Software-defined storage vendor Vdura unveiled the latest version of its flagship platform with additions aimed at capturing demand from GPU clouds and so-called AI factories.

Data Platform Version 12 can scale between eight to 100,000 GPUs on a single software stack. The vendor touted improvements including up to 20-times faster metadata, context-aware tiering so data moves between flash and hard drives based on how it's used, and more than 60% lower total cost of ownership versus competing architectures at the same feed rate.

The platform is available for Vdura's V5000-class data storage arrays and can be deployed in a multitenancy configuration with a single storage pool split across isolated tenant environments, each of which comes with their own performance guarantees, encryption keys, and virtual LANs (VLANs). Vdura touted the platform for firms renting GPU access, like neoclouds.

The crux of Vdura’s V12 update is ensuring better performance per watt and lower cost, with GPUs kept busy while using less power and rack space.

“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," Vdura CEO Ken Claffey explained. “V12 is that list, shipped. It is the same mixed-fleet, software-defined model the hyperscalers run inside their own clouds, delivered on Supermicro systems our customers already buy. Every watt and every rack unit we give back is another GPU the operator can put into service.”

Central to V12 is its High-performance, Yield-optimized, Distributed, Resilient Architecture (HYDRA) that the company says it has extended with the multitenant, API-driven capabilities GPU clouds require. For example, the vendor claims a single platform of roughly 20 petabytes usable can span a 35-fold performance range, from a capacity-optimized mixed fleet with 2% flash drawing 18 kilowatts to an all-flash configuration delivering 1,000-megabytes per second per terabyte.

V12 also features what Vdura billed as remote direct memory access (RDMA) "data paths” that enable direct transfers between the GPU and storage, cutting out the CPU. The vendor’s DirectFlow parallel client, which runs on the GPU server, eliminates the need for an extra hop to the processor, a process Vdura claims takes roughly 191 megabytes of dynamic random-access memory (DRAM) and no CPU cores.

“Every kilowatt spent on storage is a kilowatt not spent on compute," the vendor noted. "V12 was engineered for that operating model: keep every GPU fed, keep every tenant isolated, keep every cluster online, and do it on hardware operators already standardize on."