AI-Infused Applications with AMD’s Matt Unangst

This is Textron tv. Hey guys, thanks for the throw. We're here with Matt Unangst, who's senior director for, uh, workstations and client systems in the commercial side of a MD.

And we're talking about this survey that they did about, there's a gap between what's happening with AI on the business side and perhaps where the IT folks are, and we're gonna dive into that in a minute. Matt, welcome to show. Great to be here.

Thanks for, uh, making the time. So, Matt, it kind of feels like the business side is having, shall we say, a certain amount of irrational exuberance around what the possibilities are for ai. And the IT side of the house is looking at it and saying, well, hold on there, Kathy, there's a few things we gotta do first before we get to this promised land.

So what's your read on where are we on this journey? Is this just another example of the, of the divide between IT and the business, or is something else afoot here? Uh, it's a great question, Mike, and I think it's a, a challenge that a lot of organizations are trying to work through right now.

I think on the business side, there's a lot of potential around some of these AI solutions businesses see the opportunity to make their employees more productive, to give them more tools, um, to make 'em more efficient and, and make better decisions on a regular basis. And a lot of the, the promise of ai, you know, shows high probability that we're going to get there. At the same time, IT organizations right now are looking at what is a very new technology, and they're trying to fit it into, uh, you know, constrained budgets, uh, constrained data center, you know, footprints.

And they're working through a lot of the typical, you know, IT focused areas such as security and deployment. Um, and so, you know, there is a bit of a gap right now between the desire from the business side and where IT departments are, uh, from a deployment perspective, but that's a gap that I think we're gonna see close, um, you know, relatively rapidly over the next year or two. Now, your survey that you guys did kind of suggested the IT side of the house is trying to build something that feels like a five year plan.

The business side is, of course, they're all the same. Their biggest fear is being left behind and they're worried that the competitors are gonna move forward. Are these two goals incompatible or can we kind of get to something in the short term but still have a long-term plan and, um, we can be adults about and kind of count on as what we want to do?

I certainly think there's a, a number of near term opportunities here. I mean, if you think about, um, even some of the early deployments of a I PCs, uh, these start to deliver some incremental benefits around certain applications that we use every day, like video conferencing, for example. Um, with some of our AI technologies, we're able to deliver longer battery life, which does translate into a better employee experience and in some cases productivity.

Um, so there's gonna be some near term opportunities to take advantage of the early versions of this AI technology. Uh, but of course, the businesses are always going to be pushing for the next greatest thing and the next new, you know, advancement in productivity. And so I think it's a space where we're gonna see the businesses pushing their IT departments to, you know, advance their capabilities, advance their solutions, um, as rapidly as possible.

And I would not expect that that's a trend that we see change, you know, for three to five years is I think, you know, this AI journey that we're on. Uh, we're in the very early stages of this and I expect that it's gonna continue to be pretty dynamic, uh, for the foreseeable future. Do you think this will drive a wave of infrastructure upgrades both on the client and the server side?

Because it seems to me a lot of this AI stuff is really data intensive and we're not quite set up to handle that. The AI stuff is very data intensive. These AI models, you know, uh, take up a, a tremendous amount of, uh, data and compute capability.

Um, and so yes, there is going to have to be an upgrade cycle around both the data center and the client side of things. Um, frankly, I think, you know, deploying a lot of these AI tools is going to be scalable only if a number of the, the tools and applications are available on the client side or on the endpoint and the edge. Um, but I also think there's gonna be a big focus on, uh, building out the data center capabilities as well.

And, um, you know, as we look at, you know, different regions, uh, you know, are, are different, but we still see that data center capacities are, are very high today. And so that's one of the challenges that, you know, a lot of organizations have with deploying these tools is they need to consolidate their data centers to more efficient servers so they can open up space that then allows them to, you know, actually use that space and power capability for some of these ai, uh, specific applications. What do you think that the inference engines are gonna look like for these things?

Because we talk about AI all the time and people are like, well, we gotta train them and we got all these GPUs, but on a practical level, it's the inference model is the thing we deploy, and that may be only in terabytes and maybe getting more efficient as we go along. So, um, do we have a firm understanding of what's required when we talk about ai? I think we do.

I think it's an evolving space, and I think it's a space where a lot of these AI models we're seeing become more efficient almost daily. Um, we're seeing, you know, new models come out. You know, I think if you take a step back, almost, you know, really about a year ago, you know, chat GPT came out and it kind of, it, it shocked the world in terms of just the, the user experience and the capability that it had.

It's amazing how far we've come in, you know, just about a year. And now you have a lot of, you know, additional models or new models that are smaller in terms of their data sets, more targeted in terms of their use cases. Um, and that's going to allow, you know, some of these capabilities, um, and experiences to scale much more broadly, um, than, you know, anybody really thought possible again, you know, six months or a year ago.

So I do think, um, you know, there's a lot of new capabilities there. I think getting more and more efficient around these models while we continue to advance the hardware that we deliver. Um, again, across the data center, the edge and the endpoint, that combination is going to allow a lot of these AI tools and models to be more deployable and more scalable.

Uh, instead of being, you know, isolated just to the data center. It seems like this fits in with a larger trend where we're trying to, um, process and analyze data at the point where it's created in the first place. And the AI model needs to sit alongside that.

Um, who's in charge of kinda understanding those infrastructure questions and who's kind of being the architect for all of that? Well, first off, I, I do agree with the trend of, you know, more and more of these models wanting to process the data kind of where they're at. I think there's a lot of considerations there.

When you push a large sets of data up into the cloud, that can be time intensive. You could kind of get a slower response. Um, and so you in a lot of cases get a better user experience or faster, uh, and better performance by doing it local.

Um, of course organizations also have security considerations or privacy considerations when they think about, you know, processing this data locally versus pushing it into the cloud. Um, in terms of who's in charge, it's really up to each organization, um, and their IT department to figure out what's the optimal way for them to deploy a set of AI solutions. And one of the things that we believe from an A MD perspective is there's no one size fits all, um, approach to ai.

So certain models are more ideal to be run in the data center, others are gonna be more ideal for edge solutions, and yet others will be more ideal for like the PC endpoint, um, type of device. And so it's really, you know, our guidance when we work with our customers is figure out what your AI strategy is, figure out exactly what AI tools you want to deploy, and then let's have a conversation around what's the right place to go deploy those tools depending on what you're trying to accomplish. And there's a lot of nuance in that.

'cause while I may use GPUs to train the AI model, the inference engine can run almost on anything. And sometimes those GPUs are expensive or hard to find. So, um, will people need to figure out that balance of, um, training versus execution of the AI model or kind of completely different motions, Our different motions.

And, and especially on the inference side, I think, you know, there are certainly opportunities to deploy those inference solutions with G GPUs. Um, there's also, you know, we increasingly are, uh, you know, upleveling our rise, what we call rise in ai, and we have a, a dedicated inference accelerator, uh, built into our mobile processors and, and now our new desktop processors. And there's a lot of models that are ideally run on that, uh, that accelerator that we call an NPU.

Um, so some of these models are gonna be best run on those types of inference, accelerators, other models may be best run on A GPU. Um, and we're working closely with a large set of ISVs to figure out for each of their applications and models what is the best, you know, hardware and and model for them to, uh, to run on. And so that's something that, you know, is an, a big investment focus for us now and will continue to be for, you know, really the, the next few years.

Right. Does this mean hardware's gonna be cool again? I think for a long time we took it for granted, but you know, there's a lot of subtlety in this space now.

There Is, there is no question that this is driving a significant uptick in innovation around hardware design. And I think that's true across, you know, the entire portfolio. Again, the data center, the edge, the endpoint.

You know, we're even having conversations around how we can use some of our, uh, expert class workstation products as as really ideal, uh, edge AI solutions just because of the memory footprint that they support and the size of the data models they can support. Um, so it is driving, you know, a very different conversation. It's not just kind of the typical rinse and repeat that we've seen for many years.

Um, and so as we continue to see advancement in the software, in the models, the hardware's gotta be right there with it. And so that's a big focus for us is to make sure that the hardware stays exciting, relevant and, uh, really kind of leading edge. So based on the survey results and what you see going on out there, what's your best advice to organizations to bring together all their hardware and software folks to kinda drive this?

Not to mention the few data scientists that are running around and some data engineers, but how do I wrap my arms around all these people? Step number one is you have to figure out how you want to use AI to enable your business. Um, and that's, that's just foundational because if you have certain goals from a business perspective and you can figure out how AI tools can help you achieve those goals, then we can start to have more targeted conversations on what is the right set of hardware, what are the right set of applications and AI models that you wanna deploy.

Um, and I think where, you know, a lot of organizations have struggled, at least in the very early stages of this, is there's a lot of AI hype, there's a lot of kind of, you know, hand waving, talking about AI is gonna revolutionize all of these, uh, you know, experiences and the employee capabilities. Uh, but it's been very vague, right? And so taking that conversation, getting down to a very specific set of goals around how we want to use this technology to help the business, then we can translate that into very tangible, you know, next steps in terms of figuring out the right hardware and the right applications.

All right, folks, I heard it here. All the software in the world doesn't run unless it has some good hardware underneath it. So we gonna have to do these things together hand in hand.

Hey Matt, thanks for being on the show. Thank you, Mike. I appreciate it.

All right, and back to you guys in the studio.