IBM has a broad, multifaceted approach to AI that encompasses platforms and models (e.g. watsonx, Granite), open-source initiatives, a range of education, training and consulting capabilities, and a comprehensive partnership strategy with key ecosystems players such as NVIDIA.
In-house technology development is a core pillar of IBM’s approach to AI, and its efforts in the storage and data realm are a prime example. On March 27, the company launched a “content aware storage” (CAS) capability that will form part of its hybrid cloud infrastructure offering, IBM Fusion. It will be offered as part of IBM Storage Scale (formerly known as General Parallel File System, or GPFS), IBM’s high performance, scale-out file system, and will also leverage a range of NVIDIA capabilities, including the NVIDIA Data Platform, NVIDIA Spectrum-X networking and NVIDIA NIM.
The aim of CAS is to enable enterprises to more fully take advantage of AI inference capabilities that sit on top of LLMs but leverage their own, proprietary data sets through capabilities such as retrieval-augmented generation (RAG).
Currently, the task of preparing and ingesting enterprise data for RAG is typically costly and time-consuming. Data must be copied from its original source into multiple other destinations—data lakes, cloud services and so on—for preparation and vectorization. This limits both the volume of data that enterprises are able to move into RAG, and the frequency by which they can refresh this data, which ultimately limits the value of the inferencing process, yielding low-quality results, high costs, increased risk (through multiple copies) and operational challenges.
By contrast, IBM believes that by “bringing the AI to the data” it can address all of these limitations in one fell swoop, drastically improving the quality of responses in the process. Its approach with CAS essentially transforms IBM Storage Scale from a passive storage system storing 1s and 0s into an intelligent infrastructure layer that’s an integral part of the AI data preparation process.
Leveraging significant innovations from IBM Research around natural language processing, CAS extracts semantic meaning from unstructured data—such as PDFs, chats, emails, audio and video files, legal, financial, and other business documents—from within the storage infrastructure itself. This reduces the number of steps required for inferencing, meaning it can be done more quickly, more frequently, and with greater overall efficiency, by minimizing data movement and latency. A neat aspect is that users can set up “watch” folders to identify data changes as they occur, helping ensure that data is always current for AI applications. What’s more, Storage Scale’s support for third-party storage systems—such as from Dell, NetApp, and others—means that CAS will be able to support data stored in a wide range of corporate storage repositories without having to move or migrate this data.
The initial release of CAS will support a purpose-built AI pipeline running an IBM version of NVIDIA NIM and the NVIDIA multimodal PDF extraction blueprint, finely tuned to ensure that AI assistance and agents consistently provide enterprise-grade accuracy.