Dell announced Project Lightning has parallelized PowerScale, to speed file read and write performance, along with more PowerScale and ObjectScale news, at SuperComputing 2025.
Dell says it, with these and other AI Factory announcements, and Nvidia are “delivering the future of enterprise AI with advancements to the Dell AI Factory with Nvidia.” It’s announcing multiple server, storage, and networking product enhancements and we look at the storage ones here. PowerScale is Dell’s scale-out, clustered file system, with its OneFS software, developed from DEll’s acquired Isilon precursor. ObjectScale is Dell’s distributed, microservices-based, multi-node, scale-out, and multi-tenant object storage software with a single global namespace that supports the S3 API. At SC25, Dell announced parallel PowerScale, PowerScale SW licensing, NIXL library KV Cache offload integration for PowerScale and ObjectScale, and AI-optimised search capabilities for ObjectScale.
Dell vice-chairman and COO Jeff Clarke said: “The Dell AI Factory with Nvidia solves the problem every enterprise is facing: how to move from AI pilots to production without rebuilding their infrastructure. We’ve done the integration work so customers don’t have to, which means they can deploy faster and scale with confidence.”
PowerScale
PowerScale’s Project Lightning has been a near 18-month development initiative to add parallel IO to PowerScale’s OneFS operating system, and, as speculated, it’s doing so by using pNFS (Parallel NFS), with a metadata server, and Flex Files layout support added to its data server nodes.
Parallel PowerScale has, Del says, 2-way communication between metadata server and clients, allowing for better parallel distribution of data across multiple nodes in a PowerScale cluster. This delivers throughput, performance gains and linear scalability with parallel I/O across multiple pathways in the cluster.
PowerScale now supports NIXL (Nvidia Inference Transfer (Xfer) Library); Nvidia code to accelerate point to point communications in AI inference frameworks such as Dynamo, its low-latency KV cache offload engine, and provides an abstraction over GPU and CPU memory, and file, block and object storage, with a modular plug-in architecture. PowerScale’s NIXL support enables scalable KV Cache offloading. It achieves a 1-second Time to First Token (TTFT) at a full context window of 131,000 tokens – 19X faster than standard vLLM.
Many storage suppliers, as well as Dell, have signalled Dynamo support, such as Cloudian, Cohesity DDN, Hitachi Vantara, HPE, IBM, NetApp, PEAK:AIO, Pure Storage, VAST Data, and WEKA.
PowerScale is now available in SW-only, subscription license form, on qualified PowerEdge servers, like the Dell PowerEdge R7725xd. This will help customers buy PowerScale compute facilities separate from PowerScale HW+SW appliances.
ObjectScale
The two AI-optimised search features are S3 tables and vector search APIs and intended to maker search faster for analytics and key AI workloads like inferencing and retrieval-augmented generation (RAG). S3 tables store structure data, while vectors are mathematical representations of unstructured data. The two together provide high-speed access to complex data stored directly on ObjectScale.
Amazon says “S3 Tables are purpose-built for storing tabular data, such as daily purchase transactions, streaming sensor data, or ad impressions. Tabular data represents data in columns and rows, like in a database table.” There are various ways to import data from relational databases into S3 tables. For example, you could use AWS Glue or EMR with a Spark job routinely moving the data. See an AWS overview here.
Availability
Dell ObjectScale and PowerScale with Nvidia Dynamo are globally available now. Software-defined PowerScale will be globally available in 1H26. Dell PowerScale parallel NFS support will be globally available in 2026.
Bootnote (1)
PowerScale currently runs the OneFS operating system, combining the file system, a volume manager and data protection into a single distributed software layer running on a OneFS powered storage cluster. The cluster has an internal or backend network, consisting of either 10 Gb, 40 Gb, or 100 Gb Ethernet, or low-latency QDR Infiniband (IB).
Clients connect to the cluster using Ethernet connections (10 GbE, 25 GbE, 40 GbE, or 100 GbE) that are available on all nodes. Additionally, OneFS 9.10 and later releases support either 200 GbE front-end connectivity or HDR Infiniband plus 100 Gb Ethernet on the all-flash F910 and F710 platforms.
A single cluster consists of multiple nodes, minimum 4, maximum 252, which are rack-mountable enterprise appliances containing: memory, CPU, networking, Ethernet or low-latency Infiniband interconnects, disk controllers and storage media. As such, each node in the distributed cluster has compute as well as storage or capacity capabilities.
Platform nodes come in three types: Performance, Hybrid/Utility, and Archive.
OneFS support for pNFS with Flex Files requires OneFS to adopt a separate Metadata Server (MDS) software entity, separate from the data serving node entities.
Bootnote (2)
At its Dell Technologies World 2024 event in Las Vegas, Dell’s Varun Chhabra, SVP for ISG Marketing, told his audience: “We’re excited to announce Project Lightning which will deliver a parallel file system for unstructured data in PowerScale. Project Lightning will bring extreme performance and unparalleled efficiency with near line rate efficiency – 97 percent network utilisation and the ability to saturate 1,000s of data hungry GPUs.”
Adding pNFS to PowerScale achieves that.
However, Project Lightning has another aspect to it. A Dell spokesperson told us there is “a proprietary client-side driver for representing the entirety of the filesystem as ‘local’ to the client. This requires more resource consumption on the client, and management of said proprietary driver. The benefit means extreme direct-to-drive performance to fully saturate the network–even with fully random reads. Lightning is a ground-up Parallel File System with a targeted application to satisfy the very highest performance requirements where the infrastructure team has the ability to optimize the full stack.”
