Future of Data Platforms Summit

More from Future of Data Platforms Summit

Ed Beauvais

HPE

Director, Product Management AI & Cloud Data Infrastructure

Info
Transcript

Ed Beauvais, HPE

Ed Beauvais

HPE Director, Product Management AI & Cloud Data Infrastructure

In this Future of Data Platforms Summit segment, Ed Beauvais, director of product management for AI and unstructured data at HPE, joins theCUBE’s Rob Strechay to unpack how enterprises are rethinking storage and data platforms in the AI era. Beauvais details how object storage is experiencing a renaissance, driven by the demands of AI workloads that require high performance, scale and intelligence. He highlights the launch of HPE’s X10000 platform and its role in transforming object storage from a passive data repository into an active engine for unlocking in... Read more

Ed Beauvais, HPE

>> Welcome back to the Future of Data Platforms Summit. In this episode, I'm joined by Ed Beauvais, who's the director of product management, AI, and unstructured data. Welcome, Ed.

>> Great. Thanks for having me, Rob. Really appreciate it.

>> Yeah, I think what is fun about this is that there's just so many trends in the industry. What are some of the trends that you've been watching, especially as you look at AI and data platforms and what people are looking for, organizations are looking for, that are acquiring AI data platforms?

>> Right. So one of the things that we see is that from a storage perspective, object storage has been reborn with this era of AI. So when we think about it, the protocol's the same, but basically everything else has changed. So we think about it's S3, but you need a new architecture because there's new challenges. And what we're seeing is that customers need new solutions to the AI challenges that they had. They need things like high performance and scale and data intelligence capabilities. So we're really excited about that, as well as seeing the need for working with the NVIDIA ecosystem. So lots of things happening from a storage perspective as it intersects with AI, and we've got lots of interest, and we've really focused on addressing some of the key needs there.

>> So I totally agree with you. I think it's definitely breathed a lot more life back into object storage. I think we're seeing a renaissance of that, to put it mildly. But there are really a lot of different options for object storage. Even HPE had some other ones, and now you've kind of gone from that old stuff to the new stuff. When you introduced the X10000 nine months ago, why did you decide that you really needed to bring this out as a data platform versus just object storage?

>> Right, and it's a great question. And what we've thought about is that the fundamental need is not just necessarily the protocol; it's really helping our customers unlock the value of unstructured data. For us, that's sort of the core mission. And we've got a number of solutions within the HPE portfolio; as you know, it can be confusing, but what we focused on is really thinking through what do customers need from an AI platform, what do customers need in terms of being able to access and unlock the value of their unstructured data? And we've got some great products and solutions with our data fabric and the X10000 together that really helped to unlock that. But one of the things I wanted to highlight, Rob, is we're thinking about it differently. From a storage perspective, for too long, all storage did was write data and hopefully read it back. And from our perspective, storage solutions need to do more. That's why we focused on introducing data intelligence that's in line in the platform. So as you're landing data, we can help to unlock that value. And I think that is really one of the core tenants of what we're trying to do. When I said originally we want to help customers unlock value, we want to help them unlock value so that they can get to that inference, that time to first inference, much faster than they've been able to do it historically.

>> Yeah, and I think, again, for people who don't see it, there's the whole Alletra MP platform and there's GreenLake Cloud as well that goes along with that. How does that kind of fit in as well with this entire space?

>> Yeah. So one of the things that HPE is doing is obviously we want to give our customers a really easy way to consume, so we've got GreenLake Cloud. And then at the bottom of the stack from a hardware perspective, we've got our disaggregated hardware platform; that's the Alletra MP hardware platform. And the way we think about it is that you can take different storage personalities and put them on top of that common disaggregated hardware platform, and we'll give you one management plane to sort of view everything and to manage that data from a cloud perspective.

>> And that's also where the data fabric comes in and where you bring it all together as well.

>> Exactly. Then you can use that. One of the things that we've talked to customers about and it's a real challenge is you've got disparate assets that are distributed everywhere. And one of the great things about the data fabric is it gives you heterogeneous access to all of your data. So if you can see it, then you can actually use it for AI. If you don't know about it, it's actually a risk.

>> So, Ed, help us understand some of the use cases where you see that's bridging beyond even just traditional AI.

>> Great. Great question, Rob. Certainly we think AI is critical, and obviously we're partnering with NVIDIA to be able to enable unlocking that value, but we also see customers don't just buy solutions only for AI. We see great use cases in analytics and to think about an active data lake or a data lake house. Those are areas where we want to help customers unlock the value of their unstructured data as well. So you start with AI, but you can also use these types of solutions for active analytics. And we see unstructured data becoming much more like primary storage, where it becomes critical to amass that unstructured data and then query it and use the AI in the querying to be able to take action and drive business insight. And then by the way, once you've done that, you need to back it up and you need to protect it. And not only that, you're at risk from potential cyber attacks. So we want to help customers with these architectures get back up and running as quickly as possible. So you can use it for AI, data analytics, as well as data protection.

>> Yeah, and we were kind of talking about this a little bit earlier on where that was really stacking up, because we have our little thesis there that is the storage, there's the data services layer, then there's the data management layer. You're really bringing a lot to bear in all of that. With the data services, it's really important, and I think we don't hit on that enough that you want to protect the data. There's still disaster recovery and, like you said, bad people out there and things happen. But you also hit on the data management a little bit, and there's things like a lot of, I would say, open standards that you're working with, because I think you mentioned things like Spark and Iceberg and things of that nature.

>> Yeah. So as part of our data intelligence layer, one of the things that we do is we can start to look at that data in line as it flows into the system. Let me give you an example, Rob. So if unstructured data is coming in, we can do some sampling to maybe make a determination of what type of data that is. And based on that type, we can then filter and say, "You know what? We should run a model." And once we've run that model, we want to take the value that's been extracted by running that data model and expose that out so they can be either used in either AI or analytics. And one of the things that we do is that we can expose that data. We're actually sort of managing that metadata as parquet, but we expose it through an open-table format so anyone can consume it. And that's what I meant by enabling the querying of that unstructured data. And we can help you with Apache, or you can bring your own query engine. And again, open-table format's great because it really allows our customers to have the flexibility to use whatever engine makes sense for them.

>> Yeah, I totally agree with that. What kind of performance is required for AI workloads? Because I think object, in particular, used to get a bad rap from a performance perspective. What are the required IOPS and bandwidth considerations that organizations should really be thinking about?

>> Great question, Rob, and certainly... Let me just go back for a second and compare it to the legacy storage architectures. When we think about the shared-nothing architectures, if you want performance or if you want capacity, there's really one answer, right? It's you add a node. And with the disaggregated architectures now, you've got the ability to really fine-tune based on what customers need. If I need more performance, I can add compute. If I need more capacity, I can just add that capacity. I don't have to make a compromise in terms of how I've deployed. Now in terms of AI, the answer is, and I'm hesitating a little, the answer is it depends, right? Because the AI workloads are so varied. But really what it's unlocked is it's really unlocked the need for much more transactional types of performance. In terms of an IOP, we would say it's an object operation per second, and that's the new IOP. So we look at being able to deliver massive performance in terms of transactions, and we also need to do that from a bandwidth perspective. So gigabytes-a-second per node is the type of metrics that we look at. And one of the things that we're doing is we're partnering with NVIDIA. And I'm really excited; our current release that's just about to come out actually has support for RDMA for object, which is really a game-changer. If you need the maximum performance from an AI perspective, now you can do it with object.

>> Yeah, and that's important because you have these GPUs, which tend to cost a lot, and you want to keep them busy from that perspective as well. It would assume that that also helps with the RDMA stuff, will help with latency as well, which a lot of times is time to first token and things of that nature as well.

>> Yeah, so RDMA does a number of great things. It obviously gives you two capabilities. It gives you GDS, which is GPU direct or access directly to that GPU memory, but you can also use it for other applications. And certainly in the AI context, it's going to lower latency. The other thing it does is it dramatically lowers the CPU consumption, right? You're getting the TCP/IP stack in this case out of the way, and you've got a much more optimized... And we've seen 95, 98% CPU reduction from a host or a target perspective.

>> Yeah. I think that, to me, makes a whole lot of sense. You talked about the X10000 having a mission to make data really AI-ready. What does that mean? Because that can mean a lot of different things. Kind of tell me a little bit more about where you see data intelligence going here.

>> Yeah, so I would say what we see is that we're in a privileged position. People are storing their unstructured data assets on our platform. And I'll give you sort of two proof points, and this is how we think about it from a data intelligence perspective. One is we want to help customers get access to the metadata, and that could mean vector embedding extraction, it could mean understanding and running lightweight AI models against that data, particularly if it's image data, so getting more metadata out of that data and doing it faster. And we want to do it in a way that's open. So the way that we think about it, for metadata enrichment, you might run a model, but if you can't access that metadata, what good is it? So from our perspective, we're delivering that in open-table format, so you can bring your own compute, or we'll help you from a Spark perspective. The other thing that we're doing is we're optimizing RAG pipelines. And I think one of the key challenges with RAG pipelines is actually security. People are really worried that the LLM can have access to data that it shouldn't have access to, and we address this challenge. If you think about using IAM or security at a data layer, and you do that all the way through the pipeline, including the vector embeddings, you can eliminate that risk from an LLM perspective; that is, the LLM will only have access to the data it's authorized to, so you have no risk of the data leakage challenge here.

>> Yeah, I think one of the neat things when I was talking to some of the folks on the HPE team was that exact stuff, where you're talking about the pipelines that are embedded in the system itself to help people get to time to value of AI. Talk to that a bit.

>> Yeah, so let me talk about that a little bit. One of the things is a little bit different. You think about the bucket as a place where you land data. Think about the bucket as a pipeline. And now if you've started to think through that, one of the things that we're doing with data intelligence is we're going to embed MCP, which is model context protocol. So now if you've got that embedded in your storage environment, what that means is now I can get access to that metadata. Actually not me, but our agents, your AI applications, your AI agents can now just, instead of listing data that's in a bucket, we have the opportunity, I think, in the future to give you more context or to give those agentic AI platforms a lot more to work with. And that's going to allow for, again, faster time to insight, getting that data AI-ready, and enabling AI agents to really maximize the value out of those unstructured data assets.

>> Yeah, and it would seem... Because I think also one of the pipelines that you have in there for the bucket also does the vectorizing as well and can help because that brings you further down the readiness for AI perspective.

>> Yeah, one of the things we wanted to do is make it easy for customers to chat with their data. So you've got a lot of different elements in these RAG pipelines. Of course, you've got the storage, then you've got metadata management, then you've got to extract the vector embedding, so you need a vector database, you need an AI application or an LLM. So tying all these things together and getting them to be coordinated can be a challenge, and one of the things that we're doing with our data intelligence capability is we simplify that. With just a few lines of code, you can get access to the data in those buckets. That can be the pipeline, and everything's integrated together so you can be chatting with your data in no time.

>> Yeah, and I think what a lot of people... When we talk to platform engineering groups, which is the new word for IT and things of that nature, that's kind of where storage... And there's a lot of times not a dedicated storage person as much anymore, so what they have to know is wear multiple different hats. This would seem that this makes it easy for those organizations that are looking to get into this and want to build that application, that first AI or agentic application. They can bring this in, and it's a lot easier to get access to the data, to bring it up and set it up and doing all of this because you're also tied in with a lot of the manageability and observability as well.

>> Yeah, and I think, Rob, that's a great point. And what we see and what we hear from our enterprise customers is they really want solutions. They want to make it easy, easy to manage, easy to deploy the solutions. And I think the great news from our perspective, whether that's AI, whether that's data analytics, a data lake house, all those things are really enabled with this next-generation infrastructure for both unstructured data, as well as enabling these applications, and that's our goal. From a customer perspective, we want to make it easy to consume, easy to take advantage of this. And then certainly when we think about data analytics as it relates to AI, they're really closely related. So that also is another factor in considering an AI platform; that is, are you going to be able to do and get the analytics out that you need as well?

>> Yeah. And I think I know what they're doing on the compute side as well and bringing a lot of that to bear and how it all works, like you said, in a solution. Is there a single answer that solves everything for storage for AI that really provides a solution that's flexible and can really have performance, scalability, and that data intelligence? What do you say?

>> Rob, I think it's such a great question, and unfortunately there's no easy answer, right? The answer is no, but I think the great thing about HPE is we've got a range of solutions, and our goal is to find the right solution for our customers. And when you think about from an AI perspective, the workloads are so varied. There's such a wide range in needs, and that's the great thing about working with HPE. We want to focus on what are your needs as a specific customer, and then we can find that right solution for you. We've got great innovation here, but we've also got a range of solutions that we can get you the right one for your AI needs.

>> Yeah. I think we definitely see it that way as well. Ed, I really thank you for coming on board today. This has been great.

>> Rob, appreciate the time, and we're looking forward to any follow-up. And certainly if people have questions, you guys know how to track me down.

>> Yeah, and we'll put something in the description there where they can find you. And thank you for watching this episode. Stay right there for the future of Data Platforms Summit. We'll be back.