Skip to content
SDB

Cloud Technology

NVIDIA IoT & AI Reference Architecture

GPU-accelerated edge and cloud AI — Jetson and IGX at the plant, TensorRT and NIM for inference, Omniverse for digital twins, and DGX Cloud for training.

NVIDIA AI Cloud & Edge · 5 architecture layers

End-to-end data flow

01

Sensors & Cameras

Video / LiDAR / IMU

02

Jetson Edge AI

TensorRT inference

03

DeepStream Pipeline

Multi-stream video

04

Isaac ROS / Apps

ROS 2 navigation

05

Triton / NIM

Model serving

06

Fleet Command / Omniverse

Ops + digital twin

Architecture layers

L1

Edge AI Hardware Layer

Jetson AGX Orin

Up to 275 TOPS for multi-DNN robotics, AMRs and industrial inspection

Jetson Orin NX / Nano

Compact modules for cost-sensitive vision and sensor-fusion deployments

NVIDIA IGX Orin

Industrial edge AI with functional-safety pathways for certified systems

Jetson Thor (next-gen)

Higher-performance humanoid and physical-AI class edge compute for advanced autonomy

L2

Edge Software & Runtime Layer

NVIDIA DeepStream

GStreamer analytics pipelines for multi-camera industrial video

Isaac ROS

GPU-accelerated ROS 2 packages for depth, detection, pose and navigation

NVIDIA TAO Toolkit

Transfer learning to fine-tune NVIDIA vision models on custom plant data

JetPack SDK

Linux, CUDA, cuDNN, TensorRT and VPI — complete Jetson software stack

L3

Inference & Optimization Layer

TensorRT / TensorRT-LLM

Optimised inference with quantization and kernel fusion for vision and LLMs

Triton Inference Server

Multi-framework serving with dynamic batching and concurrent model execution

NVIDIA NIM

Containerised inference microservices for LLMs, vision, speech and domain models

NIM Agent Blueprints

Reference agent workflows combining NIM models, retrieval and tool use

L4

Training & Foundation Model Layer

NVIDIA DGX / DGX Cloud

Validated multi-GPU systems and cloud capacity for large-scale training

NVIDIA NeMo

LLM and speech frameworks — pretrain, LoRA fine-tune, RLHF and evaluation

NVIDIA Cosmos

World foundation models for physical AI simulation and synthetic data generation

NVIDIA BioNeMo

Generative AI for drug discovery and computational biology

L5

Platform & Simulation Layer

NVIDIA Omniverse

USD-based collaboration and simulation for factory and warehouse digital twins

NVIDIA Isaac Sim

Photoreal robotics simulation — synthetic data, nav and manipulation testing

NVIDIA Fleet Command

Deploy, manage and monitor NVIDIA-certified edge AI systems at scale

NGC Catalog

GPU-optimised containers, models, SDKs and Helm charts

Industrial AI use cases

Autonomous Mobile Robots

Isaac ROS + Jetson Orin + DeepStream

Real-Time Visual Inspection

DeepStream + TensorRT + TAO

Factory Digital Twin

Omniverse + Isaac Sim + Replicator

Edge LLM Inference

NIM + TensorRT-LLM + Jetson / IGX

Predictive Maintenance

Triton + NeMo + time-series models

Industrial Fleet AI

Fleet Command + AGX Orin + DeepStream

Key differentiators

Edge-to-cloud GPU continuum

Same CUDA software path from Jetson on the plant floor to DGX Cloud in the data centre.

Production inference stack

TensorRT, Triton and NIM turn research models into managed, portable microservices.

Physical AI & twins

Omniverse, Isaac Sim and Cosmos close the loop between simulation and deployed robots.

Robotics-native tooling

Isaac ROS and DeepStream are purpose-built for AMRs, vision and multi-sensor autonomy.

Jetson platform comparison

ModelAI perf.PowerTypical use
Jetson Nano~21 TOPS5–10WEntry vision prototyping
Jetson Orin NX 8GB~70 TOPS10–20WIndustrial inspection, compact robots
Jetson Orin NX 16GB~100 TOPS10–25WMulti-sensor fusion
Jetson AGX Orin 32GB~200 TOPS15–40WAMRs, multi-camera edge servers
Jetson AGX Orin 64GB~275 TOPS15–60WFull autonomy, multi-DNN inference