 Good morning, small cozy room and a small cozy audience. So thank you for showing up for my talk. I know you have multiple talks going on. So thank you for coming to my talk. My name is Arun Gupta. I'm the Vice President and General Manager for Open Ecosystem Team at Intel. And as part of that, my team is responsible for crafting Intel's open source strategy. We work with multiple software teams, hardware teams, API teams, specs all across the board to craft that strategy. And today we'll talk about how Intel is working in open source to bring AI to everywhere. Whatever device you are using, whatever screen you are using, essentially. So let's take a look at what has happened recently. Nothing new. This is nothing dramatic. But really, if you think about it, AI is exponentially increasing in both quality and size. If you think about the first transformer model that was done way back in 2017, about 65 million parameters. But if you see the transition, how it has gone from GPT-2 to 3 to 4, what they're looking at, we're looking at about a trillion dollar per trillion parameter. This was in 2022. And this is only growing and only going to grow more. So the quality is increasing and the size is increasing. So in five years only, this is about 15,000 times growth. Where this is going to stop? And what is the point of diminishing returns? How do you get that compute? What is it that you have the value that you're getting out of it? That's one element. So that growth has happened. But it is still inaccessible to a lot of developers across the world, whether it's on the training side or on the inferencing side. So if you look at it, for example, on the top, it shows the training cost for GPT-3 is $1.65 million. Based upon how you are calculating it, if trained on, say, Google TPU-V3. Similarly, for GPT-4, it's a $40 million training cost. Not all companies, and this is just a part of it, not all companies have that level of infrastructure, compute processing power, or the dollar amount to get that training going, essentially. And similarly, if you look on the inferencing cost for chat GPT or Bing AI chatbot, if you're thinking about it, the back end is constantly churning out data. The chat GPT is a lost business at this point of time, because the amount of dollars it takes to run a single chat GPT query is pretty significant. And that's where the chat GPT plus enterprise model, all of those are coming along, just to make sure they can actually make it a bit more sustainable. If you think about, again, in terms of the large language models, what's happening in that space? We started with, say, Dali, Bard, GPT-3, 4. The number of parameters are increasing. And it's really driving the productivity to new heights. The innovation is focused more on scaling the AI model size. Is that the right approach? And then more recently, Gemini was introduced by Google. There is Purple Lama, which we'll talk about in a second, was introduced by Meta. Is openness, is closeness a factor in there? Now, how do you make sure that we can build more transparency, build more trust in these models? That's a critical element that Intel really deeply cares about it. These are some of the use cases on how consumers are leveraging AI to their benefit. There is, of course, GitHub co-pilot that you use in your day-to-day coding assistance, which is a commodity. But people are building custom co-pilots where they're able to dig into their network, where they're like something like a retrieval augmented generation pipeline where they're able to dig into their network and their database and make it more custom to their element. Think about the usage of AI for creating hyper realistic textures. My son is doing master's undergrad and then master's in gaming technologies. So he's deep into gaming and creating those textures and biomes and all. It's still manually driven for his project, but that land is moving towards AI-generated textures as well. I listened to a lot of podcasts. One of the podcasts that I listened to is Hard Fork. It's a podcast by New York Times. And they were talking about how Mountain Dew is using an AI chatbot to go through all the Twitch streams and identify any Mountain Dew branding over there, whether you are drinking, whether you have a sticker. Anywhere Mountain Dew branding is available over there. They get your Twitch ID. They send you a direct message and then say, hey, you want to be a Mountain Dew ambassador. So that's step one. Step two is where you become an ambassador, let's say, they accept the application, where you become the ambassador. Then they continue to monitor your Twitch stream just to make sure that you have not gone away or walked away from that brand ambassador ship, where you're not promoting the brand anywhere. And then they actually send you a message that, wait, you did not display the Mountain Dew for the last three days. Do you still want to be an ambassador? So something as simple as live streaming and what product are you displaying is going to get impacted anywhere and everywhere. And these are not really large models. These are very custom models, small models, serving a very particular niche, very use case. Now with that background, let me talk about Intel's legacy in open source. I'm just laying out the landscape here first. But what does Intel do in open source? Intel has a lot of legacy in open source. We have been contributing to open source for over 20 years. Intel is a Silicon company. It's primarily a hardware company. But there are over 19,000 software engineers at Intel, several thousand of them actually contribute to open source. And the reason we contribute to open source projects is because our customers consume our Silicon, whether in a data center, whether in a cloud hyperscaler, whether in a network or an Edge device or a laptop, off the shelf laptop that you buy from Best Buy or whatever your favorite retailer is. When you buy that laptop, when you spin up a VM in, say AWS, you download a latest upstream software and you expect it to work out of the box in the most optimal manner. So as we are creating new chips, we are making sure there are latest hardware instructions in there. And then correspondingly, we are contributing to those open source projects to make sure those projects are optimized for that underlying hardware as well. So as a result, we contribute to 300 plus open source projects. These are community managed projects. And by that, what I mean is projects like Kubernetes, OpenJDK, PyTorch, TensorFlow, scikit-learn, modern, LLVM, GCC, Linux kernel. These are the projects where we are one of the largest contributors. And as a matter of fact, Intel is the largest corporate contributor to Linux kernel for over 15 years for that reason. Because as you look into hyperscaler, still the commodity is Intel platform. And as you're looking at Intel platform, we wanna make sure all flavors of the underlying distributions, whether it's a canonical, whether it's Amazon Linux, they work. And the best way to make that work is contributing to upstream so that it goes into downstream. So that's sort of our philosophy. We are one of the top 10 contributors to Kubernetes and OpenJDK. And we are also the top three contributor to PyTorch and TensorFlow. And the reason we contribute is because our customers tell us that, hey, we wanna make sure that we are able to run this in the most optimal manner. More recently, we joined the PyTorch Foundation at the governing board level and the LF AI and data foundation at the governing board level as well. Because we believe it's not just doing work in the open source project, but really making sure that the governance, the financial responsibility, the administrative responsibility is exactly is about the sustainability of the entire ecosystem. So we continue to partner over there. Let me tell you a story about PyTorch. PyTorch, as we all know, the project was created by Meta. Now, Meta is the primary maintainer of the project and more and more maintainers are getting onboarded. Outside Meta, Intel is the only CPU maintainer of PyTorch. What that means is, Intel not only contributes patches to PyTorch that makes Intel platform shine for PyTorch, but when others are contributing, whether it's AMD or ARM or other CPU vendors are contributing their patches, Meta and Intel are the only maintainers that are gonna merge those patches. And for us, it's really about lifting the entire industry. We don't wanna be selfish out there just doing our work. So that chopwood carry water is a very, very important and essential element for us to make sure we create that impact. Another example on PyTorch is we let significant designs in PyTorch, something like torch.compile. You write your code and you say, go run it. And then we let the design so that the backend is pluggable because how torch.compile is gonna operate in the backend is gonna be different for Intel versus AMD versus any other CPU vendor. So we let that design and created space for other CPU maintainers to be able to contribute over there. For us, that is part of the legacy. That is critical for us that as we are contributing to these open source projects, we continue to create space for others and enable them to be successful because that makes the total addressable market a lot bigger where more players can compete on equitable play space as opposed to locking you into a wall garden where only one vendor contributes. And they define what the innovation is gonna look like. Talking a little bit about our open source legacy, as I talked about, Intel is part of several foundations and governing board. Take a look at top left, Kathy Zhang. She is two-time elected member to the CNCF Technical Oversight Committee which drives the technical direction for cloud native computing foundation. On the top right is Krob. He is part of the Technical Advisory Council of Open Source Security Foundation and the chair for that body as well. Down at the bottom is Marlo Western. She's part of the Environmental Sustainability Group. She's a co-chair for the group and defines how CNCF projects could be more sustainable. Myself, I am on the CNCF governing board representing Intel and I'm also on the open SSF governing board and I happen to be the chair for both the governing boards as well. And that is something Intel supports very well. Like when I was running for the open SSF governing board chair, my management supported it. They said, yep, you should run for it. But then I needed somebody to do my internal job and they gave me additional headcount where they could do my internal work so that I could be that external voice and drive that entire initiative for the industry from Intel. Let's switch gears. Let's talk about what do we mean by bringing AI everywhere? I think it's important that we help the developers realize the full benefit of AI. And the only way they can do that is where we bring AI everywhere to everyone, to all of your devices and to all of your use cases. That's sort of the whole concept of AI everywhere. No matter whether you're running it on a client, on a desktop, on a cloud, on our edge, you know, no matter what your use case is, you are training, fine-tuning, running a rack pipeline, whatever that use case is, we wanna support that. And the way we support that is, and the important element is we wanna make sure we do that in a very secure, transparent, ethical, responsible manner. Those are sort of the basic pillars on how Intel has operated forever and think about it. And so the three key elements that we wanna think about is like, how do we help you accelerate the workload with you, our foundational silicon and the software? That's at the bottom layer. Now, there are general-purpose CPU, but then there are targeted CPU, GPU, and NPU that I'll talk about in a second from the hardware perspective. On top of that comes is the AI infrastructure, which is where we talk about the projects where we contribute to and how we are enabling those projects for you from whether it's on the cloud or on the client or in between in the edge, wherever you're running it. And then the last one that I'm gonna talk about is streamlining the AI workflow to enable seamless deployment and deployment, development and deployment productivity. Those are critical elements. So let's dig into it a little bit more. So what is Intel's approach? Our work is really to, our approach is really to work at and every level of the solution, as you can see here, and enable the best of breed for both our customers and developers. So if you look at the bottom layer here on the accelerate the workload layer here. So we really have a very wide range of hardware. CPU, GPU, accelerators, NPU, neural processing unit that is getting announced, that was announced more recently. So you pick that and wherever you wanna pick it. If you look at your CSPs, those CPUs are prominent in all regions in all availability zones. You look at Intel developer cloud that I'll talk about in a second, where we are providing you the pre-production hardware where you can bring your optimized stack and run them in the most optimal manner. Go a layer above. And that's where we are looking at how are we serving you in the data center or in the client? What softwares are we contributing to? And I'll talk about that in the next slide, essentially. And then with end to end applications, basically doing that in a very open manner where we have this customer first core value. We listen to customers, then we bring that in, and then we work with them together to create those solutions. So that open participation is important because that's what builds trust and that's what gives choice to the customers. So really bringing those solutions in an open choice and trust manner is critical. And when we look at generative AI, for example, when you have several developers, you wanna take them, make it easy for them to develop applications. How will you do that? So let's take a look at an overall AI stack that will get you AI everywhere. So again, at the bottom you have ubiquitous hardware, whether it's Xeon, there are 100 million plus Xeon installed all across the world. So that gives you a lot of compute power that is literally pervasive wherever you are on the compute continuum. So that's one part of it, right? Then on the open software layer, as I talked about, you pick an open source project. As you are building your GenAI apps, we are contributing to those projects that matter to you. If there is a project that we are not contributing to, we wanna hear from you. Tell us what could be different? What do we need to do different? How do we need to engage? That's exactly the charter of my team. My team works across Intel working with these different engineering teams that are contributing upstream and how are we rallying that together? Then at the top is really, is where all these large language models come in. And again, there are a lot of closed source models happening, but we are a big believer of the open source models here. And so I'm gonna share some initiatives that Intel has been driving to make that more successful. More recently, Meta announced Purple Lama. So they did the first Lama, Lama 2, but then more recently Meta announced Purple Lama model, which is an open and a transparent model. And really, the Purple really comes from that concept of cyber security, where there is a red team, which is an offensive team, and a blue team, which is a defensive team. Red and blue makes purple. So it's a purple Lama. So they really leverage that concept of what are they doing for red team and what are they doing for blue team? And that's where the name Purple comes from. So in the Lama 2 Responsible Use Guide, they recommended that all, Meta recommended that all inputs and outputs to the LLM be checked and filtered in accordance with the content guidelines appropriate to the application. So really what they're doing is, they're announcing something called as Lama Guard. That is an openly available foundation models to help developers avoid generally potentially risky outputs. So you can look at the inputs, you can see what guards need to be done, and then announce it. So on the left, for example, you can see, it shows the pipeline here. How you can determine the use case, fine tune the product, and as you are fine tuning for your product, this is where the pipeline kicks in, right? That's where you are saying, Purple Lama, what are the safety limitations that you're providing over there? And you keep doing that until you believe that your model is safe. And then again, when you are giving out outputs, at that point again as well, you are injecting the Purple Lama model to make sure the output is clearly, clearly safe. So it's really an open project, umbrella project, for open trust and safety tools. And we're very excited about this because we've been partnering with Meta on the launch of this Purple Lama model. The other initiative that I wanna share with you is the AI Alliance. This alliance was really created led by Meta and IBM, essentially. And I'm really excited about this personally because what it gives is a community of researchers, technologies, companies, and end users collaborating together to make sure that the AI stays open, stays safe, and responsible. And in AI Alliance, really, there are three key elements and how do we build and support open technologies across software, hardware, tools? And then how do we enable developers to be successful in it? And then how do we advocate for open innovation? How do we work with the standards and the foundations and the federal government? Because Executive Mandate came out that, hey, as these large companies are building foundation models, they need to report the results of their pen testing, security testing, all of that to the federal government. So this is really going by that mandate, essentially, that was given by the federal government. So lots of good partners over here. What we're gonna do is basically define benchmarks and tools, create evaluation standards for open model releases, responsibly advance the ecosystem of open foundation models, and supporting global AI skills building, because is how do we grow that total addressable market is a key element as well? Another element that Intel announced earlier this year was the Unified Acceleration Foundation. Now Intel has had one API, which is basically provides a single API to a range of accelerators, CPU, GPU, NPU, whatever is FPU, FPGA, whatever your underlying accelerator is, so that you don't have to change your application based upon the accelerator that you're running at. And that was Intel's, that is Intel's effort of one API. We contributed that to a third party foundation, this is a Linux Foundation Foundation, essentially. And so we really committed to open and scalable acceleration. And again, very excited to partner with a bunch of steering members over there. There's a steering committee, and I highly encourage you to take a look at it, join that foundation if accelerated computing matters to you. The last part that I wanna really talk about is the responsible AI, because we believe that it has to be the underlying foundation. We have long recognized the importance of ethical and human rights implications that are associated with the development of technology. The reliance of developers and general customers on the GenAI technology is becoming so much more prevalent if this is not done in an ethical and a responsible way, we're really adding a lot more danger over there. So what we do is we address this by utilizing a rigorous multidisciplinary process based on key principles throughout the development lifecycle. And now, make sure this must be done in a responsible manner, in a very open manner. Open ecosystem is the foundation of anything and everything we do at Intel. And that's what basically drives that entire initiative forward. Just an example of how we are creating that open ecosystem. We've been working with Huggingface for over three years now. Lots of the models over there are trained. For people who don't know Huggingface is basically provides an open repo for open models. So people can go take a look at it, what model is there, fine-tune it. And Intel has worked with them very closely to encourage that ecosystem more and more. Intel Developer Cloud is again another element. We talked about sort of the pervasive availability of zeons all across the compute continuum. But Intel Developer Cloud is where you will get to see the latest pre-production hardware. And you can bring your latest open source software and see, hey, how this is gonna work on Sapphire Rapids was launched earlier this year. Emerald Rapids is coming, then Granite Rapids are coming. That's the Intel hardware product line. But how could I run that in a pre-production hardware? And that's what Intel Developer Cloud is. Eventually, once you get it up and running over there, then you can move over. You can either stick to Intel Developer Cloud or you can move over to a hyperscaler when those hardware components are available over there. But it gives you an ability to test your workload in a very pre-production environment. So I just wanna conclude that we would like you to be, we would like to be trusted partner in bringing AI everywhere. We wanna accelerate innovation with broad AI software portfolio. I have that compute ubiquity across the space, whether it's CPU or GPU or accelerator, deploy anywhere and every stage of your AI journey on an Intel platform. We believe we can help you succeed your business. We can help you work with your customers and make it more successful. And this is the last obligatory slide. And that's a wrap for me. Any questions? Thank you. I'm gonna be out in the hallway if you have any questions. Thank you so much. Thank you.