Jetson Orin AGX vs Orin NX vs Orin Nano: which to pick

Choosing between Jetson AGX Orin, Orin NX, and Orin Nano comes down to three things: how much inference throughput your application actually needs, how many I/O interfaces your hardware design requires, and what power budget you’re working within. The modules share the same Ampere GPU architecture and run the same JetPack 6 software stack — the differences are in how much of that hardware is exposed and at what power draw.

Key Insights

• All three run JetPack 6 and the same software stack — switching between them is a hardware and DTS change, not a software port
• Orin NX and Orin Nano are pin-compatible on the same carrier board; AGX Orin uses a different, larger form factor
• The DLA is present on AGX Orin and Orin NX but absent on Orin Nano — this is the most important functional difference for inference workloads
• Memory bandwidth drops from 204 GB/s (AGX) to 68 GB/s (Nano) — a bigger real-world bottleneck than TOPS
• PCIe lane count limits peripheral expansion: AGX Orin can drive NVMe + Ethernet + PCIe capture simultaneously; Orin NX and Nano cannot

Full specification comparison

	AGX Orin 64GB	AGX Orin 32GB	Orin NX 16GB	Orin NX 8GB	Orin Nano 8GB	Orin Nano 4GB
CPU	12-core A78AE	12-core A78AE	8-core A78AE	6-core A78AE	6-core A78AE	4-core A78AE
GPU CUDA cores	2048	2048	1024	1024	1024	512
DLA cores	2	2	2	2	None	None
AI performance	275 TOPS	275 TOPS	100 TOPS	70 TOPS	40 TOPS	20 TOPS
Memory	64GB LPDDR5	32GB LPDDR5	16GB LPDDR5	8GB LPDDR5	8GB LPDDR5	4GB LPDDR5
Memory BW	204.8 GB/s	204.8 GB/s	102.4 GB/s	68.3 GB/s	68.3 GB/s	34.1 GB/s
Storage	64GB eMMC	64GB eMMC	16GB eMMC	16GB eMMC	16GB eMMC	16GB eMMC
PCIe	4.0 ×16	4.0 ×16	4.0 ×4	4.0 ×4	3.0 ×4	3.0 ×4
MIPI CSI cameras	Up to 16	Up to 16	Up to 4	Up to 4	Up to 4	Up to 4
USB 3.2 Gen 2	3×	3×	2×	2×	2×	2×
GbE	10GbE	10GbE	1GbE	1GbE	1GbE	1GbE
CAN FD	2×	2×	2×	2×	2×	1×
Power (max TDP)	60W	60W	25W	25W	15W	15W
Module price (approx)	~$899	~$499	~$299	~$199	~$149	~$99
Form factor	AGX (87×100mm)	AGX (87×100mm)	Module (69.6×45mm)	Module (69.6×45mm)	Module (69.6×45mm)	Module (69.6×45mm)

Prices are NVIDIA suggested retail for production quantities. Devkit pricing is higher.

The DLA is the most important functional difference

The Deep Learning Accelerator (DLA) is a fixed-function neural network accelerator separate from the GPU. It is present on AGX Orin and Orin NX, and absent on Orin Nano.

Why it matters: TensorRT can partition models to run layers on DLA and the rest on GPU, freeing the GPU for other tasks. In a camera pipeline running continuous inference, the DLA handles the inference engine while the GPU handles preprocessing, postprocessing, and other concurrent workloads. Without the DLA, all inference runs on the GPU — which works, but consumes more power and competes with other GPU work.

In practice: if you are running a YOLOv8 model on a camera stream alongside GStreamer decode and display, the DLA on Orin NX lets the inference run independently of the GPU, keeping the overall pipeline throughput higher. On Orin Nano, everything shares the GPU, and you will hit GPU utilization limits sooner.

For purely CPU-bound workloads (ROS2 processing, control loops, sensor fusion without neural networks), the DLA difference is irrelevant and Orin Nano is a perfectly capable platform.

Memory bandwidth is the real runtime bottleneck

TOPS figures are easy to compare but memory bandwidth is what you actually feel in production. The GPU and DLA share memory with the CPU on all Jetson modules. High-resolution video decode, GStreamer pipeline buffers, model activations, and postprocessing all compete for that bandwidth.

A concrete example: running two concurrent 1920×1080@30fps camera streams through a YOLOv8-M model:

• AGX Orin: saturates at about 20% of memory bandwidth — headroom for more workloads
• Orin NX 16GB: saturates at about 40% — workable for 2 cameras, tight for 4
• Orin Nano 8GB: saturates at about 65–70% — leaves little room for additional processing

If you are building a surround-view system with 4+ cameras doing concurrent inference, memory bandwidth on Orin Nano will bottleneck you before GPU compute does.

Camera I/O: how many MIPI CSI cameras fit

AGX Orin has substantially more MIPI CSI-2 lane capacity than the NX and Nano modules:

• AGX Orin: 8 MIPI CSI-2 ports, up to 16 cameras total with a deserializer
• Orin NX / Orin Nano: 4 MIPI CSI-2 ports, up to 4 cameras directly or up to 8 via GMSL2 with a quad-deserializer

For GMSL2 camera systems, the deserializer aggregates multiple cameras onto fewer MIPI lanes. A MAX96724 quad-deserializer routes 4 GMSL2 cameras over 2 MIPI ports, so Orin NX can support 4–8 GMSL2 cameras in practice. For 6+ camera surround-view or full surround-view AVM systems, AGX Orin is typically required.

For direct CSI cameras (no SerDes): Orin NX supports 4 simultaneous cameras. AGX Orin supports up to 16 with the right carrier board.

PCIe expansion: when AGX Orin is the only option

If your carrier board design requires multiple PCIe devices simultaneously — for example, an NVMe SSD plus a PCIe Ethernet controller plus a PCIe-connected FPGA or capture card — only AGX Orin has the lane budget.

Orin NX and Orin Nano expose ×4 PCIe lanes total. In practice this means one NVMe SSD OR one additional PCIe device, not both simultaneously. For designs that need full PCIe expansion, plan for AGX Orin from the start.

Which module to choose

Choose AGX Orin when:

• You need 6 or more MIPI CSI cameras
• You need multiple PCIe devices simultaneously (NVMe + Ethernet + other)
• Your inference workload requires more than 100 TOPS
• You need 10GbE on the module
• You are in a development phase and want maximum headroom

Choose Orin NX 16GB when:

• Your application needs 1–4 cameras
• You need DLA for sustained inference at low power
• Your PCIe requirements fit within ×4
• You’re building a product and 25W fits your thermal design
• You want Orin NX and Orin Nano to share the same carrier board for cost-tier SKUs

Choose Orin Nano when:

• Cost and power envelope are the primary constraints
• Your workload doesn’t require DLA (simple models, CPU-based processing, or occasional inference)
• You can accept PCIe Gen 3 ×4
• You are prototyping and will move to NX for production if performance headroom is needed

Carrier board compatibility

Orin NX and Orin Nano use the same 260-pin SO-DIMM style connector and are pin-compatible with each other and with Jetson Xavier NX. A carrier board designed for Xavier NX will generally work with Orin NX and Orin Nano — with a DTS and BSP update, not a hardware change.

AGX Orin uses a different 699-pin module connector and requires a carrier board designed specifically for the AGX form factor. NVIDIA’s AGX Orin Developer Kit carrier board is the reference design. Third-party carrier boards for AGX Orin are available from Auvidea, Leopard Imaging, Sievert Larsen, and others.

For GMSL2 camera bring-up on any of these modules, the process is the same — the module choice affects how many cameras you can connect and which NVCSI ports you route them to. See GMSL2 camera bring-up on Jetson Orin for the full setup sequence.

For help selecting the right module and carrier board for your application, the Jetson expert support service covers module selection, carrier board design review, and BSP bring-up.

NVIDIA’s official module comparison is on the Jetson modules page. The full hardware design guidelines for each module are in the Jetson Linux Developer Guide.