Resource Library

This IBM Redbook presents a validated reference architecture for an AI infrastructure that allows KV cache to be persistently stored, shared, and reused across requests, sessions, and GPU nodes. It consists of NVIDIA Dynamo for intelligent distributed KV cache management, IBM® Storage Scale Erasure Coding Edition (ECE) as the high-performance shared storage tier, Supermicro Petascale servers as the storage and networking foundation, and NVIDIA Spectrum-X Ethernet to tie it all together with the low-latency, high-bandwidth fabric that production AI inference demands.

This document represents the reference architecture for the NVIDIA Enterprise AI Factory for Supermicro SYS-212GB-FNR systems. It is a comprehensive explanation of the technology used, each component’s purpose and configuration.

As AI/ML, Agentic AI, and Retrieval-Augmented Generation (RAG) workloads grow in complexity and demand, the performance of underlying storage systems becomes mission-critical. To keep GPUs fully utilized, storage must deliver exceptional performance in both sequential and random operations. With their excellent performance, NVMe PCIe drives are ideal for mission-critical applications.