Examples
YOLO Benchmarks
Examples for a variety of YOLO versions have been provided, the following table shows benchmarks of how each version performs with respect to resource requirements running on the RK3588. Note that detection results vary depending on the YOLO version.
The data/palace.jpg image file was used for benchmarking.
rk3588
Results for the rk3588 with 6 TOPS three core NPU operating at 1Ghz.
| Model | Average Total Time | Inference | Post Processing | Rendering |
|---|---|---|---|---|
| YOLOv5s | 30.2ms | 34.9ms | 0.4ms | 2.2ms |
| YOLOv8s | 39.6ms | 41.5ms | 2.4ms | 2.3ms |
| YOLOv10s | 48.8ms | 47.5ms | 1.3ms | 2.0ms |
| YOLOXs | 38.4ms | 39.4ms | 0.2ms | 2.6ms |
| YOLOv11s | 46.8ms | 49.0ms | 0.9ms | 1.7ms |
| YOLOv8n-pose | 37.8ms | 37.3ms | 0.5ms | 1.4ms |
| YOLOv5s-seg | 108.1ms | 46.0ms | 64.6ms | 4.4ms |
| YOLOv8s-seg | 122.0ms | 53.0ms | 72.4ms | 4.3ms |
rk3576
Results for the rk3576 with 6 TOPS two core NPU operating at 950Mhz.
| Model | Average Total Time | Inference | Post Processing | Rendering |
|---|---|---|---|---|
| YOLOv5s | 29.2ms | 48.3ms | 1.0ms | 3.8ms |
| YOLOv8s | 43.4ms | 60.6ms | 4.1ms | 4.0ms |
| YOLOv10s | 49.2ms | 68.2ms | 2.3ms | 3.5ms |
| YOLOXs | 37.9ms | 56.2ms | 0.6ms | 4.7ms |
| YOLOv11s | 38.4ms | 70.1ms | 2.0ms | 3.4ms |
| YOLOv8n-pose | 33.9ms | 49.5ms | 0.9ms | 2.3ms |
| YOLOv5s-seg | 151.9ms | 65.9ms | 95.7ms | 7.0ms |
| YOLOv8s-seg | 154.2ms | 78.7ms | 126.7ms | 8.2ms |
rk3566
Results for the rk3566 with 1 TOPS single core NPU operating at 900Mhz.
| Model | Average Total Time | Inference | Post Processing | Rendering |
|---|---|---|---|---|
| YOLOv5s | 70.2ms | 70.4ms | 2.4ms | 7.2ms |
| YOLOv8s | 95.5ms | 84.6ms | 10.3ms | 7.6ms |
| YOLOv10s | 122.3ms | 116.3ms | 7.6ms | 8.1ms |
| YOLOXs | 83.0ms | 78.9ms | 1.7ms | 9.8ms |
| YOLOv11s | 117.3ms | 117.6ms | 4.9ms | 7.1ms |
| YOLOv8n-pose | 59.8ms | 62.2ms | 1.6ms | 4.5ms |
| YOLOv5s-seg | 433.64ms | 95.7ms | 349.7ms | 20.0ms |
| YOLOv8s-seg | 420.8ms | 110.7ms | 312.3ms | 19.7ms |
The Inference, Post Processing, and Rendering columns show how processing time is split across the Total Time. These figures are derived from the first run on the benchmark so when totaled together they are slightly higher than the Average Total Time.
The Inference column represents processing on the NPU, Post Processing and Rendering values are performed on the CPU.
 |
|
|---|---|
| YOLOv5s - 21 Objects Detected | YOLOv8s - 22 Objects Detected |
 |
 |
| YOLOv10s - 19 Objects Detected | YOLOXs - 29 Objects Detected |
 |
 |
| YOLOv11s - 19 Objects Detected | YOLOv8n-pose - 9 Objects Detected |
 |
 |
| YOLOv5s-seg - 21 Objects Detected | YOLOv8s-seg - 20 Objects Detected |