In this comprehensive technical session, SNIA Board Chair J Metz, Ph.D., presents how the StorageAI Community is tackling the fundamental data-movement, benchmarking, physical design, and sustainability challenges unique to AI data requirements. From SDXI zero-copy semantics and Accelerator-Initiated I/O, to Object-over-RDMA direct-to-GPU paths, standardized Direct Liquid Cooling for EDSFF SSDs with PCIe 7.0 readiness (including detailed topology and end-to-end link budgets), new AI-specific workload definitions that expose the limitations of traditional benchmarks, computational storage offloads for KV cache and similarity search, and emerging storage carbon intensity metrics—discover the interoperable open specifications that will define the next generation of high-performance, efficient, and sustainable AI data services. Ideal for architects, performance engineers, and infrastructure leads building large-scale generative AI systems.
The StorageAI initiative by SNIA is designed to provide coherent, workload-specific solutions by addressing critical bottlenecks in AI data services. A major focus is on optimizing data movement, moving beyond traditional multi-step buffer copies to more efficient "processor bypass" methods, such as zero-copy semantics via the Smart Data Accelerator Interface (SDXI). This enables direct-to-GPU paths for object storage over RDMA, reducing latency and allowing GPUs to initiate I/O, which is crucial for inference workloads such as KV caches. Furthermore, SNIA is addressing the inadequacy of conventional benchmarks for dynamic AI workloads by defining "real" AI-specific workloads. These definitions aim to enable reproducible, standardized comparisons, ensuring that performance metrics accurately reflect the efficiency of AI systems, regardless of rapidly changing hardware.
Addressing physical design and sustainability, the StorageAI community is developing open standards for direct liquid cooling of EDSFF SSDs, recognizing that high-density AI processing generates significant heat, rendering traditional air cooling insufficient. These liquid-cooled SSDs act as active thermal bridges, dissipating heat from GPUs and contributing to overall server cooling, thereby improving energy efficiency. This holistic approach extends to defining carbon-intensity metrics for storage in AI environments, with collaborations underway with organizations such as OCP to standardize these definitions. Ultimately, StorageAI aims to eliminate proprietary silos, integrate components from physical layers to software, and provide a long-term, vendor-neutral framework that enables integrators and developers to build efficient, scalable, and sustainable AI systems, including critical aspects such as checkpointing for large clusters and overall data governance.

Presented by Dr J Metz, Chair at SNIA. Recorded live at AI Infrastructure Field Day in Millbrae, California, on June 10th, 2026. Watch the entire presentation at https://techfieldday.com/appearance/snia-presents-at-ai-infrastructure-field-day-5/ or visit https://techfieldday.com/event/aiifd5/ or https://www.snia.org/ for more information.