Open Formats for AI Workloads to Exploit Computational Memory and Storage
San Tomas + Lawrence
Mon Sep 28 | 7:00pm
Abstract
The long-term roadmap for memory fabrics for AI supercomputers points to evolution beyond present-day compute-centric homogeneous scale-up hardware and symmetric memory stacks toward data-centric heterogeneous asymmetric and eventually disaggregated architectures with asymmetric memory stacks that will naturally cause suitable data-intensive operations to gravitate to emerging NDP (near-data processing) devices.
Anticipated benefits of near-data processing include improvements in performance per Watt, latency, fabric goodput, and cost-effective scaling of infrastructure. In order to move beyond prototypes, NDP software stacks need to standardize to a vendor-neutral architecture capable of accelerating a broader range of applications and provide not just improved performance but also portability across different shapes of compute-memory hierarchy encountered during deployment.
After surveying large-scale data-intensive workloads from AI, Database, and HPC communities, we have identified a widely used set of computational memory programming principles and best practices of parallel computing that exploit algebraic structures based on relations, large sparse tensors, and large graphs. Entire AI workloads -- for instance, comprising deep neural networks and LLMs, their relational data wrangling supply lines, and supporting physical world simulations and digital twins -- can be captured beautifully in algebraic expression trees that multilayered compiler frameworks can progressively lower into single or fused operators that modern day accelerator hardware can process at petaops per second speeds.
It is this foundation of three sisters -- relational algebra, linear algebra, and graph processing -- that now needs to form a pivot around which the software stack evolves to tackle growing compute-memory asymmetry and disaggregation and to exploit power-efficient NDP.
In this Birds-of-a-Feather (BoF) session on open formats, we will invite speakers who will approach and debate the path forward in software techniques for programming future open AI hardware ecosystems from several different perspectives.
Related Sessions
Birds of a Feather
DNA is in your Storage Genes!
Join us to discuss the latest developments in DNA data storage, and to explore the potential benefits of DDSA collaborating with other SNIA groups.
Birds of a Feather
Innovating GPU‑Agnostic Storage, Multivendor Interoperability and Test Tools
AI is driving the future of data, and a multivendor interoperability test approach is needed to ensure customer data integrity, any-to any high efficiency data throughput, and vendor-neutral open framework standardization. Come to this BoF to discuss future direction on SNIA’s efforts to pull together all the necessary building blocks, including contributions from Accelerated Object IO Technical Work Group (TWG), Cloud Object Storage Test Tools TWG, as well as StorageAI™ community building and Cloud Storage Technologies community which sponsors Cloud Object Storage Plugfests.
Birds of a Feather
Update from IEEE SISWG (Security in Storage WG) about Storage Sanitization
This topic provides an overview of standards and activities from the IEEE Security in Storage Work Group (from Chair). The topic will cover overview of standards activities such as IEEE STD 1619-2025, IEEE STD 1619.2-2025, IEEE STD 2883.2-2025. The session will cover highlights from the IEEE STD 1619-2025 (AES XTS) with focus on key scope for data encryption and its impact on storage device design. This could impact SSD designs, specifications such as the OCP Datacenter NVMe SSD Specification, and relevant certification processes. Also covered are the highlights from the IEEE 1619.2-2025, which covers the Wide Block Encryption Algorithm. The session will also provide an overview of IEEE STD 2883.2 (being published) which covers the recommended practices for data sanitization in cloud storage. This also covers adjacent standards such as the TCG Key Per I/O and the TCG MPA MEK as applicable to this practice. We can provide an overview of the ongoing work on IEEE STD 2883, which focuses on sanitization at the storage device level.
Birds of a Feather
Data at Scale- Lessons from Rethinking HPC and AI Object Storage with DAOS
Join storage technologists across the industry as we discuss the deployments of Distributed Asynchronous Object Storage (DAOS) in the industry so far, performance and debug learnings, and growth roadmap going forward. DAOS is an open-source software-defined high-performance scalable storage system with successful production deployments and opportunities for more companies to engage in the open-source community. It frequently ranks in the top performance of exa-scale storage workloads alongside Lustre.
Birds of a Feather
Fully Open Source Parallel NFS with PNFS-Lattice Scalable Metadata Service
At long last, there is a path to a fully Open Source Standards based Parallel/Scalable File system, PNFS with PNFS-Lattice. PNFS-Lattice provides a scalable PNFS metadata service that separated the protocol service from the metadata catalog service allowing independent scaling of the services. The solution is in user space on Linux and provides a metadata catalog service using RonDB but enables you to plug in your own metadata catalog service. The metadata catolog service can have many API's for NFS service, metadata extract, change enumperation, etc. PNFS-Lattice also will eventually allow you to implement your own policy for how data is laid out and how state is shared between the parts. A discussion of this effort and solicitation of input is the subject of this BOF.
Birds of a Feather
Open Formats for AI Workloads to Exploit Computational Memory and Storage
The long-term roadmap for memory fabrics for AI supercomputers points to evolution beyond present-day compute-centric homogeneous scale-up hardware and symmetric memory stacks toward data-centric heterogeneous asymmetric and eventually disaggregated architectures with asymmetric memory stacks that will naturally cause suitable data-intensive operations to gravitate to emerging NDP (near-data processing) devices.
Anticipated benefits of near-data processing include improvements in performance per Watt, latency, fabric goodput, and cost-effective scaling of infrastructure. In order to move beyond prototypes, NDP software stacks need to standardize to a vendor-neutral architecture capable of accelerating a broader range of applications and provide not just improved performance but also portability across different shapes of compute-memory hierarchy encountered during deployment.
After surveying large-scale data-intensive workloads from AI, Database, and HPC communities, we have identified a widely used set of computational memory programming principles and best practices of parallel computing that exploit algebraic structures based on relations, large sparse tensors, and large graphs. Entire AI workloads -- for instance, comprising deep neural networks and LLMs, their relational data wrangling supply lines, and supporting physical world simulations and digital twins -- can be captured beautifully in algebraic expression trees that multilayered compiler frameworks can progressively lower into single or fused operators that modern day accelerator hardware can process at petaops per second speeds.
It is this foundation of three sisters -- relational algebra, linear algebra, and graph processing -- that now needs to form a pivot around which the software stack evolves to tackle growing compute-memory asymmetry and disaggregation and to exploit power-efficient NDP.
In this Birds-of-a-Feather (BoF) session on open formats, we will invite speakers who will approach and debate the path forward in software techniques for programming future open AI hardware ecosystems from several different perspectives.