VDURA V12 Targets GPU Clouds With One Storage Plane

VDURA has made its V12 storage software generally available for GPU-cloud and AI-factory operators, turning the former Panasas platform into a multi-tenant, API-driven service layer for large accelerator fleets, Blocks & Files reports.

The release is aimed at providers that need to isolate customers on shared storage without splitting performance, capacity and archive data across separate systems.

V12 uses the company’s HYDRA architecture to combine per-tenant quality-of-service controls, namespaces, encryption keys and VLAN isolation on one fleet, so a single storage pool can be carved into tenant services with capacity and performance guarantees.

Automation is central to the product shift.

REST APIs, Kubernetes CSI and infrastructure-as-code provisioning are meant to let operators deploy, provision and bill storage through the same pipelines used for the rest of a GPU cloud.

That moves VDURA further from its PanFS hardware heritage and into the software-defined storage market it entered after Panasas rebranded in 2024.

The technical argument is that GPU infrastructure needs one data plane that can follow data as workloads move between training, inference and colder retention.

Context-aware tiering keeps roughly 90 percent of files on flash while about 90 percent of capacity settles on hard drives, without stub files, rehydration steps or manual tuning.

File and S3 access also sit inside one platform, with an S3 object treated as a file in the same volume rather than a staged copy.

V12 adds features designed to reduce wasted accelerator time.

A persistent key-value cache can outlive a pod, allowing inference sessions to resume instead of being prefilled again.

RDMA data paths move traffic directly between storage and GPUs while the DirectFlow parallel client uses about 191 MB of DRAM and no CPU cores on the GPU node.

Metadata and resiliency are another part of the pitch.

VeLO is positioned as the acceleration layer for small-file and namespace-heavy work, with VDURA citing a 20-fold uplift, per-Director create/delete rates of 225,000 a second, and aggregate metadata activity reaching the billions each second.

Snapshots support checkpoints and operational recovery, SMR drive optimization can unlock 25 to 30 percent more rack capacity, and AES-256 encryption covers data at rest and in flight with per-tenant KMIP key management.

Availability covers V5000-class hardware and V11 upgrade paths.

VDURA says qualified Supermicro Building Block Solutions can scale from eight to 100,000 GPUs on one software stack, pairing 1U AMD EPYC 9005-based systems for director and flash roles with hybrid nodes and a 90-bay, 4U JBOD in the shared data plane.

Capacity can grow online from three nodes to thousands, with operators changing the flash-to-HDD ratio by adding all-flash nodes, hybrid nodes or expansion shelves independently.

In one cited configuration, a roughly 20 PB usable system spans a 35-times performance range from a 2 percent-flash, 18 kW capacity-optimized fleet to an all-flash setup at 1,000 MB/s per TB.

The operating condition is straightforward: V12 tries to let providers tune cost, power and rack space without creating another storage stack when workloads change.

VDURA claims more than a twofold watt-efficiency gain and TCO reduced by over 60 percent versus rival designs at an equivalent data feed rate.