Jetson Orin Nano vs Orin NX for Robotics

NVIDIA’s Orin family spans several modules built on one SoC architecture, which is why Orin Nano and Orin NX boards can look nearly identical yet behave very differently under load. The two entry points are the Orin Nano series and the Orin NX series: the same 260-pin SO-DIMM footprint and JetPack software family, different amounts of GPU clock, CPU cores, memory, accelerator silicon, PCIe generation and video hardware.

Every figure below is NVIDIA’s published specification, read from the sources linked at the end of this guide. No independent benchmark results are quoted, and no board was tested for this article. Judgement calls are labelled as assessments.

Verified specifications

Rows come from NVIDIA’s Orin module comparison table unless marked otherwise. Power modes come from the Jetson Linux power and clock tables.

Specification	Orin Nano 4GB	Orin Nano 8GB	Orin NX 8GB	Orin NX 16GB
Overall AI performance	34 sparse INT8 TOPS	67 sparse INT8 TOPS	117 sparse INT8 TOPS	157 sparse INT8 TOPS
GPU tensor core INT8, sparse / dense	34 / 17	67 / 33	77 / 38	77 / 38
GPU	512-core Ampere, 16 tensor cores	1024-core Ampere, 32 tensor cores	1024-core Ampere, 32 tensor cores	1024-core Ampere, 32 tensor cores
GPU max clock	1020 MHz	1020 MHz	1173 MHz	1173 MHz
CPU	6-core A78AE, 1.7 GHz	6-core A78AE, 1.7 GHz	6-core A78AE, 2.0 GHz	8-core A78AE, 2.0 GHz
Deep learning and vision accelerators	none	none	1x NVDLA v2, 1x PVA v2	2x NVDLA v2, 1x PVA v2
Memory	4GB 64-bit LPDDR5, 51 GB/s	8GB 128-bit LPDDR5, 102 GB/s	8GB 128-bit LPDDR5, 102.4 GB/s	16GB 128-bit LPDDR5, 102.4 GB/s
Video	decode 1x 4K60, encode via CPU cores	decode 1x 4K60, encode via CPU cores	decode 1x 8K30, encode 1x 4K60 H.265	same as Orin NX 8GB
PCIe	1x4 + 3x1, Gen3	1x4 + 3x1, Gen3	1x4 + 3x1, Gen4	1x4 + 3x1, Gen4
Documented power modes	10W default, 7W AI, 7W CPU	15W default, 7W, MaxN	MaxN, 10W, 15W default, 25W	MaxN, 10W, 15W default, 25W

Two footnotes matter. The headline TOPS figures decompose in NVIDIA’s own tables: Orin NX 8GB is 77 sparse INT8 TOPS from the GPU tensor cores plus 40 from its single DLA, Orin NX 16GB is 77 plus 80 from two DLAs, and Orin Nano has no DLA, so its figure is GPU only. Also, MaxN on Orin Nano 8GB is available only when the board is flashed with the Super configuration.

What it means for robotics developers

Multi-camera pipelines and I/O. Both series expose 8 lanes of MIPI CSI-2 D-PHY 2.1 (up to 20 Gbps), up to 4 cameras and 8 virtual channels in NVIDIA’s module table, so camera count is set by the carrier board. The Orin Nano Super Developer Kit board provides two 22-pin CSI-2 connectors for 2-lane and 4-lane cameras, and its x4 M.2 Key M slot is PCIe Gen3 where Orin NX offers Gen4. That gap matters for NVMe capture and accelerator cards.

Video. Orin Nano has no hardware encoder: NVIDIA lists encode as 1080p30 handled by 1 to 2 CPU cores, with decode at 1x 4K60 H.265. Orin NX adds H.265 encode to 1x 4K60 and decode to 1x 8K30. Streaming robots pay for that difference in CPU time.

SLAM and VSLAM. Memory bandwidth is nearly identical (102 against 102.4 GB/s), so the differentiators are CPU cores, GPU clock and offload engines. My assessment: the extra CPU cores, the higher GPU ceiling and the DLA on Orin NX 16GB are what let a SLAM front end, a detection network and a control loop run concurrently instead of taking turns.

VLA and transformer inference. Unified memory is shared with the OS, camera buffers and every ROS 2 node, so capacity sets the ceiling. Orin Nano stops at 8GB, Orin NX 16GB doubles it. My assessment: 16GB is the comfortable target for vision-language-action models, and NVIDIA’s sparse INT8 figures are ceilings that quantization decides in practice.

Power and thermals. NVIDIA documents 7W and 15W budgets on Orin Nano, and 10W, 15W and 25W on Orin NX. The power tables show low-power modes reduce online CPU cores and, on Orin NX, disable the PVA cores entirely in the 10W and 15W modes.

Training versus deployment. The published Orin figures are inference figures. Training belongs on a workstation or data centre GPU, and Isaac ROS targets workstations and embedded Jetson from one codebase: train and simulate upstream, export through TensorRT, deploy to the module.

The software side

JetPack is the unified SDK. NVIDIA states that JetPack 7 gives full support for the Jetson Orin and Thor platforms, together with a preemptable real-time kernel, Multi-Instance GPU and an integrated Holoscan Sensor Bridge (JetPack). It bundles the CUDA Toolkit, cuDNN, TensorRT, PyTorch containers, vLLM and SGLang serving paths, Triton Inference Server, DeepStream, OpenCV and VisionWorks libraries, Nsight profiling tools, plus security and over-the-air update support. Both series inherit the same CUDA and TensorRT toolchain, so a model built for one runs on the other.

The ROS 2 layer is Isaac ROS, NVIDIA’s collection of CUDA-accelerated ROS 2 packages and AI models for perception, localization and AI robotics, compatible with existing ROS 2 nodes (Isaac ROS). Its documentation states that all packages are designed and tested with ROS 2 Jazzy, that the supported Jetson combination is JetPack 7.2 on Jetson Orin or Jetson Thor, and that 128GB or more of NVMe storage is expected on Jetson (requirements). NITROS keeps image pipelines on the GPU, which is where the DLA and higher clocks of Orin NX turn into headroom.

Choosing between them, as tradeoffs

• Orin Nano 4GB: one modest network at a time, a few watts of thermal budget, few cameras. The 51 GB/s memory path and absent DLA and PVA make this the tightest option in the family.
• Orin Nano 8GB: the same low power envelope with twice the memory bandwidth and GPU cores, but no hardware video encode.
• Orin NX 8GB: more CPU cores and clock, hardware video encode, PCIe Gen4 and a DLA to offload part of the inference graph, if 8GB of memory is enough.
• Orin NX 16GB: when memory capacity rather than TOPS binds, for larger perception models, VLA models, several concurrent camera streams and both DLA cores.
• Prototype once, deploy either. NVIDIA’s datasheet states the Orin Nano Super Developer Kit carrier board accommodates all Orin Nano and Orin NX modules, so a design can move up to Orin NX once compute or memory runs out, after checking power delivery and thermals.

What comes next

NVIDIA’s Orin page highlights the announced Jetson Orin Nano 2, described as nearly twice the inference performance of its predecessor in the same form factor and a 40W power envelope using 40% less power, and it references new Blackwell-powered T3000 and T2000 modules. Above Orin, the module lineup lists Jetson AGX Thor (T5000 and T4000) at up to 2070 FP4 TFLOPS and 128GB of memory across 40W to 130W, claimed at over 7.5x the AI compute of AGX Orin (Jetson modules). Both series keep the same JetPack and Isaac ROS migration path.

FAQ

Can I run the same software on a Jetson Orin Nano and a Jetson Orin NX? Yes. NVIDIA states the Orin family modules share one SoC architecture and one software stack, and the Orin Nano Super Developer Kit datasheet says its carrier board accommodates all Orin Nano and Orin NX modules.

Do Orin Nano and Orin NX use the same module connector? Yes. NVIDIA’s Orin module comparison table lists both series in the same 69.6 mm x 45 mm, 260-pin SO-DIMM form factor. Check power delivery and thermal design before substituting modules.

Does the Orin Nano module have a hardware video encoder? No. NVIDIA lists video encode for the Orin Nano 8GB module as 1080p30 supported by 1 to 2 CPU cores, while Orin NX lists hardware encode up to 1x 4K60 H.265.

Why do Orin Nano and Orin NX headline TOPS numbers differ when the GPUs look identical? NVIDIA’s overall AI performance figure combines GPU tensor core throughput and, where present, DLA throughput. Orin NX 8GB is published as 77 sparse INT8 TOPS from the tensor cores plus 40 from one DLA; Orin Nano has no DLA.

Can I train models on Orin Nano or Orin NX? NVIDIA publishes inference figures for both series, with DLA acceleration on Orin NX. Training belongs on a workstation or data centre GPU, and Isaac ROS deploys on workstations and embedded Jetson from one codebase.

Sources

All figures were read from these pages, fetched on 14 September 2026:

• NVIDIA Jetson Orin product page
• NVIDIA Jetson Modules
• Jetson Orin Nano Super Developer Kit
• Jetson Orin Nano Super Developer Kit datasheet, PDF
• Jetson Linux Developer Guide, Orin power and clock configuration
• NVIDIA JetPack SDK
• NVIDIA Isaac ROS
• Isaac ROS documentation, Getting Started