What is this reference architecture for?

This document provides a reference architecture for the NVIDIA Enterprise AI Factory on Supermicro SYS-212GB-FNR systems, detailing the technology stack, sizing guidance, and configuration for Kubernetes-based AI workloads.

Who is the intended audience for this document?

It is intended for IT architects, consultants, and administrators planning, designing, and deploying Kubernetes-based AI solutions using Spectro Cloud PaletteAI.

What are the core infrastructure components covered?

The architecture covers Supermicro SYS-212GB-FNR servers, NVIDIA Spectrum-X Ethernet for RoCE, NVIDIA BlueField-3 networking (SuperNICs and DPUs), Ubuntu 22.04 LTS, Kubernetes 1.33, Cilium, and foundational services for running AI workloads.

How are AI workloads scheduled and managed?

The reference architecture includes NVIDIA Run:ai for AI workload orchestration and GPU utilization management, along with the NVIDIA GPU Operator and supporting components for provisioning and monitoring GPUs.

How is bare metal provisioning handled?

Canonical MAAS is used for automated bare metal OS deployment via PXE, and Spectro Cloud PaletteAI integrates with MAAS to enable automated Kubernetes cluster deployment and lifecycle operations.

What storage approach does the architecture use?

Longhorn is used for basic Kubernetes storage needs (e.g., operator and control plane requirements). For AI workload data, an external storage solution can be added via the chosen vendor’s CSI driver.

NVIDIA Enterprise AI Factory Reference architecture for
Supermicro SYS-212GB-FNR

Get the PDF

NVIDIA Enterprise AI Factory Reference architecture for Supermicro SYS-212GB-FNR

NVIDIA Enterprise AI Factory Reference architecture for
Supermicro SYS-212GB-FNR