Jetson Orin Nano · Orin NX · AGX Orin · Jetson Thor

NVIDIA Jetson Platforms for Edge AI and Robotics

Compare Jetson platforms by AI metric, memory, sensor bandwidth, JetPack generation, carrier-board requirements, power mode, and sustained deployment limits.

  • Orin Nano Super · 67 TOPS8 GB · up to 25 W
  • Orin NX Super · 157 TOPS16 GB · up to 40 W
  • AGX Orin · 275 TOPS64 GB · up to 60 W
Compare Jetson Modules Review My NVIDIA Requirements
NVIDIA Jetson carrier platform for edge AI robotics and machine vision

Jetson Module Fit

Choose the Module by the Bottleneck in the AI Pipeline

Jetson selection should start from model memory, sensor load, runtime and sustained power—not from the highest published compute number.

Jetson Direction Reference Strongest Fit Move Up When
Orin Nano 8GB Super 67 sparse / 33 dense INT8 TOPS · 8 GB · 102 GB/s Compact vision, robotics and lighter multimodal inference Memory, concurrent pipelines or 25 W compute ceiling becomes limiting
Orin NX 16GB Super 157 sparse / 78 dense INT8 TOPS · 16 GB · 102 GB/s More concurrent inference in the 70 × 45 mm module class You need much more memory, sensor I/O or AGX-class expansion
AGX Orin 64GB 275 INT8 TOPS · 64 GB · 204.8 GB/s Large models, multi-sensor robotics and high-compute edge systems 64 GB or Orin-class compute is still insufficient
Jetson Thor / T5000 2070 FP4 TFLOPS sparse · 128 GB · 273 GB/s Physical AI, humanoid robotics and larger generative models 40–130 W power or new platform mechanics do not fit the product

INT8 · FP4 · Sparse · Dense

Do Not Rank Orin and Thor with One AI Number

Orin and Thor use different published precision and sparsity references. A larger number can describe a different workload format rather than a directly comparable increase in application performance.

Metric Reference Example What the Visitor Should Compare Instead
Sparse INT8 TOPS Orin Nano Super: 67 · Orin NX Super: 157 Target model, precision, sparsity support, batch size and sustained clocks
Dense INT8 TOPS Orin Nano Super: 33 · Orin NX Super: 78 Use when the model does not benefit from sparse acceleration
FP4 TFLOPS Jetson T5000: 2070 sparse FP4 TFLOPS Do not compare directly with Orin INT8 TOPS; benchmark the target model and runtime

Memory Residency

Model Size and Working Memory Can Force the Upgrade Before Compute Does

Jetson memory is shared by the operating system, GPU workloads, model weights, frame buffers, containers and application data. The usable headroom is therefore smaller than the headline module capacity.

Module Memory Bandwidth Typical Upgrade Trigger
Orin Nano Super 8 GB LPDDR5 102 GB/s Model weights, camera buffers and runtime no longer fit with safe headroom
Orin NX Super 16 GB LPDDR5 102 GB/s Concurrent models, larger VLMs or sensor pipelines exceed 16 GB
AGX Orin 64GB 64 GB LPDDR5 204.8 GB/s Larger generative models or multi-sensor workloads need more residency and bandwidth
Jetson T5000 128 GB LPDDR5X 273 GB/s Physical-AI workloads require substantially larger model and context memory

Compare AI-Ready Platform Requirements Beyond Jetson →

Jetson Module · Carrier Contract

The Module Defines Resources; the Carrier Defines What the System Can Connect

A Jetson module does not provide a finished deployment interface set. Carrier routing determines camera inputs, NVMe, networking, USB, CAN, expansion, power and external connectors.

Camera Path

Define CSI, GMSL bridge, USB or GigE cameras, synchronization, trigger, sensor power and driver support before carrier selection.

Storage and PCIe

Map NVMe, capture, modem and network devices to the module’s available lanes and confirm shared-resource limits.

Networking

Size GbE, 10GbE or 25GbE-class links from sensor input, logging and inference-output traffic rather than connector count alone.

Mechanical Compatibility

Orin Nano/NX use the compact 70 × 45 mm module class, while AGX and Thor require different carrier and thermal architectures.

Plan High-Speed Network I/O for AI Data Paths →

JetPack · CUDA · TensorRT

Freeze the NVIDIA Software Stack with the Hardware Revision

JetPack is part of the production platform because it binds Jetson Linux, CUDA, TensorRT, drivers, device tree and supported module generation.

Layer Freeze for Production Why It Matters
Jetson Linux / BSP Release branch, kernel, device tree, drivers Controls camera, PCIe, USB, networking and carrier compatibility
CUDA / TensorRT Runtime versions, operators, engine build process Can change model compatibility, precision path and latency
Containers Base image and application dependencies Reduces unexpected runtime changes between production revisions
OTA / Recovery Update package, rollback and recovery image Prevents an update from making deployed units unrecoverable

Super Mode · nvpmodel · Sustained Power

Size the Carrier and Cooling for the Power Mode You Will Actually Deploy

Module Power Reference Design Meaning
Orin Nano Super 7 / 15 / 25 W reference modes 25 W operation requires carrier power and cooling designed for Super-class sustained load
Orin NX Super 10 / 15 / 25 / 40 W reference modes 40 W mode changes thermal and DC power requirements even though the module remains 70 × 45 mm
AGX Orin 64GB 15–60 W Carrier, heat sink and enclosure must match the selected sustained mode
Jetson T5000 40–130 W Moves the product into a materially different power and cooling class

Jetson Migration

Upgrade Only When the Existing Module Hits a Measurable Limit

Current Module Move When Next Direction
Orin Nano Super 8 GB memory or 67 sparse TOPS is no longer enough Orin NX Super
Orin NX Super 16 GB memory, 157 TOPS or compact carrier I/O becomes limiting AGX Orin
AGX Orin 64 GB memory, 275 TOPS or Orin-generation software is insufficient Jetson Thor
Jetson Thor 130 W-class power, module mechanics or JetPack 7 migration does not fit Re-evaluate platform architecture

NVIDIA Platform Review

Send the Model and Sensor Pipeline First

Provide model, precision, JetPack target, camera/sensor topology, memory need, storage, networking, power mode, enclosure, quantity and lifecycle.

Submit NVIDIA Platform Requirements

Starting Platforms

Start from the Carrier Direction Closest to the Required Jetson Module

Starting Platform Best Reason to Start Here Freeze Before Release
NVIDIA Jetson Mini-ITX Carrier Board Standard Jetson edge-AI carrier direction Module, CSI/USB/network map, NVMe, power, JetPack, thermal solution and hardware revision
Custom Jetson Reference Platform Standard carrier I/O, sensor topology or mechanics do not fit Resource map, carrier BOM, BSP, validation scope, thermal path and lifecycle controls

FAQ

NVIDIA Jetson Platform FAQ

Does higher Jetson TOPS always mean higher FPS?

No. Precision, sparsity, TensorRT support, model operators, memory traffic, preprocessing, batch size and sustained power determine real throughput.

When should I choose Orin NX instead of Orin Nano?

Move when 8 GB memory, 67 sparse TOPS or Nano-class concurrency is limiting, while the compact 70 × 45 mm module format still fits.

Can Thor be compared directly with Orin TOPS?

No. Thor’s published FP4 TFLOPS and Orin’s INT8 TOPS use different precision metrics. Compare the actual model, memory need, latency and power.

What must be frozen for Jetson production?

Freeze the module, carrier revision, JetPack/Jetson Linux release, device tree, drivers, CUDA/TensorRT versions, containers, power mode, thermal design and recovery process.

Explore Practical Insights for NVIDIA Edge & Embedded Design

Stay ahead with in-depth coverage on Jetson module integration, real-time GPU compute, and field-tested AI deployment tactics—spanning robotics, computer vision, and autonomous infrastructure. Whether you’re building the next smart factory or deploying AI at the edge, our blog delivers the knowledge to do it right.