Comprehensive performance benchmarks for YOLO26 object detection in Home Security Intelligence.
This document provides detailed performance metrics for YOLO26 variants used in the detection pipeline. For the auto-generated benchmark results, see yolo26-benchmarks.md .
Provenance (verified 2026-09-22): The tables below are a curated benchmark from the TensorRT FP16 era (CUDA 12.4 / TensorRT 10.0 / driver 550.54.14 on the RTX A5500) and are not reproducible with the scripts at the bottom of this page -- scripts/benchmark_yolo26_latency.py now measures through the Ultralytics .pt stack, and its regenerated output (the sibling yolo26-benchmarks.md ) shows a very different profile (e.g. yolo26n at 640x640: 44.83 ms mean there vs 5.2 ms here). The deployed detector also differs from these tables: YOLO26m runs as FP32 ONNX under Triton's ONNX Runtime backend inside ai-gateway (TensorRT was dropped for Triton -- 4 GB A400 workspace limits, see NVIDIA Technology Inventory ), and TensorRT engines remain an option only for the standalone dev server and the export pipeline. Treat the numbers as best-case TensorRT reference points, not current production latency.
Executive Summary YOLO26 is the sole object detection model in the Home Security Intelligence pipeline. It offers excellent performance for real-time security monitoring with multiple variants optimized for different use cases.
Metric yolo26n yolo26s yolo26m (default) Inference Time 5-6ms 10-12ms 15-20ms FPS (TensorRT) 170-200 85-100 50-70 VRAM Usage ~100 MB ~150 MB ~200 MB mAP@50 (COCO) 37.8% 44.3% 50.9% Parameters 2.57M 10.01M 21.90M
Model Variants yolo26n (Nano) Best for: Edge devices, maximum throughput, resource-constrained environments.
Specification Value Parameters 2.57M Model Size 5.3 MB (FP32), 2.7 MB (FP16) Inference 5-6ms @ 640x640 (TensorRT FP16) Throughput 170-200 FPS VRAM ~100 MB mAP@50 37.8% Best For High-volume, multi-camera, edge devices
yolo26s (Small) Best for: Balanced speed and accuracy, standard deployments.
Specification Value Parameters 10.01M Model Size 19.5 MB (FP32), 9.8 MB (FP16) Inference 10-12ms @ 640x640 (TensorRT FP16) Throughput 85-100 FPS VRAM ~150 MB mAP@50 44.3% Best For Standard security cameras, balanced use
yolo26m (Medium) - Default Best for: High accuracy requirements, fewer cameras.
Specification Value Parameters 21.90M Model Size 42.2 MB (FP32), 21.1 MB (FP16) Inference 15-20ms @ 640x640 (TensorRT FP16) Throughput 50-70 FPS VRAM ~200 MB mAP@50 50.9% Best For Maximum detection accuracy
Accuracy Benchmarks Detection Accuracy by Class Performance on security-relevant COCO classes:
Class yolo26n yolo26s yolo26m person 52.1% 58.4% 63.2% car 48.3% 54.7% 60.1% truck 41.2% 47.8% 53.4% dog 45.6% 51.2% 56.8% cat 44.8% 50.5% 55.9% bird 38.2% 44.1% 49.7% bicycle 39.7% 45.9% 51.3% motorcycle 42.4% 48.6% 54.2% bus 54.1% 60.3% 65.8%
Confidence Threshold Impact Effect of confidence threshold on precision/recall:
Threshold Precision Recall F1 Score 0.25 72.3% 89.1% 79.8% 0.50 85.6% 76.4% 80.7% 0.75 92.1% 58.2% 71.3%
Recommendation: Use 0.50 confidence threshold (default) for balanced precision/recall.
Latency Benchmarks Single Image Inference Tested on NVIDIA RTX A5500 (24GB VRAM) with TensorRT FP16:
Model Resolution Mean (ms) P50 (ms) P95 (ms) P99 (ms) yolo26n 640x640 5.2 5.1 5.8 6.2 yolo26n 1280x1280 18.4 18.2 19.6 20.3 yolo26s 640x640 10.8 10.6 11.9 12.5 yolo26s 1280x1280 38.6 38.2 41.2 43.1 yolo26m 640x640 17.3 17.1 18.8 19.6 yolo26m 1280x1280 62.4 61.8 67.2 70.1
Latency Breakdown Component timing for yolo26m @ 640x640:
Stage Time (ms) Percentage Preprocessing 0.8 4.6% Inference 16.1 93.1% Postprocessing 0.4 2.3% Total 17.3 100%
VRAM Usage Model Loading Model VRAM (FP32) VRAM (FP16) VRAM (INT8) yolo26n 180 MB 100 MB 65 MB yolo26s 280 MB 150 MB 95 MB yolo26m 380 MB 200 MB 125 MB
Runtime VRAM (including buffers) Model Batch 1 Batch 4 Batch 8 Batch 16 yolo26n 150 MB 200 MB 280 MB 450 MB yolo26s 220 MB 320 MB 480 MB 800 MB yolo26m 300 MB 450 MB 700 MB 1.2 GB
Throughput Benchmarks Frames Per Second (FPS) Sustained throughput on RTX A5500:
Model Batch 1 Batch 4 Batch 8 Batch 16 yolo26n 192 256 280 290 yolo26s 93 128 142 148 yolo26m 58 78 86 90
Multi-Camera Throughput Cameras processed at 5 FPS each:
Model Max Cameras Total FPS yolo26n 50+ 250+ yolo26s 25 125 yolo26m 15 75
TensorRT Optimization Format Load Time Inference Portability PyTorch (.pt) 2.5s 45ms High ONNX (.onnx) 1.2s 28ms High TensorRT (.engine) 0.5s 17ms GPU-specific
TensorRT Optimization Gains Optimization Latency Reduction VRAM Reduction FP16 40-50% 50% INT8 60-70% 75% Layer Fusion 10-15% 5-10%
Hardware Scaling GPU Comparison Performance across different NVIDIA GPUs (yolo26m @ 640x640):
GPU VRAM Inference (ms) FPS RTX 4090 24 GB 12.1 83 RTX A5500 24 GB 17.3 58 RTX 4070 12 GB 19.8 50 RTX 3070 8 GB 25.4 39 RTX 3060 12 GB 31.2 32 T4 (Cloud) 16 GB 38.5 26
Multi-GPU Scaling Linear scaling with multiple GPUs:
Configuration Throughput (FPS) Efficiency 1x RTX A5500 58 100% 2x RTX A5500 112 97% 4x RTX A5500 220 95%
Recommendations Model Selection Guide Use Case Recommended Rationale 1-4 cameras, accuracy priority yolo26m Best detection accuracy 5-15 cameras, balanced yolo26s Good speed/accuracy 16+ cameras, throughput priority yolo26n Maximum throughput Edge deployment (Jetson) yolo26n Low VRAM requirement High-resolution (4K) cameras yolo26s Balanced for larger images
Latency Targets Application Target Model Resolution Real-time alerts (<100ms) <50ms yolo26n/s 640x640 Near real-time (<500ms) <200ms yolo26m 640x640 High accuracy (<1000ms) <500ms yolo26m 1280x1280
Benchmark Methodology Test Configuration Hardware: NVIDIA RTX A5500 (24GB VRAM)CUDA Version: 12.4TensorRT Version: 10.0Driver Version: 550.54.14Warmup Iterations: 50Benchmark Iterations: 1000Input: Synthetic images with random pixel values Reproducibility Run benchmarks locally:
# Latency benchmark
run python scripts/benchmark_yolo26_latency.py
# Accuracy benchmark
run python scripts/benchmark_yolo26_accuracy.py
# GPU benchmark in container
run python scripts/benchmark_yolo26_container.py
Back to Reference Hub | Benchmarks Hub