VDURA has made its VDURA Data Platform V12 generally available, with a release focused on running storage as a multi-tenant, API-driven service for GPU clouds, AI factories, and “neocloud” operators. V12 is also offered as a qualified solution on Supermicro Building Block Solutions, with VDURA describing scaling from eight to 100,000 GPUs on a single software stack.
Technically, V12 is an extension of VDURA’s HYDRA architecture (High-performance, Yield-optimized, Distributed, Resilient Architecture) aimed at the operational mechanics of renting GPU capacity: isolating tenants, automating provisioning, and keeping GPU nodes fed with data. Multi-tenancy is built around per-tenant QoS, namespaces, encryption keys, and VLAN isolation on a shared fleet. Automation hooks include REST APIs, Kubernetes CSI, and infrastructure-as-code workflows for tenant provisioning.
A central feature is “Context-Aware Tiering,” where VDURA says data lands on the right media automatically as access patterns shift between training and inference. The company also claims that roughly 90% of files stay on flash while roughly 90% of capacity settles on HDD within one platform, with no stub files, rehydration steps, or manual tuning. For inference, V12 adds “persistent context,” described as a KV cache that outlives the pod so sessions can resume rather than prefilling again.
On the datapath side, V12 supports RDMA data paths for direct GPU-to-storage transfers. VDURA’s DirectFlow parallel client is specified at roughly 191 MB of DRAM and zero CPU cores on the GPU node. For metadata, V12 includes an “Elastic Metadata Engine,” with VeLO metadata acceleration described as up to 20x improvement and rated at 225,000 creates and deletes per second per Director, plus “billions of metadata operations per second” in aggregate.
V12 also combines file and S3 in one platform, where VDURA describes an S3 object as a file in the volume rather than a copy. Platform features listed include instantaneous, space-efficient snapshots for checkpoints and recovery, plus SMR HDD optimization that VDURA says can unlock 25% to 30% more capacity per rack. Security features include AES-256 encryption at rest and in flight, plus KMIP key management per tenant.
Supermicro-qualified configurations called out by VDURA include the AS-1116CS-TN (1U, single AMD EPYC 9005 series processor, 12 NVMe bays) used as both the VeLO Director node and an all-flash F-Node, alongside the AS-2015HS-TNR hybrid storage node and the CSE-947HE2C 4U 90-bay JBOD for the mixed-fleet data plane. VDURA describes RDMA connectivity straight to GPU nodes with no dedicated back-end storage fabric.
“Neocloud and AI factory operators told us exactly what they need from storage: keep the GPUs fed, isolate the tenants, automate everything, expand capacity for cold data without a second system, and move that data back to flash the moment it warms up for extended context,” said Ken Claffey, CEO of VDURA.
VDURA says V12 is generally available now for all V5000 class systems and as an upgrade for V11 customers. It also says VDURA Sentinel is included with V12, with proactive service requests and parts dispatch delivered through VDURACare Premier, described as a 10-year support offering covering hardware, software, and 24×7 expert response under a single contract.
Source: VDURA
