Red Hat Summit + AnsibleFest 2025 | Chris Tobias, Intel & Ryan King, Red Hat

Clips
More from Red Hat Summit + AnsibleFest 2025

Chris Tobias

General Manager, Americas Technology Leadership and Platform ISV Account Team

Intel

Ryan King

Senior Director, Hardware Partners

Red Hat

play_circle_outline Exploring Intel and Red Hat's Collaboration History and the Impact of Enterprise AI with Gaudi 3 Accelerators

play_circle_outline Open source approach for AI innovation

play_circle_outline Security measures for customers in regulated environments

play_circle_outline Cost-effective AI solutions and business process transformation

Info
Transcript

Chris Tobias, Intel & Ryan King, Red Hat

Chris Tobias

General Manager, Americas Technology Leadership and Platform ISV Account Team Intel

Ryan King

Senior Director, Hardware Partners Red Hat

Chris Tobias, general manager of Americas technology leadership and platform ISV account team at Intel, and Ryan King, global head of AI and infrastructure ecosystem at Red Hat, join theCUBE’s Rob Strechay and Rebecca Knight at Red Hat Summit 2025 to explore the enduring Intel and Red Hat alliance and its impact on enterprise AI development.

Tobias and King examine how their collaboration continues to push open-source innovation forward, with new advancements like the Gaudi 3 accelerators and Intel Xeon processor integration leading the charge. The d... Read more

explore Keep Exploring

What is the history of collaboration between Red Hat and Stack, and what are they currently collaborating on? add

What drove you to become involved in the open source community and how did you transition into the cloud computing industry? add

What are some examples of the collaboration between Intel and Red Hat in terms of security measures for customers operating in highly regulated environments with strict compliance needs? add

What is Red Hat working on with the goal of minimizing complexity in hardware architecture and infrastructure software to make it easier for businesses to focus on their applications and business problems while also being cost-effective? add

bolt Powered by CUBE AI

Chris Tobias, Intel & Ryan King, Red Hat

search

Rebecca Knight

>> Hello everyone and welcome back to theCUBE's live coverage of the Red Hat Summit, AnsibleFest here in beautiful Boston, Massachusetts. I'm your host, Rebecca Knight, alongside my co-host and analyst, Rob Stretchay. Rob, this is what is always true about Red Hat Summits is that you hear about the partnerships, the collaboration, the deep storied history between many companies.

Rob Strechay

>> I agree, and I think this is another one that has a great story leading back to what was announced on the keynote earlier today.

Rebecca Knight

>> Well, indeed. So let's with that, I'd like to introduce our next guest. We have Chris Tobias, GM, Americas Tech Leadership, and ISV team, Intel. Thank you so much for coming on the show.

Chris Tobias

>> Thank you, Rebecca. Thank you, Rob. Very pleased to be here.

Rebecca Knight

>> And Ryan King, senior director, hardware partners at Red Hat. Thank you so much.

Ryan King

>> Thank you.

Rebecca Knight

>> Excellent. So why don't we talk a little bit about this Red Hat, Intel collaboration. Just as we said at the very outset, it is a long history.

Chris Tobias

>> Yeah, it's a very long history. We've been collaborating together all the time from RHEL hitting the marketplace up until now. Stack's latest benchmark launch with Xeon and RHEL 9.5 and set a world record. So fantastic collaboration up until this day. And now we're collaborating on enterprise AI and with our latest Gaudi 3 accelerators. And as our CEO of products said today, Michelle Johnson, "Enterprise AI is Red Hat." So we are very closely tied together in our technology strategy and the belief in open community and software for AI.

Ryan King

>> It's an amazing milestone for the partnership.

Rob Strechay

>> Yeah, I mean, it seems that way. I think again, the collaboration has gone on for a while and you have a lot of joint POCs going as well. Why don't you talk to how the reaction has been from the customers because... And we look at it and I think it's been, I would say, one of the themes of the stage today was really inference. And what we look at, there's a long tail of inference where it's not just going to be Gaudi 3 and accelerators, it's going to be a lot of XENON and CPUs doing inference, especially as you get towards the edge and things of that nature. I mean, hundreds of POCs sounds pretty substantial at this point in time.

Ryan King

>> Yeah. So our POC framework has different technical use cases and now that VLM becomes more central and on the stage for Red Hat, we're seeing a lot of interest for VLM-based POCs from our customers. And for those that don't know VLM, it's a layer that sits between the models and the accelerators and it provides all those useful capabilities for them to run efficient and for things like memory management. If you read the original paper on it, it was all about page memory attention and taking concepts that are known in operating systems, but applying them to a GPU. So it made really simple for a model to be able to make itself ready day zero for how it can best run on an accelerator.

Rebecca Knight

>> So you think Red Hat, you think open source. And I know that this collaboration is really committed to fostering an open ecosystem for AI innovation. Can you talk a little bit, both of you, in terms of how you actually do that?

Ryan King

>> So we're open source first and always. It's kind of just our DNA. We think in the AI era. I draw back. When I started, I actually worked 10 years at Intel and got drawn into the open source community. We actually lived in the same town. There's more history than just the company here. But back then I was working early at Red Hat and I was working on Red Hat Enterprise Linux. It was this big enterprise push for Linux and we're stabilizing it, making enterprise ready. And then I started to hear about what was happening in the cloud and I heard about a project from Google called the Borg and I was like, "Oh God, what is that?" And then people told me the technical capabilities of it and it was like, "Holy cow, this thing can run a cluster, a single computer, it does all this great thing with containers." And he really technologically got interested in it. And it was our then CTO, Brian Stevens that worked with a few other Red Hatters and they went into Google and they brought that project out and that became Kubernetes. So there's this innovation that happens in different places and then we start to work with the people that are the innovators and we say, "Hey look, how can we bring a more open approach to this?" And then once that becomes an open project and it becomes a standard for everybody, we work on kind of our standard pattern, which is how do we make that enterprise ready, hardened and ready and suitable for the Global 1000.

Rob Strechay

>> When you look at this, and you brought up vLLM and we had Brian on earlier actually, and we were talking very deep into some of the stuff around it. It would seem, and even from an Intel perspective, this is extremely helpful in getting more out of it because what customers seem to be concerned about is ROI and TCO in addition to the performance that you're talking about.

Chris Tobias

>> Yeah, absolutely. It lowers the complexity and ideally the cost of AI and it connects the business case to the technology a lot faster at the high level. So that's where we want to go, that's what we're working towards. And by making and supporting it through the open source methodology, we expect to create a rally point for the industry to innovate around. So it's fantastic. And as Ryan described earlier, that's the playbook we're following. It's been successful before. It's going to be successful again.

Ryan King

>> Yeah, simplifying that approach for both the front end, the model, people creating the models and say, they just want it to work well and work quickly on the accelerators, the accelerator people to expose those features and next set of capabilities, new quantization, those sort of things, new capacities in the GPUs, that's all great to bring those teams together with a single layer that we can expose. And then again, back to our pattern, we'll support it, make it enterprise. And now we're announcing LLMD, which is a layer upon that about how you can do distributed inference, which just takes it to the next level.

Rebecca Knight

>> How are you thinking about security, which remains a top priority for so many customers? How are Red Hat and Intel teaming up, particularly for customers who operate in highly regulated environments and have a lot of strict compliance needs?

Chris Tobias

>> Yeah, security is top of mind for Intel. For example, between Intel and Red Hat, we've launched some support for security measures that are inherent in the Intel silicon that will lower the attack surface as you're running your workload. So there's a technology called TDX and it basically lowers it to that application. So you have a very small attack surface there, for example. So we have extensive collaboration and both Intel and Red Hat, it's deeply embedded in our DNA as how to minimize the security risks and minimize the attack surface as we put products out in the marketplace.

Rob Strechay

>> Yeah. So I think again, and there's some exciting stuff with RHEL 10, around the post-quantum encryption stuff that, again, really exciting as well. But how does the-

Ryan King

>> The post-quantum era, it's there.

Rob Strechay

>> Yeah.

Ryan King

>> It's fascinating.

Rob Strechay

>> Well, I'm sitting there, I'm like, "Okay, now we're going to go from 3000 years to 10 seconds of cracking stuff with quantum." I'm like, "I'm glad... It depends on whose quantum computer you think and which methodology. Somebody says there's one out already, others are saying it's out, whatever. Say we're three to five years away from it being there, we still see a lot of companies really getting ready. But another thing they're getting ready for and I think maybe, and you can talk to this, is really how Red Hat AI and RHEL AI and OpenShift AI really take advantage of Gaudi 3 and XENON and how does that integration really come to bear?

Ryan King

>> So VLM predominantly focused on GPUs today, but the company started in the CPU space. So we will continue to evolve how this works across all sorts of accelerators. Today we are showing VLM with Gaudi 3 in I think two locations here on the floor and the capabilities and the performance and really price performance and also performance per watt that Intel's very proud of.

Chris Tobias

>> Yeah. So we just went through GAA with IBM Cloud with our Gaudi 3 accelerators. A third-party company, Signal Six measured cost per token and token output. Very competitive with all the other leading market accelerators out there. And compatibility is there for the full Red Hat stack from RHEL through Red Hat OpenShift. And it's something we've collaborated on. And one of the things we're very deeply collaborating on is making sure that VLM layer, it looks seamless for people who are using VLM no matter what the hardware looks like underneath that.

Ryan King

>> We think of it as the kernel is to an application is to the CPU. Like, VLM is in between the model and the GPU. So it's like a new paradigm of Linux but now around vLLM.

Chris Tobias

>> And then as Ryan was saying, it originally started around CPUs. Well, at the time, CPUs didn't have some of the accelerated technologies we put in there for our advanced matrix extensions, which is big cache and a big matrix multiply, which is perfectly suited for generative AI inference. And we're getting good results. But we've had people actually migrate their workloads into the cloud from accelerators to CPUs because of the cost and performance capability of both.

Rebecca Knight

>> So what is that unlocking for customers in terms of the transformations that you're seeing and how you work with customers? What opportunities is this creating for them?

Chris Tobias

>> So I think it's addressing two areas. It's minimizing complexity. Because you hear, "Generative AI, I got to do something, what do I do?" And then, "How do I connect my business process to the technology?" And you can just get lost in that trouble. So what we're working with Red Hat to do is minimize that complexity, and what is the hardware architecture and what does all the infrastructure software look like and make that kind of seamless so you can just worry about, "Hey, what kind of application do I want to go with and what kind of business problem do I want to solve?" And then ideally that gets you into a cost- effective solution because right now you can spend your entire budget on AI and if you're not sure what you want, you're not sure what you're going to get. So by minimizing complexity and minimizing cost, you make it much more accessible to effective enterprise business processes. And that's what they wanted to do to be more effective and compete in the marketplace.

Ryan King

>> I think it depends who you're speaking with. And I'm an ecosystem, so when I talk to a GSI, it's all about them speaking to the CEO and the C staff and saying, "Here's the business processes we want to transform." And they're looking down into the agentic architecture and into the infrastructure and say, "What's the best way to implement?" And they have patterns and blueprints around that. And we pipe into that. And then I talk to the hardware folks and it's about how do we create repeatable patterns and then how do we expand the aperture? Because if you're looking at your infrastructure overall, you're going to look at both vert and AI and containers in one space. And what's unique and differentiating about Red Hat is Kubernetes is a management plane across all of those is really an amazing way of looking at your transformation because when we talk to partners, they say, "Here's how we can take our current footprint in terms of power and we can bring this down to a better power footprint profile and we can make room for what we need in the AI space and look at it as a row of racks versus just a distinct, what's my vert strategy, what's my containers, what's my AI?" So from infrastructure-wise, we feel like we have a really strong position with our lead with OpenShift across all three areas.

Rob Strechay

>> That would seem like, and we've talked about this a little bit, and I think with the current political things, sustainability has kind of fallen out of favor, but I think saving power and watts is not out of favor. Because data centers are not getting more power. I mean, Eastern Virginia can't build data centers in Eastern Virginia right now because there is no power. So that must be one of the focuses as well.

Chris Tobias

>> It absolutely is. So when you do the TCO calculation, power becomes the ongoing biggest contributor of that. And right now with the AI options out there, power is following an exponential curve depending on where you choose. So vLLM is actually a key tool in making the model size more effective at getting more out of the model size and getting better power performance overall. So it's easier to implement for the IT shop. It fits in their ongoing OPEX budget, because people used to think about CapEx as they moved to cloud, went to OPEX. The power equation is a big, big deal and it comes up almost immediately no matter who you're talking to. It used to be just at the cloud providers, now it's everybody.

Rob Strechay

>> Yeah. Having been in one of those cloud providers in a past life, I can tell you everybody is. My friends who run data centers today are trying to figure out what governments they have to talk to get more power for some of their data centers, which is insane.

Ryan King

>> It's interesting because it's like the neocloud world of, "Hey, we're looking at inference itself." They're seeing what we're doing with LLMD and they're seeing how it's a better way of serving up a bunch of inference overall. So they're having conversations with that as well as sovereign AI and what they do in that space and they love the security and the ability that it's Red Hat software, it's all open source. And so it's really kind of opened the door some new strategic conversations between us and different partners that maybe have been just building stuff and how they're saying, "Hey, what you're doing here is differentiating." And to go back to your question on the power thing overall, on one end it's like we just continue to optimize this and these are great strides forwards, but also on the other hand, reasoning models come in and agentic and multi-model interaction and we're looking specifically at VLM and around reasoning because we want to... I mean, you see this go up eight X in tokens. You use a deep research model, it does not come back quickly. Right?

Rob Strechay

>> No.

Ryan King

>> So there's a lot of work still to do because the technology just continues to take strides and taking more and more compute in order to achieve its means.

Rob Strechay

>> Yeah. No, I think to me, and the final kind of question from us is that, like you mentioned, the size of models and things of that nature and the compression that vLLM can do on the model size. There's also I think and we look at it, we have a power law distribution of the GenAI that I kind of mentioned earlier, and one of the things that we see is the long tail is inference. And what we also see is the size of models. Not going to be these trillion parameter models out at the edge or far edge or something like that. They're going to be smaller, specialized, more SLMs, smaller, specialized models out at the edge. This would seem like having both the CPU and the GPU to be part of that AI ecosystem really seems a good way to drive that efficiencies as well because again, it's a cut to your point, it was a cost and power and throughput type of discussion.

Chris Tobias

>> Yeah, absolutely. We're in the first rounds of generative AI discussion and the amount of innovation has been amazing. I mean, vLLM started, what, two years ago and now it's been a company that's been bought by Red Hat and fully integrated into their software strategy. So the rate and pace is amazing. And we've seen software techniques to get to smaller model sizes. We've seen hardware techniques to make that more efficient. So what are people going to use in their day-to-day inference architecture? I think that's still being figured out. Historically, up until now, inference on machine learning and deep learning was mostly CPUs. And then we got this ChatGPT moment and it took a huge model to expose the capability there, which was wonderful. But we realized in the day-to-day usage you could get away with much smaller footprints, which has implications, power, managing the security of the model, et cetera. So there's lots of ways for efficiency, which you could either take off the table as more capability or you could take off the table as TCO savings. So it's very, very exciting to be part of this, and the game is far from settled on how this will play out.

Rob Strechay

>> That's great.

Rebecca Knight

>> Yeah, absolutely. Chris, Ryan, thank you both so much for coming on theCUBE.

Chris Tobias

>> Yeah, thank you.

Ryan King

>> Thank you.

Rebecca Knight

>> Stay tuned for more of theCUBE's live coverage of the Red Hat Summit, AnsibleFest 2025. You're watching theCUBE, the leader in enterprise tech news and analysis.