NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Welcome, everyone. On behalf of Red Hat and NVIDIA, we welcome you to this interactive webinar, and I'm excited to have you join us. Today, we will cover the topic of Solving Unstructured Data and Real-Time Analytic Challenges in Financial Services. I'm Aric Rosenbaum, the Chief Technologist in Red Hat's Global Episode Practice and will moderate today's session. We're joined by Prabhu Ramamoorthy, Global Partner and Customer Manager of Financial Services. He holds a CFA, FRM and PMP certifications, and previously worked at KPMG and EY as a director. He's helped 100-plus institutions in AI/ML transformation projects and works on use cases for NVIDIA acceleration. Additionally, he manages the NVIDIA ecosystem where he works with clients and partners. We're also joined by Marius Bogoevici, Chief Architect in the Financial Services team at Red Hat. Marius has more than 20 years of experience in open source as both a practitioner and a leader. As a technology strategist and AI/ML expert, he's helping the top 30 banks in the U.S. and Canada apply artificial intelligence as part of their digital transformation journey. During the next 60 minutes, Prabhu and Marius will jointly present on how enterprise financial service firms can solve unstructured data challenges and accelerate time to value using a solution based on Red Hat and NVIDIA technology. [Operator Instructions]. Marius, I will hand it to you to start.

Thanks a lot, Aric. So let me just walk you quickly through what we're going to talk about today. This presentation comes from our experience of working with customers and for our strong conviction in the power of unstructured data. We're going to talk quickly about how Red Hat and NVIDIA work together to help customers build their AI platforms. We're going to tap into the power of unstructured data, and we're going to share why we're excited about and what solutions we propose for customers to accelerate the implementation of business solutions based on that. And we're going to wrap it up with a sample solution, a deep learning-based document analysis developed by 1 of our partners.

Thank you, Marius. As of November 2022, NVIDIA is in the top 10 listed company in U.S. stock market by market cap BBA leadership. Here, I'm going to highlight the fact about how we are revolutionizing this space. We are referring to the fact that Moore's Law is dead, and Moore's Law referred to the ability for chip's advancements to deliver twice the performance at the same cost or at the same performance half the cost every year. And quite frankly, this is not progressing as we expected. And here comes NVIDIA as the answer where we provide the heterogeneous compute layers where we pair the CPU with GPU. And this has been the key to AI's success, where NVIDIA shines on unstructured data and real-time analytics, which is the topic of this webinar, as well as you see that we have revolutionized both the hardware and the software world, where the GPU parallelization with the software such as CUDA helps here. And we invest a lot in R&D and research, where we have close to invested $5.3 billion in this space. The end of last year, we had shipped over 1 billion NVIDIA GPUs and along with our CUDA software for over 10 million downloads. We have an integrated hardware plus software that's covering those 3 areas for NVIDIA HPC [ Quant Finance ]; NVIDIA AI Enterprise, which is our framework for NVIDIA AI projects; and as well as NVIDIA Omniverse, which is our project for building the real 3D world, where you can simulate existing worlds as well as the unreal world. You can design the next Doctor Strange Madness of World. And this integrated software layer with HPC QuantFinance, AI Enterprise and Omniverse enables you to build the cutting-edge software applications for the future. To name a few at the top, we have these applications, NeMo neural net modules for building conversational AI; RIVA our tool for building conversational AI applications; Maxim, a tool for augmented reality; Avatars, which can help you converse with your customers, integrating NVIDIA AI Enterprise and Omniverse. We aim to be what the other company, the big phone company that's for B2Cs very well where they integrate hardware and software. We aim to do that for our enterprise customers. And this is key to driving the next generation of AI applications.

Just in a nutshell, we are helping customers. We're building this -- like the strength of our joint portfolios to help customers build their, as I said, enterprise-ready machine learning architecture. What does that mean? It means that machine learning processes are typically complex workflows as the 1 that you see at the top that involve a number of personas? And it's not just about 1 part. It's actually an entire stack that contributes with software-defined infrastructure, machine learning data pipeline and services and a wide variety of machine learning and software tools. And as you can see here, we are integrating at every level of the stack with different capabilities. And we'll tell you later about what these capabilities are. One thing I want to draw attention on is that at the heart of this entire joint stack is the combination of NVIDIA Enterprise and Red Hat OpenShift, which offers -- taps into the power of Red Hat OpenShift as a hybrid cloud platform with self-service capabilities, that allows running hardware-accelerated solution, whether they're training, fine-tuning inference using NVIDIA AI Enterprise tools, libraries and free trade models anywhere; on-premises, environmental and virtualized environments as well as in public cloud environments. You can take advantage of your hardware acceleration in all these places. So what's ultimately more -- most important here is that we provide a platform for converging all your high-performance compute, machine learning and artificial intelligence for cloud on a unified stack. Let's take a little bit of a look at what kind of key customer applications and workloads we enable on this joint platform. And we talked about OpenShift and GPUs. A lot of our customers, for example, use training directly on RHEL, for example, and traditional client finance use cases directly on bare metal using a combination of RHEL and GPUs. And they're looking at making the transition into containerized workflows running an OpenShift as well as running them side by side with other types of workloads, like machine learning, data science and MLOps. Also, a big portion of these use cases involve the deployment of different types of models, whether trained in-house or pretrained and fine-tuned from third parties and ISVs. And I think this is 1 part that's very, very exciting and this is 1 part where we're going to focus in our presentation later is because there's an entire wealth of such models that can be -- that our clients can take advantage of and they're part of our joint ecosystem.

And we also have to mention the fact that customers are looking at the big picture here. For real AI projects to happen, there has to be multiple use cases that are solved for customers. And this involves Quant Finance, machine learning, non-neural nets and deep learning. We'll go into neural nets and define what that is for you later in the presentation. NVIDIA, on top of that, adds the ability to provide the AI layer called NVIDIA AI Enterprise. We provide a number of pretrained models as well as app frameworks. And these include tools such as cuGraph, RAPIDS, where you can do your machine learning modeling; train, adapt and optimize, which is a tool kit for fine-tuning your applications or deep learning models; TensorRT that optimizes your deep learning models to run better on hardware; as well as NVIDIA TITAN. The key item here is to ensure that we are able to develop quickly for the multiple use cases at hand. For example, such as with NVIDIA neural networks model called NeMo, with which you can deploy conversational AI. So together, Red Hat and OpenShift have provided this layer where you're able to build your stack like a computer and you are able to start from the scratch, describe your problem at hand, take a prepicked tool and then further adopt it with your data on OpenShift platform and get to the use cases on top. What is your own experience?

My experience has been that, in general, customers want a platform that can address all these. Like the question is A, B, C or D? It's no, it's all of the above. They want a consistent platform where they can run the different parts of their machine learning workflow and they want to run the different use cases consistently. They want to focus on what brings them value, which is solving the real problems as opposed to putting together and stitching together the platform in each and every case. So I think that's this consistency, this convergence that we're talking about here is extremely, extremely important.

And we also see that OpenShift provides that containerized platform, right, where you can have like many containers deployed on your OpenShift pods, each of which serves its own purpose and you can use specific kits there and bring all these applications together to work seamlessly for the customer use case.

That is very correct. And in fact, actually, I have -- we have a good example of that happen, right? Like we've not only kind of talking about this is an aspirational thing. We're actually kind of -- we have customers having real-world stories of how they achieve great results doing that. So if you allow me to talk for a moment about one of our common success stories, right? And it's very well publicized. You can read all the announcements the Royal Bank of Canada, who's a very advanced -- AI advanced bank is using a joint Red Hat OpenShift on NVIDIA -- sorry, EGX systems platform to become more productive, right? And all these containerized machine learning applications and services that we're talking about, all this consistency of a private cloud platform enabled by the NVIDIA operator actually serves to accelerate everything that they do. They're not only doing things faster, days instead of months, they're not doing more, they're able to run 10x more experiments, but they're also being able to run a wide variety of models and use cases on the platform. Like we're talking about 1,000-plus models. So this is a great example of what we talked about earlier, how this consistent platform actually helps customers realize business results. In reality, there is a large number of use cases that benefit from acceleration. They're ranging from financial revenue-enhancing use cases like Quant Finance for example, the ability to address market and counterparty risk to do back testing, which are typically high-performance compute high-volume computation, not necessarily AI-enabled, they still benefit a lot from the acceleration part. Down to applying AI, helping faster decisioning with real-time analytics and analyzing alternative data, things like news and sentiment data [indiscernible]. Also, there are use cases around operations, doing know your customers and anti-money laundering and fraud and all this is better, down to intelligent automation. And of course, there are customer experience enhancing use cases, things like conversational AI and chatbots and personalization. Today, we're going to focus on how unstructured data helps a lot of these use cases. In fact, unstructured data is kind of present in a lot of the boxes that you see on the screen. And in particular, we're going to focus on intelligent automation.

One another area that we have seen is this is not constrained. You have to make the unstructured data work with your structured application. It has to work with your core banking application. It has to work with your ERP layer. It has to work with your database layer. And we can see that a customer is able to tap on to all this if they are able to bring the big picture together. And we live in this quite an interesting age where chatbot GPTs have become popular. We have the new concept of generative AI, all of which it is driven on neural nets, and we will define on what these terms are in the next slide and understand how our platform solves these use cases. So let me quickly touch on how Red Hat and NVIDIA together solve this big picture for the end client. On the right-hand side, you see it starts with our accelerated stack where we can call it as IAS. We provide you all the hardware tools. But more importantly, right, it's key for us to provide the software tool. And this is where, together we, Red Hat OpenShift and NVIDIA, where we are able to provide the platform as a service for helping you build these key applications. And what we have seen is that AI applications are stuck in a limbo. Executives have not been able to deliver these projects or deploy them. And that's where NVIDIA AI Enterprise comes in on the platform as a service, where we are able to develop advanced models together quickly with the use of NVIDIA pretrained models, where we bundle toolkits such as NeMo, RIVA discussed earlier. And the customer is able to quickly apply it to their AI projects, where they're able to build on their own as well as with strategic partners with pretrained models. So essentially, we are able to address the data scientist user all the way to the end business domain. And you can see that this range from quantitative analytics projects on the left-hand side, the business delivers around machine learning projects, data transformation and loading ETL projects where NVIDIA toolkits such as RAPIDS exists. And ultimately, right to the task, topic at hand where we are discussing neural nets, which is a new form where it behaves very similar to your brain. And it has -- it gives you the magical power to read through documents, interpret tables and natural language processing called AI RPA, AI intelligent dock automation, as well as deep learning-based natural language analysis for investing risk, ESG, environmental, social, governance where companies are able to tap into alternative data and generate alpha to drive more front-line revenues that we discussed earlier. We are also increasingly seeing areas such as graft neural nets for AML and fraud. And Marius, what do you see, right, along these unstructured neural nets? What makes you excited about other applications here?

What makes me -- honestly, what makes me excited is basically the new potential that it opens, right? Like they're very -- I think financial institutions have been doing, not only financial institutions, enterprises in general, are trying to tap into their data. They're trying to do more with the data that they have. Hence, they've been very good and very focused on using structured data solutions, practically kind of extracting data from their core banking systems, analyzing transactions, analyzing trades, like anything that comes from their repositories where this kind of data comes in a tabular format and applying all kinds of statistical methods to analyze this and build solutions. And they've been doing that for a long, long time. But what's exciting is that neural networks and opened the door for processing a different type of data, which is just as valuable.

We spoke about unstructured data in the last couple of slides. Let's quickly understand why this has been key and very important. Organizations are currently sitting on valuable data on their premise, be it news and media, earnings and call, imagery satellite, that's very key for climate risk; customer onboarding farms for products where the valuable customer information that can be useful for fraud and analytics; payments and invoice applications; 10-K, 10-Q financial statements, analysis, documents and mails. We could endlessly talk about a lot of these things. And all this current information is left untapped. Where current customers are focused is structured data, which is sitting in a siloed manner. And when you combine this with unstructured data, you could make sure your core banking systems, ERP systems, supply chain systems work seamlessly where you can extract data from unstructured data, downstream to be used for your structured models, which will get you the ROA. And you can also make sure that you use such data for both internal and external customer experiences. But there are some challenges.

If that's asking like, what do these organizations do? Why do they're focus on structured data, right? It's basically because it's easier. It's easy to extract the data from -- it's already prepared, it's already kind of built together. You probably heard about like trillions of the new data being created. But as Prabhu mentioned, a lot of these data, a lot of the data that, for example, lines of business have, comes in the form of documents that they possess, right? And the challenge is that the traditional machine learning methods that have been developed for -- to work on structured data, they don't necessarily work very well for unstructured data. You need to do a lot of complex transformations, which are manual and never prone to basically transform the way we look at an image or we look at the document, using something that works with the statistical methods, which don't mean a traditional machine learning. Here it helps like the kind of new methodologies like deep learning, which essentially mimic the structure of the human brain to kind of try to get a holistic understanding of a complex object like an image or a document work better on unstructured data. And we have emergent methodologies like speech detects or natural language processing or computer vision, right? Try to think about how a kid learns, learns by example, learns by processing and kind of tries to understand the world in a more holistic manner, right? The challenge is that these deep learning models require large quantities of data, right? If I understand, for example, the structure of the language, you need to learn a lot of stuff written in that language, right? And luckily, for example, there are pretrained models in the public domain were developed by organizations that do this kind of work, like they feed a lot of data into these models, preparing them to be able to understand the structure of a sentence or understand how different objects look like. The challenge with those models, of course, is that they're more general. They lack specificity, either for understanding how like a business domain or for understanding the specifics of an organization. So this is where techniques like fine-tuning, for example, help us take the great work that's been done to train these objects and these models and applying in the context of a specific organization.

And here, I would add like similar to the example that Marius discussed, we have a kid who is trained in the language. He knows English, but he doesn't know how to apply that English in the specific task at hand for writing in a history paper. And this is where the next steps come in place. We have 3 steps here, the pretrained model followed by fine-tuning, otherwise called as transfer learning; and essentially deploying those 2 trained models into inferencing. This is the premise and mantra that we are laying out here for our customers to follow. If you're a large customer, you could start training that base model. You have a large number of data scientists, and you want to train it to your specific task or begin with your data. But most of the customers belong to that world where you don't want to rediscover the fire. You don't want to reinvent the wheel and build it. You want to start with something that is already built, available for you to leverage. You just have to train it on that particular data, which is for your use case at hand, how we treat this classroom with multiple kids and they have to be trained. And you can consider that each of them, kids have to be taught for a specific skill set. Somebody is good at music, somebody is good at reading, somebody processes natural language better and somebody passes vision models better. And Marius, I mean, how have customers done on this project based on your own experiences?

Yes. I mean, they have a number -- they need to do a number of things, right? They need to gather the data. They need to put it together. You need to set up a process for the model to be trained. They need to run things like experiments, for example, to figure out what is the best way to train -- to fine-tune a specific model. And then what they need to do is to go and put these models and run them in production. And it doesn't stop there, right? Like the -- I think one of the biggest challenges is being able to continuously adapt and to continuously learn from as more documents and more data come through and through. So applying this process of continuously updating the models is actually a key piece here.

Let's quickly look into how our stack helps us do those tasks. For example, fine-tuning. We are -- NVIDIA provides the toolkits suchas NeMo, neural nets model, train, adapt, optimize style RIVA for your conversational AI applications. With all this data, you can take your base model, which can be a transformer model, such as Bird or any other model that NVIDIA provides, and you could train it on your 10-K, 10-Q data, SEC filings. And Marius spoke about inferencing, right? The essential ice on the cake is, we are able to take these trained models and now boom, you have deployed it on OpenShift, where you can scale to multiple customers, the workload can be batched real-time and it can go to multi-node, multi-GPU systems. Power of OpenShift can scale it to cater to a large number of customers. Let's go into details of the training, fine-tuning and inferencing.

Yes. Yes, I think there is something to be said here about like the training process and the inferencing and kind of the technical details, how this process actually comes together, right? Hence, you will see a working example in a few minutes like a solution that actually built by 1 of the NVIDIA partners, that helps -- that illustrate how this process works in practice. But let's -- I just want to kind of spend a couple of minutes discussing the advantages and benefits of training and doing the training process, in particular, for fine-tuning on a platform like OpenShift. And the reality is that when you kind of -- when it comes down to training for unstructured data, you're still training neural networks, right? You still need accelerated hardware such as GPUs to be able to do your work. You still need to have the efficiencies of doing a multi-node, multi-GPU train, right? And this is where OpenShift is a great platform for not only kind of providing you seamless access to GPUs, but you also allow you the working -- like this process to work at scale. It's been kind of -- in order to get the efficiencies of running, for example, something like OpenShift, sometimes you need -- you can benefit from things like a secondary scheduler that's GPU optimized. So this is where we work with partners, for example, like Red Hat AI, right, which is a joint mutual partner to enable this type of efficiency. And I would say making kind of a pun, fine-tune OpenShift to be able to do training and fine-tuning for these models.

We also discussed this -- I mean, like the kid is sitting in the classroom. Obviously, they're learning all the skill sets from everywhere. But the kid has to be corrected and fine-tuned. These models are working in their real-life correction. And they have drift. These models are not performing well. So similar to a kid in the class, they have to be corrected on the spot. And this is where like OpenShift shines. You are able to do that fine-tuning and retraining to make sure that there is an active learning pipeline. You are making sure that, that happens, this training is happening when they are in school. And this is one of the key advantages of OpenShift. You have like a production-level grade deployment environment, where this architecture is suitable for training, fine-tuning as well as inferencing.

See, the end goal of the entire process is not just to train models, it's actually putting them in production, you want to be able to run these models. You want to be able to run these models. You want to run them on a wide variety of use cases. And as Prabhu mentioned earlier, you want to be able to scale out to meet, for example, the needs of end users, if you're doing customer-facing applications or to scale out to tens of thousands of documents. So this horizontal scaling, it's one of the great capabilities that Kubernetes and OpenShift provides. We also want to be able to run heterogeneous workloads, right? These models are not always running their refines of GPUs, although some of the more complex neural networks require that. They're also running on CPUs, for example. Trident, for example, is a great tool for that. And then they need to run side by side with the other application services that enable this functionality, right, like databases or streaming applications and so on. So all in all, OpenShift provides a great platform to run all these inferencing jobs, together with the business applications, together with the services and provide a complete solution to your problem.

Right. As we discussed, we are looking at converge HPC plus machine learning plus AI workloads. So here is where together, NVIDIA plus OpenShift are helping you in Quant Finance application. Deep learning applications that we are focusing here with inferencing. And these need real-time responses. The documents that we are talking about here need to be answered in milliseconds. And the models that are being used here are large transformer models, big deep learning vision models, which have millions and billions of parameters. And these need instantaneous response, or it also needs a batch kind of workload where multiple documents are processed quickly as well as it can be scaled to thousands and 10,000 customers. And cluster can get to that maximum multi-node, multi-GPU media accelerated stack of OpenShift.

That's a great point, Prabhu. And I would add here something that what we talk -- we start these 2 platforms side-by-side, right? We talk about training and inferencing kind of as being 2 different facets of the same problem. But one of the powers of this unified platform is that at the end of the day, you can run all these jobs together. You can run your training and inferencing and everything you need to do on the same platform and essentially accelerating this process of fine-tuning and deriving value from data that you have and as part of this workflow. And in what's going to come next, I'm going to hand it over to Prabhu to walk us through this real-world example.

Quantiphi is a key NVIDIA elite partner having done the extensive engagement for insurance, banking and capital markets. The team will showcase an AI document digitization example, continues to be learning machine models. This example illustrates base model training, fine-tuning transfer learning with customer data sets and inferencing that we discussed earlier. To you, [indiscernible] from Quantiphi team. Thank you.

Thank you, Prabhu, on briefing about Quantiphi. Quantiphi started in 2013 at the same time when AI is moving from academic discipline to more of tactical applied science. We have over 3,800 professional and 100 within NVIDIA practice, and we have more than 250 professionals working within R&D development. Quantiphi have very strategic partnership with NVIDIA, where we are portioned as an elite service delivery partner. We work closely with customers to build custom AI solution using NVIDIA AI stack for on-prem, cloud or even hybrid deployments. Our key area of expertise are document understanding, NLP with knowledge, grafts, custom AI analytics use cases. Here, we have a dock AI accelerator solution package, which is built using NVIDIA Enterprise with a set of software development kits under NVIDIA Enterprise on top of the data OpenShift. Customers can use this solution accelerator package, and they can easily build the custom AI projects or document understanding. Document digitization and document understanding is a common challenge across all industries, whether it be insurance, banking and capital markets. In banking and capital markets, we have use cases such as LIBOR transition, where exposure to such interest rate products are examined from contracts. Customer e-KYC, Know Your Customer, are used for various banking and products for customers that are usually available in digitized manner. For example, mortgage agreements, mortgage verification and other contracts. Other examples are payment and invoices data from suppliers and customers that have to be processed. Document digitization is usually customer-specific AI project where it means to be customized according to each of the customer needs. In financial space, usually the documents are complex. Let's see an example how a document format would look like. As you could see here, this is a sample invoice form. Here, you could see there are different layouts to it, a quoted text. We have hand it in text. We have different layout attached to this documents. To extract this information out of this document, we should have modules like [indiscernible], optical character recognizer, layout processor to exact layouts likes tables, et cetera. All this model should work in tandem to extract information out of this document and store it and use it for further downstream applications. With that, now let's look to look into our solution. Here, you could see our document digitation solution is completely deployed and Red Hat OpenShift with [indiscernible] Let's run it through the solution. Here, we'll upload a sample form. And we'll have a base model trained on a base created data set, which will help us to extract information out of this form. Here we're using the base model, which is trying on a curated dataset. Our model is not trained on a customer-specific curated dataset as we don't have access for it. So the accuracy will be a little low. Our idea is to keep the pretrained model trained on a custom curated dataset so that it can be used by customers to retrain it on their own dataset so that it can be up to the mark and bring in to the confidence and accuracy. Here, let's see this example. I'll submit this invoice and let's see how this information out of these documents have been extracted. As you could see here, the setup models we just discussed earlier are working in tandem and extracting this information out of this document. Here you could see initially, these checkboxes have been extracted. [indiscernible] operation. All these have been extracted. The key value has been extracted. The key value is been extracted, application name, company name, e-mail, phone number, produce name, underwriter name, et cetera. And third, we are extracting the table. All the data present in this table have also been extracted. Once these data are extracted, we'll store this data in the database and we'll reuse this data for a downstream application. As we discussed earlier, all these models are deployed in the Red Hat OpenShift platform with the help of [indiscernible]. As this is built using in NVIDIA Enterprise and certified by Red Hat OpenShift, customer can deploy with confidence without worrying about much of compatibility issues. As you could see here, all our models are deployed in the Red Hat OpenShift. We have multiple components, which is running in tandem to make sure our pipeline is working. Here we have a document digitization back end; document digitization front end; document digitization freight and inference server, where all our models have been deployed. Freight and inference server is a part of NVIDIA Enterprise stack, which helps us to deploy our models in production. We also have a retraining notebook running where customer could go and retrain the models on their own curated datasets. As we could see, our solution is cloud-native. Customer can without worrying about much of deployment, they can deploy across infrastructure, be it cloud, on-prem or even hybrid. With integration of NVIDIA Enterprise, Red Hat OpenShift and NVIDIA GPUs, customer can focus on solutioning instead of focus on setting up the infrastructure conquering the files, setting up the SDKs and libraries. NVIDIA AI Enterprise plus Red Hat OpenShift also supports multiple MLOps tools. Customers can easily set up their MLOps pipeline to make sure the ML project like maintained well, right, from the data collection, model retraining and deploying the model in production and monitoring the models in real-time. Here, our solution is not just about training the models and deploying it in production. It is more about how we integrate with the existing business application. With NVIDIA Enterprise and Red Hat OpenShift, we can seamlessly integrate this pipeline into the existing business application to generate more value. Customer want to try out this solution now, NVIDIA provides a best platform using LaunchPad. You can go and register and try out this solution. The reference links have been added to the source section. You can reuse those reference links and register for this. With that, I hand it back to Red Hat and NVIDIA team. Thank you.

Thanks, Quantiphi team. That was a great example. Let's try to recap quick what we talked about today. First, we're very, very excited about this, the potential that unstructured data provides. It's founded upon emerging technologies and advancing in deep learning, in things like natural language processing with Bird's and so on. But what's very, very important is that we're showing a way to close the gap between taking advantage of these innovations and move it into real-world use cases. You've seen how NVIDIA Enterprise provided a joint platform for solving these type of problems and how it helps you accelerate, moving from applying these models to actually creating real-world use cases.

So this is where completely, NVIDIA is on the same page as OpenShift, and we are doubling down on these use cases. As we discussed earlier, we need to look at the big picture. There is a way where clients can transform their front-line revenues. For example, we discussed alternative data, and NLP for financial services, where these teams are looked at for risk investing or ESG, environmental, social and governance, that's been a key area. In addition to dock automation, there are also other areas such as chat GPD, digital avatars where customers want to appeal to the millennial customer, increase the number of retention, increasing number of users so that traditional banks and capital markets can compete against fintechs. And the way to do this would be we are doing such structured AI products, which cover this big picture around unstructured data along with structured data, greenfield and brownfield projects driving huge revenues to customers.

I fully agree, Prabhu. I think what's key here is the wide variety of use cases that we can open in different areas, in different business areas, like we're talking about document processing but improving customer experience, improving call center platforms and many, many other business use cases is what this platform opens a door to.

So now that we have discussed all the 3 key aspects: retraining, fine-tuning and inferencing, along with a great number of use cases, let's quickly take a step back on how NVIDIA AI Enterprise plus OpenShift help us here. We now start with a point where we have a tool for training to a specific task at hand. So this accelerates your AI development project. You are able to start with a MLOps consistent model that you are able to deploy for a specific AI use case at hand. And the customer, you are able to bring your own data and fine-tune this model to greater precision and apply this task of fine-tuning and transfer learning to your specific task at hand. And we quickly discuss why this is needed in AI. AI has been stuck because customers have not been able to reach this point previously. But we have given you the mantra, the secret success and the recipe where we start with our NVIDIA toolkits and pretrained models, and you fine-tune it to the task at hand with the help of our SDKs further. And we can quickly see how your barrier to entry is lowered further where our platforms together make this happen.

It's also a very important point here, which is this is also a way to lower the entry barrier for projects. Like we see it as critically important. Like if you were -- if you're trying to build a business solution, right, you're trying to get -- as you said, you're trying to get these results as quickly as possible and you're trying to maximize the efficiency of what you're doing. And this is what we're offering here. We're offering a recipe to -- for institutions to kind of realize these solutions very, very quickly. Also, an often neglected aspect of this is security. And like the entire -- especially in financial industry, which is a very regulated industry, this asset is critically important. And using libraries that are certified, that are validated and then using the security controls built into the platform, the lines of business and financial institutions can deploy these projects not only quickly, not only with precision, but also with the confidence that they will treat the sensitive data that they're handling in a very secure manner.

OpenShift has addressed these areas where a lot of these business applications have been already deployed for brownfield engagements. We are doing this, we are also doing these greenfield projects under those areas. And together, we are making it a success. We are going to discuss how you can access these labs on our platform, NVIDIA LaunchPad on OpenShift. We have a number of labs that range from conversational AI, where you are able to experiment with NeMo, which is our SDK for neural net modules for conversational AI. Others such as RIVA, Maxim for augmented reality. And all these tool kits are available for you, the customer, to go and try it out. And these are available on certified systems, on LaunchPad, on OpenShift. And there's a no-cost trial where you can play around with these projects to check whether you could reach your end goal of customization for our specific task at hand.

Thanks a lot, Prabhu. And I strongly encourage everyone to go and try the LaunchPad Labs with Red Hat OpenShift. Also, we share a number of resources, where you can learn more about NVIDIA, about Red Hat, like Red Hat's approach in data science and MLOps and, of course, solutions for financial services. So want me to stay in touch with us and learn more. And with that, I'm going to hand it over back to Aric for the Q&A. Thank you very much.

Thank you, everybody. Thank you, Marius. Thank you, Prabhu. We're now in the live Q&A portion of the presentation. [Operator Instructions] So first question I have here, the HPC AI/ML stack appears to address scalable platforms with job schedulers. Is this a platform that could be deployed and useful on the Edge, for example, for credit card fraud detection? Do we need to use Enterprise, traditional HPC scheduler such as [indiscernible]? Marius, can you address that?

Thanks, Aric. I want to thank everyone for attending. So that's a great question because it comes down to applying AI and applying models in practice. And you're absolutely right, the traditional stack for training models has been using schedulers and being using things like [indiscernible] for example, especially for distributed multi-node, multi-GPU training scenarios. Hence, I think when you look at -- kind of to address the first question about Edge, there is definitely a space there for having things running side by side for example, with models being trained on traditional HPC platform and then deployed for inference at the Edge where kind of the PowerEdge solutions, for example, if you look at what Red Hat has been doing in the Edge space, definitely kind of tries to address that. At the same time, when you think about AI, there's definitely been a trend of moving some of the workloads from -- necessarily from traditional scheduling scenarios where, by the way, the stack of like [indiscernible] LSF, anything running on something like RHEL, for example, running on NVIDIA GPU works very, very well. And we've been very, very successful at our customers with doing that type of thing. The trend is to move and start exploring more towards the containerized Kubernetes, OpenShift for that type of work, especially for training. And the part of the reason is because, and it's been brought up in other questions as well, the training workloads become more and more complex. It's not just about running tasks. It's not just about running very compute-intensive job, but also there are questions like how do we recover? There are questions like, especially with deep learning, there are hyperparameter tuning scenarios when you want to run multiple experiments and get the results of that. So I don't want to elaborate like a lot here, but in some ways, the complexity of the workloads themselves lends itself very, very well to kind of a more complex and distributed platform. And there is an inherent advantage in kind of using a unified platform for doing all things from training to the -- from data exploration to training to inference and observability. So -- but again, always it's about kind of choosing the right tool for the task.

And Marius, we also addressed this, right? It's like a small drop is to make an ocean. So there are different customization needs, fine-tuning transfer learning needs, right, on the Edge, or on on-prem, on OpenShift, wherever on OpenShift and that triggers this need for training where OpenShift trends.

Thank you. Thank you, Prabhu. An important aspect to AI/ML is the concern about bias. And so we have a question here, how are you dealing with bias and models? It seems to be an active conversation in AI research, and it's not a solved problem. Generally the rates seem to be an issue with both NLP and cross-model foundational models. Prabhu, can you address that question? Prabhu, you're still with us? Okay, Marius, can you address it?

I'll get started until Prabhu reconnects. So definitely comes down to the complexity of the processes. There are tools that are explainability tools. There is -- there are a means of testing the models specifically for bias, there are means for testing the data against bias in data. We get them as good as a model as the data that you're using. So of course, there are the human processes...

Let me jump in for a second. But traditionally, if there was bias and decisions being done or bias in that original data, are you not just compounding the issue and automating and make it faster?

That's why you need to test the results. That's why you have to -- like you have tooling, for example, for -- and we work with a number of partners that provide that type of tooling. You need to incorporate that testing for -- against bias as part of your pipeline, as part of your MLOps on MLOps pipeline. And again, it comes out to the complexity of your process. You need to incorporate them and make them a part of your model.

Please go ahead and repeat it. I mean, that was the part where I had kicked off a question button then I got logged off and joining in because I'm a human.

Good to hear you're not a deep NVIDIA fake as far as your RIVA. So you're here, you're live and you're a real person. So the question is about bias in models. Active conversations about how do we not have bias on our models. And when we have bias in our original data, in our training data, how do we make sure that, that bias is not automated through the AI/ML models?

This is actually a very, very applicable question, right? We have been discussing about GBD chatbots coming out with wrong answers. A lot of what is in AI is to deal with the data, right? You're trading with a large amount of data, various alluded to that, right? And the AI models picks up a lot of these biases that are in the data, right? So I would make sure that the data is cleaned, right? And this is an intensive effort where we need to pull in all items. So that needs to be revisited. There is also like a field of AI called [ ex-AI ], where you ensure that the explanations for these models exist. You kind of understand the rationale. And that would be what we would do here to ensure that customers can adopt such AI areas, make sure that your data rate is not already bias. And the follow-up would be a strain on such data and then use such tools. And one another area is that [indiscernible] by now, AI itself has proven in cats and dogs, right? It's very great grade for image classification tasks like where it can make out what's a cat, what's a dog. So maybe start using AI in such areas where the compliance area is not that high, right? And you could start seeing the benefits, for example, dock AI automation is an example use case, right, where these kind of items do not matter as much as it should. And we could start with those engagements and slowly move on to these harder problems, right, and apply these tools to get to the AI use case for business analytics.

That's a good segue, too. I mean, are there pretrained models that people can be leveraging that are maybe more observed that have less risk of, say, bias? And where can people start working with these models, these pretrained models to: A, accelerate themselves in terms of learnings; and also accelerate time to market?

Yes, there are a number of pretrained available models that are available both with NVIDIA as well as the public. And again, you want to [indiscernible] you want to use the right pretrained model, right? And we discussed like a pretrained model that was available for dock AI. There are also -- if you go to NVIDIA LaunchPad, we also make other application frameworks such as NeMo, neural network module available or fine-tuning to our conversational AI application, for the natural language application. So -- and what I would also like to tell is, right, there are also these ISDs, other partners, right, who are offering these pretrained models. They have put in a lot of work and they have ensured that this can be applied specifically to the use case at hand. So it's just about identifying where your organization lies, what's the effort that you want to put, and decide the problem and come and use these pretrained models. We'll mail out these resources so that you can try out and you can also log on to LaunchPad. But ultimately, it would be great if you define what your -- if you know what your problem area is, start with that pretrained model and make sure that the pretrained model, for example, NeMo or RIVA has these various fine-tuning opine parameter options to customize real training needs, and that will work. And that's where sometimes you want to apply AI that's proven, right, for example, even chat GPD bot, right? We could make sure that it trains on your own data. For example, if it's a list of customer facts, we could make it a closed model, starting with such pretrained models. So it just answers based on your fact and it will respond, it doesn't no. If it doesn't no rate if you cannot find it in the chart. So that's why it's key that you use the right pretrained model which is better.

Yes. I would like to add something here, which is, I think, interesting and it kind of ties into even a hard G&A in open source. Like, one thing that excites me about pretrained models is the extent to which they democratize access to knowledge and to capabilities in many, many ways. Of course, there is a lot of -- like there's a lot to do. You can pick up a bad model like from the public domain, and you have to protect yourself against transparency and like against bias and other things like that. So you can't just go and pick up anything. But at the same time, this ecosystem of models kind of allows you access to a lot of knowledge that's built into these pretrained models that you can tap into. So I think that's one of the biggest changes from kind of the traditional machine learning that we've seen like in the past. Like it's not about just organizations sitting on top of their data and kind of treating that as a secret, but collaborating openly to get -- to build better models. Let's start addressing different types of problems. And one example, I think that's kind of been very, very exciting in the past is the space of, for example, climate change. Starting like collaborating on models that kind of model, analyze the kind of the environmental impact that companies have and for the analyzed carbon emissions. I think that's a very, very exciting space where that demonstrates like the power of collaborating in the public domain around like these types of pretrained models.

Great. Thank you. I think we have time for one more question. I'm going to direct this over to Marius. So the question that we got is, how is deploying an OpenShift lowering the entry barrier? Is it not more complex than running on bare metal?

In terms of what you're saying, like where you think the complexity lies. If you think about like a bunch of processes running on an operating system, for example, that's a very, very simple thing to do. What we do see, though, is that the training and the inference process has become in and of themselves more and more complex. We hinted earlier on a wide variety of tools, for example, for running experiments, for running workflows that not just take a bunch of data, run for a bundle of time and take out a model, but actually can analyze what's happening with the model. You have, for example, in deep learning things like early completion, we decided the model is good enough and you want to stop right then. You wanted to have more complex workflows with -- again, coming back to the question of bias and transparency, you want to incorporate tools that analyze your data before it's fed into the system and you want to have tools that analyze your models in real time, look at the types of responses that they give and try to like, use explainability tools to actually assess how healthy those responses are. That's something that needs to happen in real time. And that's where you need to operationalize the entire process. So when you kind of look at the entire ecosystem of things, this is a large distributed system that runs continuously. And here, it looks like OpenShift, for example, OpenShift has been a great platform for running exactly these types of systems. We kind of demonstrated it with DevOps and we're really talking about applying the same principles from DevOps into machine learning.

I'll also add here, it is, OpenShift on Kubernetes is a trend, right? And AI is a trend. And these needs multiple components like you mentioned, right? So you see 1 and 1, right? And it's very easy to put 1 and 1 plus 2 together to understand that OpenShift on Kubernetes is enabling a lot of these AI items. And because of this containerized workloads, it makes it easier to get other tools in, right? And it can become like this universal platform where you could bring in the tool as you want on the cluster on OpenShift, and that could address your ex-AI needs, machine model training needs, as well as it could address to your earlier question, Aric, like I mean why do you want to go from bare metal, right? You could do this fine-tuning transfer learning for different use cases at hand and deploy it in production, right, and do active learning. And so it's about us for to put that piece together, right, that this trend, many people think that it's a separate trend, but it's the same trend together happening together.

I would like to add some. Yes. One more thing, which I think is very, very important. All you're describing right now, what we see with our customers, for example, is a desire to move off-premises into the cloud and start exploring the cloud as a platform for doing that type of work. So on the one hand, concurring that complexity is a benefit, but also having the portability that containerization provides that are we simply pulled a plug for ourselves. But the portability that OpenShift provides as a platform for running in the hybrid cloud and different types of clouds and the ability to kind of run hardware like -- software like NVIDIA Enterprise, they kind of sold these types of problems. On any kind of environment, I think that makes a very, very strong case as to why you want to think in that -- in this direction of moving to OpenShift and doing the work that way.

Right. I mean anything comes back to the other person's question with regard to country running at the Edge as well. You have a common experience, whether it's on the developer workstation, whether it's for training, whether it's in production, in the cloud, on-prem or on the Edge. So I want to thank you both. We've had a number of great questions. Unfortunately, we couldn't get to all of them. But I want to thank Marius Bogoevici from Red Hat, I want to thank Prabhu Ramamoorthy from NVIDIA for an engaging conversation, and also the partner is over quantified for a great ML. So thank you, everybody.

Thank you, Aric. A pleasure to always catch up with the team. And hopefully, Aric, we shared all that we have and keen to help more customers out there. Thank you.

And I wanted to add my own thanks. And just a quick kind of, just a quick hint. Thanks for asking those great questions, and we'll try to answer as many of them as we can offline, and maybe post something like a blog too, that we can go and read the answers to that. Thank you very much.