· Research & technology· 日本語で読む
Seeed Wiki compares 4K industrial vision processing on Jetson Orin Nano and Orin NX
Seeed Wiki compared 4K industrial vision processing on the Jetson Orin Nano 8GB and Orin NX 16GB. The test used four 4K H.264 video streams and included YOLOv8n inference, recording and browser display. According to the reported results, the Orin NX delivered 1.33 times the inference throughput and 69.7% lower end-to-end latency.
What happened
On November 14, 2026, Seeed Wiki published a benchmark of 4K industrial vision processing using the Jetson Orin Nano 8GB and Jetson Orin NX 16GB. The test used four 4K H.264 video streams and combined TensorRT FP16 YOLOv8n inference, recording and a browser dashboard. According to the article's measurements, under the same video-file, sequential-inference and recording conditions, the Orin NX delivered 1.33 times the inference throughput, 63.3% lower normalized CPU load and 69.7% lower end-to-end latency.
Why it matters
Seeed Wiki's measurements show the difference when 4K video inference, recording and display are handled simultaneously in a configuration without recording functionality and one using hardware NVENC. For industrial vision system design, the results illustrate that selecting a configuration based on recording method and CPU load, as well as inference performance, can affect latency and available processing headroom.
Analysis
The comparison focuses not on standalone image-inference speed but on overall processing when video decoding, inference, recording and preview delivery run simultaneously. The Orin Nano does not have NVENC and uses CPU-based x264 emulation, while the Orin NX uses hardware NVENC. The reported difference should therefore be understood as a difference between the overall configurations, including recording, rather than inference processing alone. For those deploying systems that handle industrial camera footage, the comparison also shows how recording affects CPU load and latency, in addition to inference performance.
However, the two systems did not use identical software configurations. The published table lists JetPack 6.2, Jetson Linux L4T 36.4.3, CUDA 12.6.10 and TensorRT 10.3.0 for the Orin Nano, and JetPack 5.1.3, L4T 35.5, CUDA 11.4 and TensorRT 8.5.2 for the Orin NX. Although the video files, sequential inference and recording targets were held constant, the results cannot be attributed solely to differences between the GPU or SoC. The recording methods also differed, so the figures should be treated as an example measured with a similar processing pipeline, rather than applied directly to other encoder configurations or applications.
Seeed Studio lists the “reComputer Super J3010” with Jetson Orin Nano, rated at up to 34 TOPS, a report by this newspaper, covered the maximum 34 TOPS and 10–25W power settings listed on the Jetson Orin Nano 4GB product page. The benchmark in this report covers the Orin Nano 8GB and measures processing under different conditions from the product page's TOPS figure. The key point when reading the two reports is not to equate peak-compute ratings with application performance that includes recording. Meanwhile, NVIDIA presents a Python inference method using cuDLA on Jetson AGX Orin covers an inference method on a different Jetson product and does not corroborate the comparison between the Orin Nano and Orin NX.
Future evaluations will focus on whether the systems can be compared under identical software configurations, cooling conditions during extended operation, power consumption and processed frame counts. The text provided does not specify figures or conditions for these factors, and it is not possible to determine whether the reported results would be reproduced with different video content or recording settings.
Sources
Translated from the Japanese original: /news/seeed-wiki-jetson-orin-nano-to-orin-nx-no-4k-sangyouyou-bijo-e9fb11. The Japanese version is authoritative if they differ.