The transformational launch of GPT-4 has accelerated the race to build AI data centers for large-scale training and inference. While GPUs and high-bandwidth memory are well-known critical components, the essential role of storage devices in AI infrastructure is often overlooked. This presentation will explore the AI processing pipeline within data centers, emphasizing the crucial role of storage devices in both compute and storage nodes. We will examine the characteristics of AI workloads to derive specific requirements for storage devices and controllers
 

This Birds of a Feather (BoF) session brings together the community to discuss, share experiences, and brainstorm the increasing role of object storage for AI workloads. As AI applications continue to grow, object storage is being utilized as a primary storage solution for many of these applications.

This presentation will show how AI and storage interconnects can support 400Gbps lane speeds. Feasibility of copper cabling across intra and inter rack interconnects will be supported using Active Copper Cables (ACC). Channel models, including co-packaged copper, flyover cables, backplane connectors, backplane cables and front panel pluggable modules will be shown and analyzed for 400G performance. 

This presentation is aimed at providing insights into the transition from UEC Specification to Productization. We will discuss how customers are perceiving the shift from RoCEv2 to UET format, and outline the criteria for moving from UEC concepts to requirements. Additionally, we will compare NIC-specific features with DC switch-specific features. The presentation will also highlight the value creation of UEC from the perspective of a networking vendor. 
 

The Grand Unified File Indexer (GUFI) is a state of the art file/storage system indexing too that offers both user and storage administrator access in a way that each user can only see the metadata for the files they have access to.  Imagine having an exabyte of data that is in many file system trees in a trillion files in10 billion directories that uses POSIX permissions (UID/GID/rwxrwxrwx) with inheritance.

Data enhances foundational LLMs (e.g. GPT-4, Mistral Large and Llama 2) for context-aware outputs. In this session, we'll cover using unstructured, multi-modal data (e.g. PDFs, images or videos) in retrieval augmented generation (RAG) systems and learn about how cloud object storage can be an ideal file system for LLM-based applications that transform and use of domain-specific data, store user context and much more.

Ultra Ethernet (UEC) is the hot new technology that supports today’s AI and HPC workloads. For AI, it is well understood that the expensive accelerators (GPUs) must be constantly supplied with data. In many cases data continues to reside on traditional storage networks which are not designed for the performance and scalability required for today’s HPC and AI workloads. These networks are expected to migrate to UEC for improved bandwidth, tail latency, and scale.