Image classification
ImageNet validation set · 50,000 images · Top-1. 1,647 Embedl builds of 665 models on 5 devices: which are the best trade-offs between speed and accuracy, and which you can run now.
Latency against Top-1
- Best trade-offs
- Other builds
Every build as a list
Try the best trade-offs on our device farm
Builds on this frontier with a device online that can take them now: the same run as on the model page.
The best trade-offs on every device
No other build here is both faster and more accurate than these. Fastest first; each row opens its build. Some builds' accuracy was measured on the L4 engine.
| # | Model | Device | Build | Latency | Top-1 | Speed-up |
|---|---|---|---|---|---|---|
| 1 | MobileNetV1 (100h) | L4 | Embedl-optimized | 0.085 ms | 74.02% | 1.34× faster |
| 2 | MobileNetV1 1.25× | L4 | Embedl-optimized | 0.100 ms | 75.01% | 1.35× faster |
| 3 | timm/gernet_s.idstcv_in1k | L4 | Embedl-optimized | 0.141 ms | 75.63% | 1.37× faster |
| 4 | MobileNetV2 1.4× | L4 | Embedl-optimized | 0.159 ms | 75.84% | 1.21× faster |
| 5 | EfficientNet-ES | L4 | Embedl-optimized | 0.177 ms | 76.73% | 1.48× faster |
| 6 | timm/gernet_m.idstcv_in1k | L4 | Embedl-optimized | 0.183 ms | 79.83% | 1.62× faster |
| 7 | ResNet-50-D | L4 | Embedl-optimized | 0.324 ms | 80.27% | 1.50× faster |
| 8 | MobileNetV4 Conv Large | L4 | Embedl-optimized | 0.362 ms | 80.65% | 1.71× faster |
| 9 | EfficientNet-Lite4 (TF) | L4 | Embedl-optimized | 0.411 ms | 81.42% | 1.77× faster |
| 10 | PiT-S (224, distilled) | AGX Thor | Embedl-optimized | 0.545 ms | 81.86% | 1.26× faster |
| 11 | ViT Little/16 (256, reg1, GAP) | L4 | Embedl-optimized | 0.559 ms | 82.96% | 1.25× faster |
| 12 | EdgeNeXt Base | L4 | Embedl-optimized | 0.749 ms | 83.24% | — |
| 13 | FlexiViT Base | L4 | Embedl-optimized | 0.979 ms | 83.35% | 1.37× faster |
| 14 | ViT Betwixt/16 (256, reg4, GAP) | L4 | Embedl-optimized | 0.986 ms | 83.68% | 1.33× faster |
| 15 | PiT-B (224, distilled) | AGX Thor | Embedl-optimized | 1.047 ms | 83.93% | 1.24× faster |
| 16 | EVA-02 Small/14 (336, MIM) | AGX Thor | Embedl-optimized | 1.093 ms | 83.95% | — |
| 17 | HGNetV2-B5 | L4 | Embedl-optimized | 1.150 ms | 84.21% | — |
| 18 | ViT Medium-D/16 (256, reg4, GAP) | L4 | Embedl-optimized | 1.232 ms | 85.93% | 1.20× faster |
| 19 | Swin Large/4 (224, window 7) | L4 | Embedl-optimized | 2.471 ms | 86.27% | 1.26× faster |
| 20 | ConvNeXt Large (mlp, CLIP, IN-1k fine-tuned) | AGX Thor | Embedl-optimized | 2.678 ms | 87.07% | 1.21× faster |
| 21 | EVA-02 Base/14 (448, MIM) | AGX Thor | Embedl-optimized | 3.335 ms | 87.88% | — |
| 22 | EVA Large/14 (336) | AGX Thor | Embedl-optimized | 4.052 ms | 88.15% | 1.14× faster |
| 23 | ConvNeXt XXLarge (CLIP, soup, IN-1k fine-tuned) | AGX Thor | Embedl-optimized | 6.683 ms | 88.18% | 1.47× faster |
| 24 | EVA Giant/14 (224, CLIP, IN-1k fine-tuned) | AGX Thor | Embedl-optimized | 8.455 ms | 88.19% | 1.31× faster |
| 25 | EVA-02 Large/14 (448, MIM) | AGX Thor | Embedl-optimized | 10.638 ms | 89.39% | 1.17× faster |