Skip to content

Add model benchmark data #2025

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
samfreund merged 4 commits into from
Aug 3, 2025
Merged

Add model benchmark data #2025

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
354ea93
Add model benchmark data!
boomermath Aug 3, 2025
59412a8
Fix build issues hopefully?
boomermath Aug 3, 2025
8f40df0
formatting, lint
boomermath Aug 3, 2025
93fc703
Add benchmarks section
boomermath Aug 3, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions .python-version
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
3.11
8 changes: 8 additions & 0 deletions docs/source/docs/benchmarks/index.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
# Performance Benchmarks

```{toctree}
:maxdepth: 0
:titlesonly: true

rknn-model-benchmarks
```
125 changes: 125 additions & 0 deletions docs/source/docs/benchmarks/rknn-model-benchmarks.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,125 @@
# RKNN Benchmarks

## Description
This benchmark compares the performance of four object detection models: YOLOv5, YOLOv5u, YOLOv8, and YOLOv11 on the [COCO 2017 Validation Set](http://images.cocodataset.org/zips/val2017.zip). The main purpose is to assess and compare the inference speed and detection accuracy of these models when deployed on the Orange Pi devices using the RKNN framework and int8 quantization.

## Methodology
- **Dataset**: [COCO 2017 Validation Set](http://images.cocodataset.org/zips/val2017.zip) (5,000 images)

- **Platform**: Orange Pi 5 with RK3588

- **Quantization**: int8 using 20 randomly selected images from the validation set

- **Framework**: RKNN Toolkit 2

## Operator-Level Benchmark Results

The following tables break down the average CPU time, NPU time, and total execution time (in microseconds) for each operator used by the models. Each value represents the mean ± standard deviation across 5,000 inferences.

### YOLOv5

| OpType | CPU Time (μs) | NPU Time (μs) | Total Time (μs) | Time Ratio (%) | Number of Times Called |
|-----------------|---------------------|----------------------|-----------------------|---------------------|-----------------------|
| ConvExSwish | 0.00 ± 0.00 | 10968.81 ± 1126.00 | 10968.81 ± 1126.00 | 73.06 ± 0.94 | 57 |
| ConvSigmoid | 0.00 ± 0.00 | 1243.49 ± 67.66 | 1243.49 ± 67.66 | 8.33 ± 0.57 | 3 |
| Concat | 0.00 ± 0.00 | 1080.68 ± 259.40 | 1080.68 ± 259.40 | 7.09 ± 0.87 | 13 |
| Conv | 0.00 ± 0.00 | 732.15 ± 29.42 | 732.15 ± 29.42 | 4.92 ± 0.42 | 1 |
| Add | 0.00 ± 0.00 | 473.71 ± 131.48 | 473.71 ± 131.48 | 3.10 ± 0.50 | 7 |
| MaxPool | 0.00 ± 0.00 | 272.40 ± 110.52 | 272.40 ± 110.52 | 1.76 ± 0.51 | 6 |
| Resize | 0.00 ± 0.00 | 147.61 ± 38.89 | 147.61 ± 38.89 | 0.97 ± 0.15 | 2 |
| OutputOperator | 106.60 ± 15.00 | 0.00 ± 0.00 | 106.60 ± 15.00 | 0.72 ± 0.13 | 3 |
| InputOperator | 8.64 ± 1.79 | 0.00 ± 0.00 | 8.64 ± 1.79 | 0.06 ± 0.02 | 1 |
| **Total** | **115.24 ± 16.16** | **14918.85 ± 1735.45**| **15034.09 ± 1734.28**| | **93** |

### YOLOv5u

| OpType | CPU Time (μs) | NPU Time (μs) | Total Time (μs) | Time Ratio (%) | Number of Times Called |
|-----------------|---------------------|----------------------|-----------------------|---------------------|-----------------------|
| ConvExSwish | 0.00 ± 0.00 | 16828.24 ± 1332.73 | 16828.24 ± 1332.73 | 83.04 ± 1.61 | 69 |
| Concat | 0.00 ± 0.00 | 1265.94 ± 250.24 | 1265.94 ± 250.24 | 6.17 ± 0.69 | 13 |
| ConvSigmoid | 0.00 ± 0.00 | 613.88 ± 62.97 | 613.88 ± 62.97 | 3.03 ± 0.15 | 3 |
| Add | 0.00 ± 0.00 | 553.75 ± 131.17 | 553.75 ± 131.17 | 2.69 ± 0.44 | 7 |
| Conv | 0.00 ± 0.00 | 298.61 ± 72.72 | 298.61 ± 72.72 | 1.45 ± 0.25 | 3 |
| ConvClip | 0.00 ± 0.00 | 256.02 ± 64.48 | 256.02 ± 64.48 | 1.24 ± 0.23 | 3 |
| MaxPool | 0.00 ± 0.00 | 178.68 ± 58.72 | 178.68 ± 58.72 | 0.86 ± 0.23 | 3 |
| Resize | 0.00 ± 0.00 | 170.87 ± 40.14 | 170.87 ± 40.14 | 0.83 ± 0.13 | 2 |
| OutputOperator | 126.89 ± 16.53 | 0.00 ± 0.00 | 126.89 ± 16.53 | 0.63 ± 0.10 | 9 |
| InputOperator | 8.69 ± 1.45 | 0.00 ± 0.00 | 8.69 ± 1.45 | 0.04 ± 0.01 | 1 |
| **Total** | **135.57 ± 17.51** | **20165.99 ± 1963.70**| **20301.56 ± 1965.88**| | **113** |

### YOLOv8

| OpType | CPU Time (μs) | NPU Time (μs) | Total Time (μs) | Time Ratio (%) | Number of Times Called |
|-----------------|---------------------|----------------------|-----------------------|---------------------|-----------------------|
| ConvExSwish | 0.00 ± 0.00 | 13017.04 ± 1165.76 | 13017.04 ± 1165.76 | 75.66 ± 1.96 | 57 |
| Concat | 0.00 ± 0.00 | 1489.94 ± 257.22 | 1489.94 ± 257.22 | 8.58 ± 0.53 | 13 |
| Split | 0.00 ± 0.00 | 681.47 ± 166.62 | 681.47 ± 166.62 | 3.89 ± 0.53 | 8 |
| ConvSigmoid | 0.00 ± 0.00 | 596.08 ± 75.01 | 596.08 ± 75.01 | 3.45 ± 0.18 | 3 |
| Add | 0.00 ± 0.00 | 443.60 ± 118.05 | 443.60 ± 118.05 | 2.53 ± 0.41 | 6 |
| Conv | 0.00 ± 0.00 | 269.61 ± 78.65 | 269.61 ± 78.65 | 1.54 ± 0.30 | 3 |
| Resize | 0.00 ± 0.00 | 236.79 ± 37.74 | 236.79 ± 37.74 | 1.37 ± 0.08 | 2 |
| ConvClip | 0.00 ± 0.00 | 231.82 ± 68.44 | 231.82 ± 68.44 | 1.32 ± 0.27 | 3 |
| MaxPool | 0.00 ± 0.00 | 156.85 ± 56.94 | 156.85 ± 56.94 | 0.89 ± 0.23 | 3 |
| OutputOperator | 124.86 ± 20.74 | 0.00 ± 0.00 | 124.86 ± 20.74 | 0.73 ± 0.15 | 9 |
| InputOperator | 8.47 ± 1.66 | 0.00 ± 0.00 | 8.47 ± 1.66 | 0.05 ± 0.01 | 1 |
| **Total** | **133.33 ± 21.95** | **17123.19 ± 1985.72**| **17256.52 ± 1986.77** | | **108** |

---

### YOLOv11

| OpType | CPU Time (μs) | NPU Time (μs) | Total Time (μs) | Time Ratio (%) | Number of Times Called |
|-----------------|---------------------|----------------------|-----------------------|---------------------|-----------------------|
| ConvExSwish | 0.00 ± 0.00 | 16034.00 ± 1331.95 | 16034.00 ± 1331.95 | 69.90 ± 1.55 | 77 |
| Concat | 0.00 ± 0.00 | 1888.89 ± 293.99 | 1888.89 ± 293.99 | 8.17 ± 0.51 | 17 |
| exSDPAttention | 0.00 ± 0.00 | 1210.88 ± 17.73 | 1210.88 ± 17.73 | 5.32 ± 0.52 | 1 |
| Split | 0.00 ± 0.00 | 908.30 ± 183.92 | 908.30 ± 183.92 | 3.91 ± 0.45 | 10 |
| Add | 0.00 ± 0.00 | 871.64 ± 212.79 | 871.64 ± 212.79 | 3.73 ± 0.60 | 12 |
| ConvSigmoid | 0.00 ± 0.00 | 617.61 ± 59.61 | 617.61 ± 59.61 | 2.69 ± 0.16 | 3 |
| Conv | 0.00 ± 0.00 | 419.72 ± 89.88 | 419.72 ± 89.88 | 1.80 ± 0.24 | 5 |
| Resize | 0.00 ± 0.00 | 272.09 ± 49.91 | 272.09 ± 49.91 | 1.18 ± 0.12 | 2 |
| ConvClip | 0.00 ± 0.00 | 260.08 ± 59.12 | 260.08 ± 59.12 | 1.12 ± 0.18 | 3 |
| MaxPool | 0.00 ± 0.00 | 181.93 ± 53.32 | 181.93 ± 53.32 | 0.78 ± 0.18 | 3 |
| OutputOperator | 131.48 ± 22.93 | 0.00 ± 0.00 | 131.48 ± 22.93 | 0.58 ± 0.12 | 9 |
| ConvAdd | 0.00 ± 0.00 | 126.79 ± 35.28 | 126.79 ± 35.28 | 0.54 ± 0.11 | 2 |
| Reshape | 0.00 ± 0.00 | 56.61 ± 18.03 | 56.61 ± 18.03 | 0.24 ± 0.06 | 3 |
| InputOperator | 8.66 ± 1.59 | 0.00 ± 0.00 | 8.66 ± 1.59 | 0.04 ± 0.01 | 1 |
| **Total** | **140.14 ± 24.26** | **22848.54 ± 2351.95**| **22988.68 ± 2355.97**| | **148** |


## Model Summary and Accuracy Metrics

The table below summarizes the mean average precision (mAP) and total inference time for each model. These metrics provide a high-level view of how each model performs in terms of both detection accuracy and runtime efficiency.

### Mean Average Precision (mAP) by Model

| Metric | YOLOv5 | YOLOv5u | YOLOv8 | YOLOv11 |
|--------|------------|------------|------------|------------|
| **mAP** | 0.2243 | 0.2745 | 0.3051 | 0.3251 |
| **mAP50** | 0.3538 | 0.3834 | 0.4145 | 0.4406 |
| **mAP75** | 0.2432 | 0.2997 | 0.3349 | 0.3568 |
| **mAP85** | 0.3054 | 0.3472 | 0.3867 | 0.4068 |
| **mAP95** | 0.3708 | 0.4822 | 0.5483 | 0.5858 |

### Model Execution Time and Call Frequency

| Model | Total Time (μs) | Number of Processing Calls |
|---------|------------------------|----------------------------|
| **YOLOv5** | 15034.09 ± 1734.28 | 93 |
| **YOLOv5u** | 20301.56 ± 1965.88 | 113 |
| **YOLOv8** | 17256.52 ± 1986.77 | 108 |
| **YOLOv11** | 22988.68 ± 2355.97 | 148 |

## Conclusion

The benchmark reveals a clear performance trade-off between inference time and detection accuracy:

- **YOLOv5** is the fastest model with the lowest total inference time, making it well-suited for situations where speed is more important than high detection precision.

- **YOLOv11** achieves the highest accuracy (mAP) across all IoU thresholds but comes with the longest inference time, which may limit its use in real-time applications.

- **YOLOv8** offers a strong balance between speed and accuracy, making it a practical choice when both factors matter.

- **YOLOv5u** improves accuracy compared to YOLOv5 but falls behind YOLOv8 in both speed and detection quality.

When choosing a model for edge devices like the Orange Pi 5, it’s important to weigh how much latency your system can tolerate versus how much accuracy you need. A faster model may give quicker results, while a more accurate one may offer better detection reliability, but at the cost of speed.
1 change: 1 addition & 0 deletions docs/source/index.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,6 +127,7 @@ docs/troubleshooting/index
docs/additional-resources/best-practices
docs/additional-resources/config
docs/additional-resources/nt-api
docs/benchmarks/index
docs/contributing/index
```

Expand Down
Loading