IBM is to begin shipping the latest release of its parallel file system, aimed at users managing AI inference workloads.

According to HPCwire, the tech giant will begin shipping Storage Scale 6.0.0 this month, with software licensing for the service to be made available this week. The solution runs on the Nvidia-certified Storage Scale System 6000, IBM’s all-flash hardware storage offering, and as a parallel file system, stores data across multiple networked servers to allow multiple and simultaneous access for workloads.

In its latest update, the solution includes the Storage Scale System Data Acceleration Tier (DAT), which uses high-speed NVMe-over-Fabric (NVMeoF) technology to provide ultra-fast, low-latency storage access for real-time AI inferencing.

According to IBM, DAT uses Storage Scale’s Asymmetric Data Replication technology to address input/output operations per second (IOPS) limitations, which can hinder real-time AI inference performance, with delays or dropped data resulting in tangible performance effects.

With the asymmetric replication support, the DAT maintains two complementary data replicas, with one a performance replica that enables fast read access, and the other a reliable replica to ensure data protection.

The solution supports two deployment models, both of which keep the reliable pool on the Storage Scale System 6000. In the Centralized DAT configuration, the performance pool is also hosted on the 6000 system and accessed through NVMeoF. This approach aims to simplify management and provide the IOPS performance required for most AI workloads.

In contrast, the Distributed DAT configuration places the performance pool on client-local storage, with performance determined by each client’s disk setup and compute resources. This model is designed to deliver maximum performance for AI workloads with exceptionally high IOPS demands.

Storage Scale 6.0.0 also includes a one-button GUI for upgrade installations, supported by improved prechecks and a unified protocol deployment that aims to simplify management tasks.

Also touted by IBM were API-driven improvements to the control plane to facilitate automation for advanced capabilities such as quota management. Diagnostics for issue resolution have also been enhanced to make root cause analysis and remediation more efficient.

The release maintains certified high-performance solutions for GPU-accelerated workloads, including content-aware storage (CAS), and on the Nvidia side, container-native storage access (CNSA) support for GPUDirect Storage, enhanced Base Command Manager support, Nvidia Nsight integration, and Nvidia-certified storage certifications.

IBM’s solution comes as AI inferencing becomes a dominant use case, joining similar offerings from Weka, FriendliAI, and Akamai Technologies.

The interest in inference has risen with the demand for enterprise AI. Lumen last year expected the next wave of AI market pressure to come from the AI model inference phase, with demand led by financial services, retailers, and healthcare.

In an interview with SDxCentral from this year, the company predicted that the AI inference push will spill over into hybrid architectures, seeing data held on-premises and the use of inference models associated with trained models in the cloud.

On the hardware side, the inference push has seen UK startup Fractile benefit from a $15 million funding round last year, with the firm promising faster AI model inference on its chips.