Jetson Orin vs Xavier: AGX Orin and AGX Xavier Compared

Jetson Orin vs Xavier is a generational comparison: Orin replaces Xavier’s Volta GPU, Carmel CPU and LPDDR4x memory with an Ampere GPU, Arm Cortex-A78AE cores and LPDDR5, and the top AGX module moves from 32 to 275 TOPS. For most workloads that means several times the real inference throughput, more memory bandwidth and a software line that carries on into JetPack 6 and 7.

Both generations come in two sizes. The Jetson AGX Xavier gave way to the Jetson AGX Orin on the same 699-pin connector family, and the Jetson Xavier NX to the Orin NX on the SO-DIMM footprint. This guide compares them side by side with NVIDIA’s published figures, explains why the headline 8× is not the gain you will measure, and sets out what changes when a design moves from one to the other.

Xavier and TX2 are now at end of life, so for fielded systems this is also a supply question. Our guide to Xavier and TX2 end of life and migration to Orin covers the last-time-buy and refresh planning; this page covers the hardware and software differences behind that move.

The table puts the two AGX generations side by side using NVIDIA’s module datasheet figures. Orin AI figures are sparse INT8; Xavier figures are dense INT8, which is how NVIDIA quoted each generation.

NVIDIA also offers an Industrial version of each generation. The AGX Xavier Industrial is rated at 30 TOPS in 20–40 W with ECC memory; the AGX Orin Industrial reaches 248 TOPS in 15–75 W, with 64 GB of ECC memory and a module rating from −40 °C to 85 °C. Our answer on what replaced the Jetson Xavier pairs every Xavier and TX2 module with its Orin successor.

Two rows deserve a second look. Memory bandwidth rises by about 1.5×, which often matters as much as TOPS for camera-heavy pipelines. And hardware encode falls: AGX Xavier can encode four 4K60 streams in H.265, AGX Orin 64GB two. Check NVIDIA’s current datasheets for exact stream counts at each resolution before you migrate a recorder or video gateway.

NVIDIA’s headline is up to eight times the AI performance, from 32 to 275 TOPS. The figure is real, but it compares two different things, and the speed-up you see on your own model will usually be smaller.

Worked example. A detection model that runs at 30 frames per second on AGX Xavier is limited by the GPU. On a dense basis AGX Orin 64GB has roughly four times the tensor throughput, plus 1.5× the bandwidth, so a gain of around three to four times is a reasonable planning figure, reaching 100 frames per second or more. With a sparsity-trained model and current TensorRT the gain can be higher. Treat these as estimates: the only reliable number is a benchmark of your model, at your resolution, in the power mode the enclosure can sustain.

Precision changes the figure too. Our explainer on what TOPS means in AI covers how TOPS are counted, and the difference between CUDA cores and Tensor Cores explains which part of the GPU does the work. The generational step continues above Orin: Jetson Thor vs Orin sets out what the Blackwell modules add.

AGX Xavier and AGX Orin share a size and a connector family, but almost every block inside the module is new. These are the changes that matter in practice.

Ampere GPU with third-generation Tensor Cores

Xavier’s Volta GPU introduced Tensor Cores to Jetson. Orin’s Ampere GPU moves to third-generation Tensor Cores with more CUDA cores per streaming multiprocessor, better performance per watt and support for structured sparsity, which skips the zero weights in a pruned network to roughly double throughput. Each streaming multiprocessor runs CUDA threads in parallel and holds the Tensor Cores that accelerate the matrix maths at the heart of deep learning, so more and newer SMs translate directly into faster inference.

Ampere Tensor Cores also support more data types, including TF32, alongside FP16 and INT8. Lower precision moves less data and runs faster, which is where TensorRT gets much of its speed on Jetson.

Arm Cortex-A78AE CPU

Orin replaces NVIDIA’s own Carmel cores with Arm Cortex-A78AE cores, 12 on AGX Orin 64GB and eight on the 32GB module, against eight Carmel cores on AGX Xavier. NVIDIA quotes about 1.7 times the CPU performance. Each A78AE core has 64 KB instruction and 64 KB data L1 cache and 256 KB of L2, with shared L3 per cluster, and the cores run at up to about 2.2 GHz on AGX Orin. The CPU matters more than it looks: image decoding, tracking, ROS nodes and application logic all run there.

Memory and storage

AGX Orin uses 256-bit LPDDR5 at up to 204.8 GB/s, against 136.5 GB/s of LPDDR4x on AGX Xavier, and doubles the onboard eMMC from 32 GB to 64 GB. The move to LPDDR5 is also the reason Orin is unaffected by the LPDDR4 shortage that ended Xavier early.

Video encode, decode and JPEG

Both modules carry a dedicated video encoder (NVENC), decoder (NVDEC) and JPEG engine (NVJPG), which take video work off the CPU and GPU. Orin’s decoder handles H.265, H.264, AV1 and VP9, and its JPEG engine handles compression, decompression, scaling and colour conversion. As the table shows, encode capacity is lower on AGX Orin than AGX Xavier, which mainly affects recorders and multi-stream video gateways.

Deep learning and vision accelerators

Both generations add fixed-function engines beside the GPU. Orin’s second-generation NVDLA engines run supported network layers at low power, and its second-generation Programmable Vision Accelerator (PVA v2) handles classic computer vision such as filtering, feature tracking and stereo, leaving the GPU free for deep learning. Our explainer on the wider family of AI accelerators sets these engines in context.

Higher performance raises the power ceiling from 30 W to 60 W, so cooling has to be revisited too. Our Jetson thermal design guide covers how to cool an AGX Orin without thermal throttling.

Largely, but not as a simple swap. AGX Orin uses the same 699-pin connector family and module size as AGX Xavier, and NVIDIA designed it as the successor, so many carrier boards were offered for both. Power rails, high-speed lanes and some pin functions differ, and the higher power draw needs a stronger supply and heatsink, so every carrier needs a design review and a board support package for Orin.

Carriers in our range show the pattern. Forecr’s DSBOARD‑AGXS is offered with the AGX Orin 32GB, 64GB and Industrial modules on one board, rated from −40 °C to 85 °C, and our Jetson AGX Orin carrier boards list which modules each board supports. Our guide to choosing a Jetson carrier board covers what to review.

Should a drone use Orin Nano 8GB or Xavier NX 16GB?

This is a common question on drones and small robots, because the two modules share a footprint. Orin Nano 8GB has far more AI compute, 67 TOPS in its Super mode against 21, but half the memory and no hardware video encoder. Xavier NX 16GB keeps 16 GB and encodes up to two 4K60 streams in hardware. For a new design the better replacement is usually the Orin NX 16GB, which has both: up to 157 TOPS, 16 GB and a hardware encoder. Our comparison of Orin Nano vs Orin NX covers that choice. If the AGX tier is also in question, our guide to Orin NX vs AGX Orin sets out when the larger module pays.

Does AGX Xavier support JetPack 6? No. Xavier’s software line stops at JetPack 5.1, while Orin runs JetPack 6 and moves on to JetPack 7.2. That is often the bigger part of a migration than the hardware. Our guide to NVIDIA JetPack explains how each release ties the operating system and libraries together.

In practice the application code usually carries across, because CUDA and the libraries above it keep the same programming model. What has to be rebuilt is the image, the device tree for the carrier and every TensorRT engine, because an engine is optimised for one GPU and one TensorRT version. Our explainers on TensorRT and the board support package cover the two pieces most likely to need work.

Yes. NVIDIA brought forward the end of life of its LPDDR4-based modules, including Xavier NX, AGX Xavier and the TX2 family, because of the shortage of LPDDR4 memory. Last-time-buy orders closed on 1 July 2026, and NVIDIA ships the final modules against placed orders until 15 July 2027.

For a new design, choose Orin or Thor. For a fielded Xavier system, the usual plan has two tracks: enough last-time-buy stock for production, repairs and spares, and an Orin version designed and validated in parallel. Our answer on how long NVIDIA Jetson is supported sets out the lifecycle of every generation, and the Jetson modules comparison helps choose the Orin module to move to. Existing Jetson Xavier computers remain available only while stock lasts.

We have shipped both generations into the field, and the migration is not theoretical for us. AGX Xavier powers autonomous optical detection and tracking and the IP67 Jetson computers in a lobster farming facility. AGX Orin runs spot spraying with GigE cameras in agriculture and ISR drone video analytics for Airbus Defence and Space.

In each case the move to Orin brought the same benefits: more models on one computer, more cameras at full frame rate and a software line that will keep receiving updates. It also brought the same work: a carrier review, a new JetPack image and a thermal check at the new power mode. For harsh sites the rugged Jetson computers in our range are the usual starting point, and we configure, image and test each one before it ships.

The facts below come from each product’s own specification.

The Industrial-module computer gives up some peak TOPS for its temperature range, which is the usual trade on cold or exposed sites.