SNIA dives into Storage.AI standards project

The Storage Networking Industry Association (SNIA) is working on an open standards project for efficient data services related to AI workloads, focusing on industry-standard, non-proprietary, and neutral approaches.

The SNIA is a not-for-profit global organization made up of corporations, universities, startups, and individuals who collaborate to develop and promote vendor-neutral architectures as well as international standards and specifications. It looks at technologies related to the storage, transport, optimization of infrastructure, acceleration, format, and protection of data. Prominent standards developed by the SNIA include SMI-S (Storage Management Initiative Specification), NVMe Management Interface, Swordfish scalable storage management API, and CDMI (Cloud Data Management Interface).

Dr. J Metz, SNIA Chair, stated: “The unprecedented demands of AI require a holistic view of the data pipeline, from storage and memory to networking and processing. No single company can solve these challenges alone. SNIA’s Storage.AI provides the essential, vendor-neutral framework for the industry to coordinate a wide range of data services, building the efficient, non-proprietary solutions needed to accelerate AI for everyone.”

The Storage.AI project will work to build broad ecosystem support with SNIA’s partners, including UEC, NVM Express, OCP, OFA, DMTF, and SPEC. The SNIA says AI workloads are extraordinarily complex and constrained by issues related to latency, space, power and cooling, memory, and cost. It asserts that addressing these problems through an open industry initiative is the fastest path to optimization and adoption.

The initial industry players who have signed on to Storage.AI include AMD, Cisco, DDN, Dell, IBM, Intel, KIOXIA, Microchip, Micron, NetApp, Pure Storage, Samsung, Seagate, Solidigm, and WEKA.

A fundamental problem is that storage is not connected to GPUs and other accelerators:

Storage.AI has identified six technology areas that can help fix this

  • AiSIO – Accelerator-Initiated Storage IO
  • CNM – Compute-Near-Memory (E.g.computational storage)
  • FDO – Flexible Data Placement
  • GDB – GPU Direct Bypass (File/object over RDMA)
  • NVMP – NVM Programming Model
  • SDXI – Smart Data Accelerator Interface

A diagram shows their inter-relationships and placement;

They should prompt the formation of new technical workgroups (TWGs) as well as fit into existing SNIA workgroups. 

There are two competing GPU suppliers signed up: AMD and Intel. But there is one obvious big AI industry beast which is not present in the project, GPU hardware and AI software supplier Nvidia.

One of the main areas concerns the delivery of data to GPUs and here Nvidia, with its massive share of the GPU market, dominates with its proprietary GPU Direct for files and objects, enabling direct storage drive RDMA access and bypassing the drive host storage server CPU and memory which are normally involved in data access by an AI processing accelerator across a network. AWS has also recently announced its S3 Express One Zone tier to accelerate small object data access times.

Using GPU Direct protocols locks companies in to Nvidia. There appears little prospect of Nvidia making its GPU Direct protocols openly available. Can Storage.AI convince Nvidia to support it? In the past, Linux succeeded against proprietary Unix versions, NVMe took over from private SSD access protocols and SATA/SAS, and SCSI now dominates over replaced proprietary disk interfaces like Enhanced Small Disk Interface (ESDI) and Intelligent Peripheral Interface (IPI).

AWS’ S3 has not been replaced and had become a de facto standard because of AWS’ market dominance. 

In general open, vendor-neutral standards encourage technology development, lower costs, provide freedom from supplier lock-in, data portability, and other good things. Industry-wide, neutral standards should mean the AI accelerator market grows faster than it would were it to continue with a proprietary standard.

Companies and industry organizations interested in participating in the Storage.AI project should contact storage.ai@snia.org.