VDURA DATA PLATFORM / V12

The High Performance Data Storage Infrastructure for AI Factories & Neoclouds

One platform for the entire AI data pipeline.

Ingest, training, checkpointing, inference, and long-term retention on one platform, with isolated namespaces for multiple tenants. Flash for speed, HDD for capacity, one control plane over both, on commodity hardware.

Talk to VDURA Read the V12 White Paper

✓ 2.7 TB/s per rack, all-flash ✓ 6 nodes to thousands ✓ Up to 8 nines durability · 6+ nines availability

THE VDURA DATA PLATFORM

Up to 2.7 TB/s per rack · isolated namespaces for multiple tenants

2.7 TB/s

PERFORMANCE

Per rack, all-flash, with 45M IOPS. A true parallel file system keeps GPUs saturated.

60%

ECONOMICS

Lower TCO frees up capital for more GPUs. Shared-nothing, software-defined, on commodity hardware.

1 day

SIMPLICITY

To deploy, minutes to expand, and roughly 0.5 FTE to operate the whole system.

WHY A PLATFORM AND NOT A TIER

The GPUs are the budget. Storage decides how much of it you keep.

A 5,000-GPU cluster running at 98% storage availability loses 876,000 GPU-hours a year, roughly $2.6 million in idle accelerators. Most storage answers this by adding a fast tier in front of a cheap one, which means two software stacks, two failure domains, and a copy of the dataset in each.

THE TWO-SYSTEM TAX

A fast tier bolted to a cheap one

A separate flash system in front of a separate object store means two software stacks, two consoles, two support paths, and a dataset copied into each of them.

VDURA: One control plane, one data plane, one namespace across NVMe flash and SATA HDD. Nothing to copy between tiers.

THE IDLE GPU

Availability priced in accelerator hours

When storage stalls, GPUs wait. At cluster scale the loss is measured in hundreds of thousands of GPU-hours a year, not in downtime minutes.

VDURA: Client-side dual-parity erasure coding keeps I/O serving through node and drive failures, with no rebuild storm to ride out.

THE HARDWARE PREMIUM

Paying twice for the same resilience

Controller pairs, dual-ported drives, and firmware RAID exist to make a proprietary array survive a failure. You pay that premium on every expansion.

VDURA: Shared-nothing nodes and software protection mean standard servers and single-port drives are sufficient.

THE HYDRA ARCHITECTURE

A true parallel file system, in four layers.

HYDRA separates metadata from data and protection from hardware. That separation is what lets the same platform serve a metadata-bound inference fleet and a bandwidth-bound training run without reconfiguration.

DATA ACCESS

DirectFlow Client

True parallel file system client: POSIX, cache-coherent, with erasure coding computed on the host. Every GPU server reads and writes to every Storage Node at once, and NFS, SMB, and S3 are served from the same namespace.

METADATA

Control Plane · VeLO

A distributed, flash-optimized key-value metadata engine on Director Nodes, handling billions of inode operations. Metadata never competes with bulk data for the same media, which is what keeps small-file and inference workloads from collapsing.

DURABILITY

Data Plane · VPODs

Virtualized Protected Object Devices abstract NVMe flash and high-density HDD into erasure-coded pools. Shared-nothing storage nodes own their own data, so there is no controller pair, no RAID firmware, and no failover set to buy in pairs.

PLACEMENT

VDURA Context Aware Tiering™

Flash is a performance medium, not a capacity medium. Every write lands on NVMe at line rate; placement then follows file size, access pattern, and temperature with zero manual tuning, and cold records drain to HDD without ever going offline.

ACROSS THE PIPELINE

The same namespace, tuned by access pattern rather than by tier.

INGEST

Land data once

Write through POSIX, NFS, SMB, or S3 into the namespace the training job will read. There is no staging copy and no import step.

One landing zone

TRAINING

Keep the GPUs fed

Parallel striping and RDMA paths sustain the sequential bandwidth large runs need, while VeLO absorbs the metadata load of millions of small samples.

Bandwidth and metadata together

CHECKPOINT

Write bursts without a stall

Checkpoints are a synchronized write storm from every rank at once. Client-side erasure coding spreads that burst across nodes instead of funnelling it through a controller.

No controller bottleneck

INFERENCE

Serve from the same data

Model weights, embeddings, KV cache, and RAG corpora are read from the namespace they were trained in, with metadata-heavy random access served off flash.

No promotion step

RETENTION / DATA LAKE

Keep everything, online

Cold records drain to the HDD tier without leaving the namespace. S3 access and V12 snapshots make the same data a governed lake for the next run.

One tier down, never offline

VDURA FOR AI FACTORIES AND NEOCLOUDS

Three workloads, one data platform.

AI TRAINING

Keep every GPU fed through the run.

Parallel striping, RDMA paths, and VeLO metadata absorb both the sequential bandwidth of large runs and the small-file storm of millions of samples, with checkpoints that never stall the fleet.

Bandwidth and metadata together Explore →

AI INFERENCE

Serve models from the data they were trained on.

Low-latency random reads for weights, embeddings, and RAG corpora, served off flash in the same namespace. No promotion step between training and production.

Sub-millisecond metadata Explore →

PERSISTENT CONTEXT

Persist KV cache beyond GPU memory.

Session context and KV cache land on the platform instead of being recomputed, so long-running agents and multi-turn inference resume where they left off.

Resume, not recompute Explore →

PLATFORM CAPABILITIES

Nine capabilities, one V12 software package.

FEATURE

Multi-Tenancy

Isolated namespaces, per-tenant QoS, quotas, and access policies, with the same API driving provisioning and chargeback.

Read the page →

FEATURE

API

REST, Kubernetes CSI, and CLI interfaces for every provisioning and telemetry operation.

Read the page →

FEATURE

RDMA

RDMA data paths that cut host CPU out of the GPU feed.

Read the page →

FEATURE

Resiliency

Multi-level erasure coding at client and node tiers, serving through failures without a rebuild storm.

Read the page →

FEATURE

S3-compatible object access to the same global namespace the parallel file system serves.

Read the page →

FEATURE

End-to-End Encryption

AES-256 at rest and in flight across every tenant data path, with key control staying inside your environment.

Read the page →

FEATURE

VDURA Context Aware Tiering™

Placement by file size, access pattern, and temperature across NVMe and HDD, with zero manual tuning.

Read the page →

FEATURE

Open Hardware

Certified commodity servers, flash, and HDD from multiple vendors, bought at your price with no mark-up.

Read the page →

FEATURE

VDURA Sentinel™ Support

Fleet telemetry on a dashboard in your network, with proactive service requests opened before a ticket exists.

Read the page →

THE ECONOMICS

Commodity hardware, by design.

Because protection runs in software across shared-nothing nodes, the platform needs no controller pairs, no dual-ported drives, and no vendor-locked media. Open hardware at your buying power, with zero mark-up applied by VDURA.

✓ Standard servers, single-port NVMe SSDs, SATA HDD, and standard Ethernet or InfiniBand

✓ Certified across AIC, Supermicro, Dell, Western Digital, Seagate, Phison, Kioxia, and Solidigm

✓ Add flash for throughput or HDD for capacity independently, online, with no data migration

✓ Ride both media cost curves as prices move, buying only the resource that is short

✓ Software subscription and hardware are separate line items, so neither one hides the other

✓ On-node data reduction: compression and dedupe run inside Storage Nodes, never on clients, toggled via GUI or CLI

✓ V12 snapshots for point-in-time protection of datasets, checkpoints, and model versions

HOW IT IS CONSUMED

V12 software SUBSCRIPTION

The full platform, including the parallel file system, protocols, multi-tenancy, encryption, Context Aware Tiering™, snapshots, and data reduction, as one package.

Certified hardware YOUR PROCUREMENT

Buy the servers and media through your own channel at your own pricing, or have VDURA quote it. Either way there is no VDURA mark-up on the metal.

VDURACare Premier 10-YEAR COVERAGE

VDURA Sentinel™ Support reads fleet telemetry against every failure pattern resolved across thousands of systems, opening service requests before a ticket exists. Device replacement is included for ten years.

VDURABILITY GUARANTEE

VDURACare Premier underpins the VDURAbility Guarantee: contractual availability and durability SLAs across the whole platform, plus 10-year coverage with proactive device replacement built in, independent of which vendor's drives are in the chassis.

IN PRODUCTION