Storage.AI

Solving AI Data Challenges:

Storage.AI is an open standards project for efficient data services related to AI workloads. We are addressing the most difficult challenges related to AI workloads, closing existing gaps in processing and accessing data.

AI exposed urgent problems in data services, especially around how storage interacts with compute.
Data pipelines are inefficient— round-tripping between storage and compute is wasting power and performance.
GPUs go idle when data isn’t available at the right place and time—and that’s incredibly expensive

Storage.AI will focus on vendor-neutral, interoperable solutions for issues like memory tiering, data movement, storage efficiency, and system-level latency that take advantage of GPUs and AI acceleration.

Ongoing SNIA Technical Work	Enables
SDXI (Smart Data Acceleration Interface)	Vendor-neutral DMA acceleration across CPU, GPU, DPU; memory-to-memory copies and transforms
Computational Storage API & Architecture	Allows SSDs and storage devices to perform computation (e.g. filtering, inferencing, format conversion) directly
NVM Programming Model	Unified software interface for accessing NVMe, SCM, and RDMA memory/storage tiers
Swordfish / Redfish Extensions	Unified, Redfish-compatible storage management for hybrid and disaggregated infrastructure
Object Drive Workgroup	Standard interfaces for object storage devices to operate in RDMA / hyperscale environments
Flexible Data Placement APIs	Works with computational storage and GPU-aware storage pipelines to optimize layout and streaming throughput
Future Work Planned
File/Object over RDMA	Enables hybrid file/object storage backends, allowing traditional applications to benefit from S3-like scalability, combined with the speed and efficiency of RDMA
GPU Direct Access Bypass	Bypasses CPU for data movement between GPU memory and RDMA/NVMe, enabling near-zero-latency I/O to accelerators
GPU-Initiated I/O	GPUs can initiate and complete their own I/O operations (e.g. checkpoint writes, data loading), reducing CPU bottlenecks. Also known as Accelerator-Initiated Storage I/O (AiSIO).