CoChat

Benchmarking YOLOv8 Variants for Object Detection Efficiency on Jetson Orin NX for Edge Computing Applications

Open access

The DOI matches its Crossref record; 49 of the 56 references it cites were found in a registry or an index; 2 of the 8 sources its claims cite were checked for support.

Abstract

Edge AI is redefining the deployment of computer vision systems by enabling real-time inference directly on resource-constrained edge devices. This shift offers significant advantages in terms of reduced latency, data privacy, and operational autonomy in bandwidth-limited computing environments. In this paper, we present a systematic performance benchmarking of multiple variants of YOLOv8 on the NVIDIA Jetson Orin NX platform, focusing on object detection tasks. We evaluate inference latency, frame throughput, and computational resource usage across varying input sizes and model complexities. Furthermore, we validate the deployment effectiveness through practical use cases, such as vehicle and package detection. Our findings show that the TensorRT model outperforms PyTorch by 17.7% at a batch size of 2, although PyTorch presents greater stability at larger batch sizes (e.g., 8), where TensorRT encounters resource constraints. In terms of memory usage, it increases linearly with batch size: 69% from batch 1 to 4, with TensorRT requiring 429.20 MB at batch size 2 compared to PyTorch’s 451.24 MB. Furthermore, the processing time per image decreases by 42% when scaling from batch size 1 to 4, highlighting a critical saturation point for edge resources. In summary, the results provide insight into the trade-offs between model size and speed, offering guidance for selecting detection architectures tailored to real-time edge applications.

Overview

CoChat Summary

TL;DR

This paper benchmarks five YOLOv8 variants on an NVIDIA Jetson Orin NX 16 GB for edge object detection, comparing PyTorch and TensorRT across batch sizes, input resolutions, and two use cases. It reports TensorRT speedups, batch-size efficiency gains, memory scaling, and practical stability limits for real-time deployment.

Key findings

  • TensorRT outperformed PyTorch by 17.7% at batch size 2, but PyTorch was more stable at larger batch sizes such as 8, where TensorRT encountered resource constraints. For vehicle detection, TensorRT reduced runtime by 52.4–63.3% versus PyTorch; YOLOv8x showed the largest absolute improvement at 36.48 s faster and a 63.3% reduction, while YOLOv8n.engine was fastest at 13.218 s.
  • Scaling from batch 1 to 4 in TensorRT-optimized vehicle/package systems improved frame rate by 35.4% to 32.04 FPS and reduced per-frame detection time by 39.9% to 16.27 ms; total execution time fell 26.2% from 12.72 to 9.39 s. GPU memory increased 69% from 25.46 MB to 42.96 MB, with batch 2 showing the lowest memory increase at 429.20 MB versus 453.91 MB at batch 4.
  • Vehicle detection showed an accuracy–latency trade-off: YOLOv8x achieved the highest accuracy (0.75) but required 144.636 s, while YOLOv8n was fastest (105.894 s) with the lowest accuracy (0.61). Batch processing improved runtime by approximately 16–21%, with YOLOv8m showing the highest relative improvement (21%).
  • For package detection, fine-tuned YOLOv8n achieved 81.5% precision and 72.2% recall across all classes; the package class reached 86.4% precision and 77.3% recall, with mAP@0.5 of 81.7% and mAP@0.5:0.95 of 67.3%.

Methods

  • Hardware/software: NVIDIA Jetson Orin NX 16 GB in MAXN (25 W) mode, JetPack 5.1.1 (L4T 35.3.1), CUDA 11.4, TensorRT 8.5; training performed in Google Colab.
  • Models/data: five YOLOv8 variants (n/s/m/l/x, 3.0–68.2 M parameters) initialized with COCO pretrained weights, SGD with lr0 = 0.01, 300 epochs, patience 50. Vehicle detection used 2644 frames; package detection used a Roboflow dataset with 12,042 training, 1187 validation, and 16 test images; stratified 80/10/10 split with Mosaic, MixUp, and HSV augmentation.
  • Optimization/evaluation: PyTorch weights exported to ONNX (Opset 12), compiled with trtexec to TensorRT engines using FP16; custom Python wrapper with GPU NMS; each run repeated ten times. Metrics included mAP@0.5, end-to-end latency, FPS across batch sizes 1, 2, 4, 8, 16, 32, 64, and 128, and resolutions 640 × 640 and 1200 × 1200, comparing PyTorch versus TensorRT.

Limitations

  • The paper states performance metrics came from a limited dataset; real-world variable object densities may yield different latency profiles. The baseline-crossing counting logic assumes generally unidirectional traffic, does not account for vehicles reversing after crossing the line, and may miss counts or switch IDs under significant partial occlusions at the counting line.
  • The paper states experiments used a custom active cooling kit; passive cooling enclosures could thermal-throttle earlier and reduce sustained FPS below reported values. Direct comparisons with other edge AI studies are difficult due to hardware configuration variations such as JetPack versions, power modes, and background process overhead.
  • Evident from the text: the paper reports conflicting batch-size stability limits—recommending capping TensorRT batch sizes at 4 after OOM failures above batch 8 in one section, while another section recommends capping at 64 after OOM failures above batch 64 for YOLOv8l/x. The abstract’s PyTorch memory figure of 451.24 MB at batch 2 is also not reflected in Table 5.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/computers15020074

Cites

49 of 56 references verified0 citation errors · 7 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 1 Not valid · 0 Not checked · 69 Own results and interpretations

45 more claims

Create a free account to see all 71 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 18
  • Probably not a substantial claim. 5

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Cited by

See all 24

Metadata

Authors
Aljami, H. M., Alrowais, N. A., AlAwajy, A. M., Alhrgan, S. O., Aldwaani, R. A., Alsawadi, M. S., Saqib, N. U., Alam, S. S., & Alsubaie, R.
Affiliations
  • King Abdulaziz City for Science and Technology (SA)
  • Riyadh Elm University (SA)
Corresponding author
Alsawadi, Motasem Samer
Venue
Computers
Publisher
MDPI AG
Published
2026-02-01
Volume
15
Issue
2
Pages
74
ISSN
2073-431X (ISSN-L)
OpenAlex
W7126409579
Cited by
24
References
56
Citations by year
2026: 24
FWCI
68.61
Type
journal article
Language
en
Open access
Locations
  • Computers (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Networks and Communications › IoT and Edge/Fog Computing
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Object Detection Techniques
Keywords
YOLOv8, object detection, edge computing, inference latency, memory usage, batch size, TensorRT, PyTorch, vehicle detection

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-30 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.