Jetson Orin Nano and Orin NX

The Orin Nano and Orin NX are the cost-conscious half of the Orin generation. Same Ampere GPU and Cortex-A78AE CPU as AGX Orin, but in a smaller SODIMM-style package with lower memory, lower power envelopes, and a hundreds-of-dollars price point that makes Jetson approachable. Orin Nano targets 5 to 15 W; Orin NX targets 10 to 25 W. Both share the same connector and carrier reference, scaling across four variants (Nano 4 GB, Nano 8 GB, NX 8 GB, NX 16 GB).

This page covers the silicon, the power modes, real workloads, and the comparison against the obvious cost-sensitive alternative: Raspberry Pi 5 plus Hailo-8.

Jetson Orin Nano and Orin NX at a glance

Vendor: NVIDIA Corporation
Released: Orin Nano: March 2023; Orin NX: October 2022
Module form factor: SODIMM 260-pin, 70 x 45 mm, pin-compatible across Nano and NX
CPU: Orin Nano: 6x Cortex-A78AE at 1.5 GHz; Orin NX: 6 or 8x Cortex-A78AE at 2.0 GHz
GPU: NVIDIA Ampere, 1024 CUDA cores, 32 tensor cores (both Nano and NX)
DLA: Orin Nano: 0; Orin NX: 2x NVDLA v2
INT8 TOPS: Orin Nano: 20 to 40; Orin NX: 70 to 100
Memory: LPDDR5; Nano: 4 or 8 GB; NX: 8 or 16 GB
Software: JetPack 6 (L4T r36, Ubuntu 22.04 LTS, CUDA 12.x, TensorRT 10.x)
External links: developer.nvidia.com/embedded/jetson-orin-nano, developer.nvidia.com/embedded/jetson-orin-nx

The silicon

Orin Nano and Orin NX share the same Tegra T234 die as AGX Orin, with portions disabled (fewer CPU cores, smaller GPU cluster, DLAs fused off on Nano). NVIDIA's peak TOPS is the sum of GPU INT8 peak and DLA INT8 peak; real workloads rarely sustain the published peak.

The CPU is six or eight Cortex-A78AE cores in a single cluster. Orin Nano 4 GB and 8 GB ship six cores at 1.5 GHz; NX 8 GB unlocks 2.0 GHz; NX 16 GB adds an eight-core configuration. The cores carry automotive-enhanced features (lock-step, ECC, fault-detection) but these are not typically enabled outside automotive deployments.

The GPU is 1024 CUDA cores of Ampere across 8 streaming multiprocessors, 32 third-generation tensor cores supporting INT8, INT4, FP16, BF16. Same architecture as RTX 30-series and A100, scaled down. CUDA kernels written for desktop run unmodified.

Orin NX adds two NVDLA v2 deep-learning accelerators — fixed-function INT8 engines for standard convolutional, pooling, and dense layers. TensorRT targets them via --useDLACore=0 or --useDLACore=1. A typical multi-stream pipeline runs the primary detector on the GPU and per-camera secondary inference on each DLA. Orin Nano omits the DLAs.

Power modes — nvpmodel

NVIDIA configures the operating-point envelope through nvpmodel. Each mode is a tuple of CPU core count, CPU clock cap, GPU clock cap, DLA clock cap, and memory clock cap. Modes are defined per-module in /etc/nvpmodel.conf. Switching modes takes a few seconds.

Module	Mode	CPU cores	CPU max	GPU max	Memory	Power cap
Orin Nano 4GB	0	6	1.5 GHz	625 MHz	2133 MHz	7 W
Orin Nano 4GB	1	6	1.5 GHz	625 MHz	2133 MHz	10 W
Orin Nano 8GB	0	6	1.5 GHz	625 MHz	3200 MHz	7 W
Orin Nano 8GB	1	6	1.5 GHz	625 MHz	3200 MHz	15 W
Orin NX 8GB	0	6	2.0 GHz	765 MHz	3200 MHz	10 W
Orin NX 8GB	1	6	2.0 GHz	765 MHz	3200 MHz	15 W
Orin NX 8GB	2	6	2.0 GHz	765 MHz	3200 MHz	20 W
Orin NX 16GB	0	8	2.0 GHz	918 MHz	3200 MHz	10 W
Orin NX 16GB	1	8	2.0 GHz	918 MHz	3200 MHz	15 W
Orin NX 16GB	2	8	2.0 GHz	918 MHz	3200 MHz	25 W

# Query the current mode
sudo nvpmodel -q
 
# Switch the Orin NX 16 GB to the 25 W maximum-performance mode
sudo nvpmodel -m 2
 
# Lock all clocks to the maximum allowed by the current mode
sudo jetson_clocks
 
# Inspect the resulting power consumption and clocks live
sudo tegrastats --interval 500

jetson_clocks is the second half: nvpmodel sets the cap; jetson_clocks instructs the governor to run at that cap rather than scaling dynamically. Benchmarking runs both: switch mode, lock clocks, capture results, unlock.

Power numbers above are SoM power. A carrier with cooling, cameras, gigabit Ethernet, and NVMe adds 3 to 8 watts.

Real-world workloads

What the Orin Nano and NX actually run, with rough numbers from published JetPack 6 benchmarks and independent measurements:

Workload	Orin Nano 8 GB	Orin NX 16 GB	Notes
YOLOv8n object detection at 1080p, FP16	60 fps	130 fps	Single class, standard input size
YOLOv8m object detection at 1080p, INT8	30 fps	90 fps	Useful for product-class accuracy
Whisper-base speech recognition (real time)	yes	yes	1x real-time on Nano, 3x on NX
ResNet-50 image classification, INT8	250 inf/s	700 inf/s	TensorRT, batch 1
BERT-base NLI inference, INT8	50 inf/s	180 inf/s	Batch 1, sequence length 128
Stable Diffusion 1.5 (512 x 512, 20 steps)	n/a	12 s/image	Tight memory on NX 16 GB
VSLAM (ORB-SLAM3, 30 fps stereo)	yes	yes	CPU + GPU combined
4-camera batched inference (1080p, YOLOv8n)	n/a	30 fps each	Uses 1 DLA + GPU concurrently

The standout workloads are YOLO-class detection at real-time frame rates and end-to-end Whisper transcription. Both fit comfortably on Orin Nano 8 GB, the most common deployment.

Comparison with Raspberry Pi 5 plus Hailo-8

The most direct competitor in 2025 is a Raspberry Pi 5 (Cortex-A76 quad-core at 2.4 GHz, 4 to 8 GB LPDDR4X) paired with a Hailo-8 or Hailo-8L NPU via the PCIe M.2 module. The combination targets the same price band as Orin Nano 8 GB. The comparison favours Jetson on developer experience and Pi-plus-Hailo on raw compute per dollar.

Trait	Orin Nano 8 GB	Raspberry Pi 5 + Hailo-8
CPU	6x Cortex-A78AE at 1.5 GHz	4x Cortex-A76 at 2.4 GHz
GPU	1024 CUDA cores Ampere	VideoCore VII (no general compute)
Dedicated NPU TOPS	40 (GPU INT8)	26 (Hailo-8) or 13 (Hailo-8L)
Memory	8 GB LPDDR5	8 GB LPDDR4X
All-in module / board price	$299 dev kit	$80 Pi 5 + $250 Hailo M.2
Software stack	JetPack, CUDA, TensorRT	Raspberry Pi OS + Hailo SDK
Model deployment	TensorRT from any framework	Hailo Dataflow Compiler
Operator coverage	Comprehensive	Limited; verify on each model
Camera path	NVIDIA libargus, GStreamer	libcamera, V4L2
Ecosystem maturity	Production-class	Improving rapidly

For a developer with a CUDA-trained model who wants to ship it, Orin Nano is the path of less resistance. For an application from a known model Hailo's compiler supports, Pi-plus-Hailo is cheaper. The choice comes down to whether the team can absorb the porting cost.

Carrier boards and developer kits

NVIDIA's Orin Nano developer kit ($249 retail for the 8 GB variant) includes a carrier with HDMI, two MIPI-CSI slots, M.2 for NVMe and Wi-Fi, gigabit Ethernet, USB 3.2. The same carrier accepts an Orin NX module, though NVIDIA does not sell a bundled Orin NX kit — NX ships as a bare SoM.

Third-party carriers include Connect Tech (rugged Photon and Boson lines), Auvidea (compact drone-focused JNX30/JN30), Forecr (cost-focused). Production usually uses a third-party carrier or a custom in-house design. The carrier design guide in the Jetson developer documentation specifies pinout and trace constraints.

Cross-references

NVIDIA Jetson landing page — full family overview.
Jetson AGX Orin and Thor — the higher-end siblings.
Cortex-A78AE — the automotive-enhanced application core.
Edge AI system design — broader architectural context.
Raspberry Pi 5 — the competing low-cost platform.
NVIDIA Orin Nano developer site: developer.nvidia.com/embedded/jetson-orin-nano.
JetPack 6 documentation: docs.nvidia.com/jetson/jetpack.
Carrier board design guide: published in the Jetson Orin Nano design files at developer.nvidia.com/embedded/downloads.