Dell’s AI story electrified by Lightning

Dell has uprated its AI game, with Nvidia-powered AI engines and Lightning extreme scale parallel storage to deliver a 4-layer boost to its AI factory story

Its AI announcements at Nvidia’s GTC event emphasize that flash is not the only storage media you need. Rather, Dell argues, customers need separate file, parallel file, and object storage storage personas; high-quality and highly-available data; andbetter, natural language-oriented data search facilities. 

Travis Vigil, SVP, ISG Product Management, Dell Technologies, said: "The number one problem enterprises face when moving AI pilots to production is curating the data they already have and putting it to work. The Dell AI Data Platform with Nvidia automates the entire data lifecycle and delivers the speed and scale AI workloads demand. We've done the integration work, so customers deploy faster, scale with confidence and see real returns."

Dell has built a 4-layer AI Data Platform with Nvidia, featuring three storage engines at the base: PowerScale for normal file access, Lightning for parallel file access, and ObjectScale for object data access.

Lightning is a whole new parallel filesystem product, separate from clustered PowerScale, and built for extreme high-performance. The three storage engines are for the attention of IT admins.

Dell positions the three in a hierarchy, like so:

It emphasizes that no one storage system can meet all of a customer's AI needs. It is ridiculous, it says, to use highly expensive flash, heading towards a 21:1 price premium over disk, for bulk data and archival storage. A highly parallel and extreme performance system like Lightning is needed to keep GPUs busy for training but wasted, relatively speaking, on inferencing. In that field, there are separate needs for file storage, such as Dell's PowerScale, and object storage, such as Object Scale.

Dell’s David Noy, VP of Product Management told us: “Why do we use both? We see a lot of object storage adoption in the Neoclouds and at scales that file systems simply wouldn't get to. For example, we're now seeing things like one-exabyte or three-exabyte type environments that are being asked for in some of the larger Neocloud environments. You would never want to do that with a clustered file system. Think about it – that could be potentially a thousand nodes. I don't care if it's us or Pure or VAST or you name it. You don't want to put a thousand nodes together in a file system where you have to worry about things like cash coherence and lock management. That's just insane. Object storage is way better solution for billions and billions and billions of objects at that scale.”

Although PowerScale has an S3 interface, “we would never tell a customer to go build a 3 exabyte Power Scale. That would be insane.”

Noy tells us: “There is an extreme level of training required for building foundational models. Tens of thousands of GPUs put together that require parallel file system performance that's unbelievably performant out of a small data set. That's what Lightning is.”

Lightning provides 150 GBps of read throughput from a 1 RU enclosure, and, Noy said: “it requires three network cards just to hit that performance because we're bottlenecking on the network here.”

Geeta Vaghela, who oversees strategy and planning for AI & technical computing with Dell High Performance Storage, enlarged on this: “About 97 percent of whatever line rate we can get, we are able to saturate. And we've done a bunch of testing to be able to scale out clients and number of NICS to see that saturation point continue as we scale up.”

Stick a bunch of these in a rack and, Noy said: “This thing is insane … The intent is going to do roughly six terabyte per second per rack.” (The match stacks up too; 40 x 150 GBps = 6,000 GBps.)

Dell VP and CTO Rajesh Rajaraman wanted us to know, ”It's not just about supporting multiple terabytes per second, but it's actually the case for Lightning is multiple terabytes per second in the smallest footprint. PowerScale and Object can support the same performance as well, but it's the rack density” that matters.  “The expectation is that there's only one rack, one rack of storage for thousands of GPUs, one rack, that's it.”

Noy explained more about Lightning’s architecture: “PowerScale uses OneFS," he explained. "The Lightning product uses a much more optimised solution for small block IO because that's what you need in these parallel file systems when you're distributing lots of little chunks of data across the file system. And it is also designed to have absolutely zero copy. So RDMA directly from the GPU to storage. Think of this as a Lustre competitor.”

Additionally, he said, “This is four times the performance of what an ObjectScale or even PowerScale can do per node. So it's wicked fast. But what you also have to remember is not everyone needs this level of performance. This is off the charts crazy. So what we're telling our customer base is 95percent of you are going to be just fine in a PowerScale/ObjectScale world. You'll be more than happy. This is for really crazy AI training environments or some HPC that's off the charts like fluid dynamics and nuclear simulations and things of that nature.”

He said: “Now data is going to move its way between PowerScale as a performance tier to Lightning. So the data stays on Lightning, it's accessed, it's streamed into the GPU, zero copies directly RDMA. When you're done with it, you want to move it back to PowerScale or back to Object Scale, one or the other.”

Schedulers will be used for the data movement here.

The second layer comprises data engines for analytics, streaming, unstructured data search, and a vector database, targeted at data engineering people.

Rajaraman told us this about the search engine: “Our search engine not only does just semantic search, which is one of the things people talk about, but it also does lexical search in combination.... What’s special about that is it actually unifies vector similarity. It does keyword precision, it does metadata filtering, it does hybrid ranking all in one engine, which is what the customer's use cases are. It's not just a one-trick pony of any kind, like pure vector databases do for example.”

He was keen to explain the analytics engine features as well: ”It's not yet another analytics engine. It's an engine built on Trino, [and] has been in the market for a very long time. It does industry-leading parallel and distributed joins, which is a hallmark of current engines. It does this across heterogeneous databases and sources. For example, it has 100 plus connectors to bring data into your system as opposed to having to copy it. For example,, it does distributed joint optimizations and consistent query semantics across all the incompatible storage and data engines.” 

There’s more: “The SQL capability that's built on top of the analytics engine allows for a very seamless natural language query interface. [It’s] really about democratising the ability for people to start leveraging SQL as their way into their enterprise data sources.” 

Above this is a new data orchestration engine aimed at data scientists, to integrate data pipelines, models, and humans, and automate workflows. AI engineers are the personas Dell has in mind for its top layer, comprising RAG, agents, chatbots and blueprints.

The storage engines, data engine and data orchestration engine layers are deeply enmeshed with Nvidia software libraries and products: GPUDirect, CUDA, CMX - the new name for Nvidia’s Inference Context Memory Storage Platform (ICMSP), and SpectrumX and Connect-X networking. 

Rajaraman said this about CMX: “Essentially, it's context memory systems, which allows us to start using KV Cache and other technologies to essentially help LLM inferencing workloads where context windows are becoming much, much larger as reasoning models become much more token-rich. So those things are starting to provide significant performance benefits. We are19 times faster here, with H100s, and internally, obviously we're looking at even faster numbers than that as we keep going with faster GPUs, faster storage.” 

And: “Basically we're blowing past any other benchmark that Nvidia is working with us on.”

Dell is announcing that Nvidia’s cuDF is used to speed data ingestion in a data processing engine, with cuDF also integrated into its data analytics engine, and cuVS featuring in its data search engine.

Vrashank Jain, Director of Product Management at Dell, said: “cuVS is a CUDA library that allows vector indexing to operate on top of GPUs. That essentially accelerates one of the biggest bottlenecks in ingestion - vectorization of documents, which is building the graph. And so that allows us to provide about a 12X performance impact compared to traditional CPU approaches.”

He added: ”cuDF, which is part of RAPIDS, allows us to bring about three times faster data processing within VDA GPUs. And we're also starting to look at cuDF on the analytics engine as a fast follow here to provide more accelerated SQL capabilities on GPUs as well.”

The data orchestration engine – Dell loves engines  – is hooked up with Nvidia’s NVAIE marketplace, NIMs microservices, and Nvidia Blueprints.

NVAIE (Nvidia AI Enterprise) is a cloud-native AI software suite, certified to deploy anywhere and run on common virtualization and container orchestration platforms. It’s described as the operating system of the Nvidia AI platform. Dell says:

• It supports the latest Nvidia AI-Q blueprint for customizable AI agents. Nvidia-accelerated data engine integrations in the Dell AI Data Platform enable high-performance data preparation, retrieval, and reasoning pipelines across structured and unstructured data. There is a growing library of pre-built blueprints and NIM microservices, along with the Nemotron 3 Super model on Dell Enterprise Hub on Hugging Face. 

• Dell will also support Nvidia STX, a new modular reference design powered by next-generation Vera Rubin NVL72, BlueField-4 DPUs, and Spectrum-X Ethernet networking. 

Lastly, Dell is supplying a single exascale storage hardware base that can have ObjectScale, PowerScale or Lightning storage software personas executing on it. 

Noy explains: “[We’ll] give you a single hardware platform, and we're announcing that this week, that basically that server platform will allow you to deploy any personality on it. So if you want to put object on it, you want to put PowerScale on it, or put Lighting on it, that's fine. But that server is a performance-optimised server, and it will allow you to deploy the software personality that you need. So from a customer perspective, they buy the CapEx once. They buy the server once, and then they decide which personality makes sense for them. It's great for CSPs” too.

The licensing scheme will facilitate this. Noy again: “Some of our customers are purchasing today for what they need three years from now. They have no idea what the correct mix of file and object are. So, by giving them a single platform that can home either product ,and by giving them a licence model that allows them to fluidly change the mix of one software personality or the other and the provisioning tools to provision them, they can go, "Okay, you 300 servers, you're going to be ObjectScales. You, 25 are going to be PowerScales and you five off in the corner are going to be Lightning. No, wait a minute, I changed my mind. I only want 200 objects and I want 150 for file.”

“You can mix and match and you can change your mind later, but that license that we give you is going to be a completely mix-and-match license that you can choose what you want today and, tomorrow, if you change your mind, you can completely change the ratios.”

Dell Lightning has support for Nvidia CX-8 and CX-9 SuperNICs and planned network connectivity up to 800GbE. 

Read more about today's announcements here and here.

Availability 

• Dell Data Orchestration Engine and Marketplace are available in Q1 CY26. 
• Dell and Nvidia Blueprints are available now. 
• Dell support for Nvidia AI-Q Blueprint is available now. 
• AI Assistant for the Dell Analytics Engine will be available in 1H CY26. 
• Nvidia GPU-accelerated data processing and data indexing in the Dell AI Data Platform will be available in 2H CY26. 
• Dell Lightning File System will be available in April 2026. 
• Dell Exascale Storage is targeted for availability in early 2H CY26. 
• Dell support for Nvidia’s latest innovations will roll out throughout the year. 

Bootnote

An Nvidia RAPIDS website says "cuDF (pronounced 'KOO-dee-eff') is a Python GPU DataFrame library (built on the Apache Arrow columnar memory format) for loading, joining, aggregating, filtering, and otherwise manipulating data. cuDF also provides a pandas-like API that will be familiar to data engineers & data scientists, so they can use it to easily accelerate their workflows without going into the details of CUDA programming."

Nvidia states cuVS “is designed to accelerate and optimize vector index builds and vector search for existing databases and vector search libraries. It enables developers to enhance data mining and semantic search workloads, such as recommender systems and retrieval-augmented generation (RAG). Built on top of the Nvidia CUDA software stack, it contains many building blocks for composing vector search systems and exposes easy-to-use APIs for C, C++, Rust, Java, Python, and Go."

New testing demonstrates that PowerScale's software-driven Parallel Network File System (pNFS) architecture delivers up to 6X faster performance with large files in enterprise AI environments compared to NFSv3.