M.A. Kukushkin1, P.A. Ukhov2
1, 2 Moscow Aviation Institute (National Research University)

1 cnegbyj99@mail.ru, 2 ukhovpa@mai.ru

Specialized hardware accelerators like Tensor Processing Units (TPUs) are promising for deploying real-time computer vision models, but their efficiency heavily depends on compatibility with a specific neural network architecture. YOLO family object detection models (v5, v8, v11), being the de facto standard, are initially optimized for GPUs. Practical questions arise: which of the modern YOLO versions adapts most efficiently to TPUs in terms of the balance between speed, accuracy, and porting complexity, and how does their performance differ from the reference GPU implementation?

The objective of the article is to conduct a comparative performance analysis of modern YOLO models (v5, v8, v11 by Ultralytics) after their adaptation for execution on TPUs, aiming to identify the most suitable model for practical application on this hardware platform.

Using a porting methodology based on TensorFlow and optimizations (quantization), YOLOv5, YOLOv8, and YOLOv11 models have been adapted and tested. A comparative analysis of key metrics (inference speed, mAP accuracy, TPU resource utilization efficiency) allowed for ranking the models and identifying the leader, which demonstrates the optimal compromise between minimal accuracy degradation and maximum speedup on TPUs compared to GPUs.

The results provide engineers and developers with evidence-based recommendations for selecting a specific YOLO model version for deployment on TPUs. This helps to reduce the time and resources required to find an optimal solution when creating resource-efficient computer vision systems for autonomous driving, video surveillance, and robotics tasks.

Kukushkin M.A., Ukhov P.A. Comparison of performance of neural network models of computer vision applied to tensor processors. Neurocomputers. 2026. V. 28. № 2. P. 44–50. DOI: https://doi.org/10.18127/j19998554-202602-04 (in Russian)

1. COCO dataset [Elektronnyj resurs]. URL: https://cocodataset.org/#home (data obrashcheniya: 01.04.2025).
2. Coral AI. Edge TPU Models Introduction [Elektronnyj resurs]. URL: https://coral.ai/docs/edgetpu/models-intro/ (data obrashcheniya: 01.04.2025).
3. Coral AI. Products Accelerator [Elektronnyj resurs]. URL: https://coral.ai/products/accelerator/ (data obrashcheniya: 01.04.2025).
4. Nebaba S.G., Markov N.G. Svertochnye nejronnye seti semejstva YOLO dlya mobil'nykh sistem komp'yuternogo zreniya. Komp'yuternaya optika. 2024. T. 16. № 3. S. 615–631. DOI: 10.20537/2076-7633-2024-16-3-615-631. (in Russian)
5. ONNX [Elektronnyj resurs]. URL: https://onnx.ai/ (data obrashcheniya: 01.04.2025).
6. Hidayatullah P., Syakrani N., Sholahuddin M.R. et al. YOLOv8 to YOLO11: A comprehensive architecture in-depth comparative review. Politeknik Negeri Bandung. Stunning Vision AI. Indonesia. 2025.
7. Raspberry Pi OS [Elektronnyj resurs]. URL: https://www.raspberrypi.com/software/ (data obrashcheniya: 02.06.2025).
8. Ultralytics documentation. YOLO11 vs YOLOv8 comparison [Elektronnyj resurs]. URL: https://docs.ultralytics.com/ru/compare/yolo11-vs-yolov8/ (data obrashcheniya: 01.04.2025).
9. Ultralytics documentation. YOLOv8 vs YOLOv5 comparison [Elektronnyj resurs]. URL: https://docs.ultralytics.com/ru/compare/yolov8-vs-yolov5/ (data obrashcheniya: 01.04.2025).
10. YOLOv6. Meituan [Elektronnyj resurs]. URL: https://github.com/meituan/YOLOv6?spm=a2ty_o01.29997173.0.0.60bcc92125ZXJa (data obrashcheniya: 02.06.2025).
11. YOLOv7. A. Bochkovskiy [Elektronnyj resurs]. URL: https://github.com/AlexeyAB/darknet? pm=a2ty_o01.29997173.0.0.60bcc92125ZXJa (data obrashcheniya: 02.06.2025).
12. YOLOv10. THU-MIG [Elektronnyj resurs]. URL: https://github.com/THU-MIG/yolov10 (data obrashcheniya: 02.06.2025).