(upbeat music) As artificial intelligence continues to gain prominence in IT, experts are raising concerns about its resource intensity, including the environmental costs of computation. Hi, I am Tom Shields, and in today's "Intelligence Report," we're going to discuss what we can do to more effectively optimize the resources used for AI. Here with me today are Russell Fishman from NetApp and Ritu Jyoti from IDC. According to "MIT Technology Review," training just one AI model can emit five times the lifetime emissions of an average American car. And the infrastructure costs for compute and storage are daunting. What capabilities can optimize compute and storage resources for cost efficiency and sustainability? I think one of the most important things is we wanna use those really precious GPU resources optimally. So there's certainly part of the end-to-end process, if you will, of training that involves me taking data and stuffing it into a GPU, right? I mean, that's an important part of it. And we got to keep those GPUs fully utilized. So saturated, if you will. So that's super important. Performance has to be good enough. That's kinda how I describe it. There's no value in being 150% when all it can consume is 100%. But ultimately, you know, finding ways to optimize your usage of these resources is critical. And what you don't want is GPUs sitting there sucking all the power down without being utilized. I hear this a lot from enterprises. They wanna find that optimal level of investment where they can combine resources, compute resources, on-prem and in the cloud, to drive theirprojects, which means, you know, I mean, of course, it comes back to a flexible data architecture, right? If you don't have that, you can't drive the compute resources where you need them to be. Say a little bit more about that premise and cloud being able to have the flexibility across both to be as efficient as you can be with the use of the resource. When I talk to customers, what I tend to hear is, they're looking for flexibility in the way that they approach the GPUs, right? And you know, this could be anything from difficulty in actually getting GPUs because of supply chain problems. It could be a challenge of not having the right data center space to have these GPUs on premises. They need a bursting capability, they need a short-term training capability, which might be best supplied by a public cloud. Whatever the answer is, the data needs to move in a very,seamless manner between these different environments. And it needs to be done in a way that is simple for the people who are actually doing the work. And that's the data scientists. That's not infrastructure people. As an industry, we've got to make that super simple for folks to take advantage of, super flexible, and very pragmatic, so that where you need the resources, and where you need the data to meet the resource, you can make that happen, and really easily. You're seeing that need as well? The last five years, you know, when we were seeing more traditional predictive AI being adopted, since then, I've consistently heard that cost is one of the biggest inhibitor. And data pipelines and data architectures, data challenge, always shot up in the top two or three challenges. And that has been exacerbated during generative AI, because it's very,hungry for data, and there's more and more additional data being created. All of this can add up to the chaos and the madness that we are talking about. But many a times, we always kind of, you know, gravitate to the fact that it's about compute. I would like to call that out that it's also very energy intensive, and every company that I'm speaking to in the last couple of years, sustainable AI is actually on top of their agenda. And when they're looking about how to kind of become more sustainable, and have more sustainable AI, they're looking through the entire stack. And it's not just the compute, it's also the storage, the networking, every part of it, right? So everyone has to play their part. But here today, we are talking about the data aspect of it, as to how can the storage play a very,critical part in bringing the right set of data to the right set of the models, at the right time, in the right quality, and not really increase the capacity? All these, you know, storing techniques, in terms of deduplication, creating efficient clones, all of that comes into very,efficient way of usage of the- It's also about the measurement itself. Exactly. The IT teams need a way to measure how much energy they're using, right? What I hear customers talk about, and Ritu's is absolutely right, having an efficient way to go about AI is critical, and having, the right tool for the right job, right? Lots of different types of data and different parts of the pipeline demands different tools. So there's a lot of focus on doing that at the design level. But of course, what's actually happening in reality when you're actually in production, and how you measure your impact, and how you can start tracking through metrics, your ability to meet your ESG goals, is super critical. So you know, what I hear customers talk to me about is this need for a dashboard or tooling that allows them to see the entirety of their environmental impact, and be able to make the right choices that lets them fine tune their use of resources. And getting back to the data, you know, obviously we're using vast amounts of data here, and copies of data. As far as managing those for efficiency, are you seeing that as an option for making this whole thing as efficient as it can be? I'll just play off something Ritu said. She started talking about some really critical foundational storage capabilities that are really useful in terms of efficiency when it comes to AI. And that means the efficient storage of data. We talked about the concept of dataset management, and the need, for example, for explainable AI to keep all the copies of the data, because when you train a model, that's how you go back and diagnose it. These storage functions, though, are storage functions. They're not functions that are intrinsically linked to the way that a data scientist does their job. So how you connect those two things up, so you essentially create the environment that allows data scientists to leverage these resources very efficiently, in this case storage, keep those copies of data, things like snapshots and clones, and make that much more efficient, and mean that you are not just saving data center space and capacity, but you're also saving electricity. And by the way, of course, at the end of the day, those two things, cost and sustainability, go very much hand in hand. So when you're being sustainable, you're generally lower cost, and that's obviously a great thing. That's an excellent point. Hey, thanks guys. Really important topic. As we further leverage AI everywhere, the focus will increase on new ways to triangulate and optimize across cost, efficiency, and sustainability parameters. The points we've discussed are just a few ideas. To hear more about how to leverage opportunities and mitigate issues facing AI adoption, tune into the other episodes in our AI series. (upbeat music)