# 普通 CNN/Transformer 性能

## 典型模型推理速度

- 本页整理边缘计算芯片上的 INT8 实测 FPS，供选型和复测时参考。数据不等同于具体项目的交付性能；
- 复测时请同时确认工具链版本、核数、NPU频率、DDR带宽等影响因素；
- 其他友商若用本章数据进行PK，请说明出处以便溯源。

### AX8850N

| 模型 | 输入分辨率 | FPS@1核 | FPS@3核 | CMM(MB)@1核 | CMM(MB)@3核 |
| --- | --- | --- | --- | --- | --- |
| Inceptionv1 | 224 | 857.63 | 2020.20 | - | - |
| Inceptionv3 | 299 | 228.83 | 598.80 | - | - |
| MobileNetv1 | 224 | 1763.67 | 3610.11 | - | - |
| MobileNetv2 | 224 | 1680.67 | 3546.10 | - | - |
| SqueezeNetv11 | 227 | 2123.14 | 5000.00 | - | - |
| SqueezeNetv10 | 227 | 1106.19 | 2890.17 | - | - |
| ResNet18 | 224 | 788.02 | 1851.85 | - | - |
| ResNet50 | 224 | 348.80 | 749.06 | - | - |
| VGG16 | 224 | 70.75 | 122.84 | - | - |
| YOLOv3 | 608 | 23.15 | 66.74 | - | - |
| YOLOv5s | 640 | 150.20 | 384.62 | - | - |
| YOLOv6s | 640 | 112.20 | 319.90 | - | - |
| YOLOv7-Tiny | 640 | 168.41 | 360.88 | - | - |
| YOLOv7 | 640 | 28.75 | 79.72 | - | - |
| YOLOv8n | 640 | 195.08 | 537.35 | 6.32 | 3.52 |
| YOLOv8s | 640 | 95.95 | 278.40 | 15.00 | 11.28 |
| YOLOv9s | 640 | 89.19 | 239.18 | - | - |
| YOLOv10s | 640 | 109.90 | 303.03 | - | - |
| YOLO11s | 640 | 89.06 | 246.85 | 19.76 | 10.06 |
| YOLO26n | 640 | 296.47 | 725.69 | 5.14 | 3.26 |
| YOLO26s | 640 | 113.70 | 315.86 | 16.90 | 10.2 |
| YOLO26m | 640 | 39.39 | 115.69 | 36.55 | 27.6 |
| YOLO26l | 640 | 30.42 | 89.49 | 37.96 | 33.88 |
| YOLO26x | 640 | 13.95 | 49.01 | 77.82 | 70.4 |
| YOLOX_s | 640 | 103.43 | 293.69 | - | - |
| YOLO-Nas_s | 640 | 82.91 | 234.19 | - | - |
| PPYOLOE+_s | 640 | 87.42 | 328.41 | - | - |
| Deit_t | 224 | 580.38 | 1282.05 | - | - |
| Swin_t | 224 | 157.88 | 401.28 | - | - |
| ViT_b | 224 | 57.57 | 174.67 | - | - |

### AX8910

| 模型 | 输入分辨率 | FPS | CMM(MB) |
| --- | --- | --- | --- |
| mobilenetv2 | 224 | 1111 | 4.43 |
| resnet50 | 224 | 292 | 25.64 |
| swin_t | 224 | 204 | 29.25 |
| vit_b | 224 | 89 | 83.59 |
| yolov5s | 640 | 141 | 11.76 |
| yolov6s | 640 | 106 | 15.86 |
| yolov8s | 640 | 88 | 16.34 |
| yolov8l | 640 | 19 | 54.86 |
| yolov8x | 640 | 12 | 83.88 |
| yolov9s | 640 | 83 | 13.70 |
| yolov10s | 640 | 97 | 14.45 |
| yolo11s | 640 | 95 | 18.02 |
| yolo26n | 640 | 246 | 5.09 |
| yolo26s | 640 | 92 | 18.27 |
| yoloworldv2 | 640 | 68 | 18.54 |
| yoloworldv2-txt | 4x77 | 52 | 134.50 |
| Depth-Anything-V2 | 518 | 15 | 50.40 |

```{note}
最新评测数据请以官方仓库为准：<https://github.com/AXERA-TECH/ax-samples/tree/main/benchmark>
```

## 能效比分析

> 待补充：功耗测试点明确后，再补充每瓦吞吐和单位功耗下的性能对比。