we talk to customers about finding a place to start, pilot rather than PC is what we tend to operate around. Think of a pilot as a execution and miniature. So you want to build your pilot in the same way you're going to go to market at deployment so that you don't find success and then have to rearchitect your workflows, your pipelines, etc. So then you can deploy. Uh so just make sure you're building for the long term as you go through the exercise of finding those initial use cases, understanding what your KPIs are, and then executing a pilot. >> Great input. >> Thank you both. I know that we're up on time now, so I understand if someone has to drop, but I want to keep Kevin on the spot for one more second if I may for another question. And the question is, are banks actually using Nvidia software for AI? And if so, why? >> Yeah, absolutely. Um, just quickly, you know, a lot of banks are looking for redundancy and resiliency in their infrastructure and that extends beyond the hardware to their software as well. And because NVIDIA's software operates everywhere and every CSP onrem, hybrid, multicloud, etc. getting your data scientists operating on one standardized platform is of great value. And then with NVIDIA AI enterprise, you know, our suite of microservices, all the application frameworks come with uh this is stability, security and enterprise support that financial services companies and other enterprises demand. Uh so NVIDIA AI enterprise has also been well adopted across financial services. >> Thank you Kevin. And with that I'll say thank you so much to the speakers here today. Thank you to everyone who attended. Uh if you still have any questions I'd highly encourage you to reach out to any of the speakers here. We will also make the recording available for you after. And yeah, with that, thank you and I wish you a great rest of the day. >> Thank you everyone.