Embedl's build embedl/dinov3-vitb16 of the original model facebook/dinov3-vitb16-pretrain-lvd1689m
Embedl optimized · TensorRT
- embedl/dinov3-vitb16
- INT8
- AGX Orin
- Labels
- Image
- Build
- Embedl-optimized
- Precision
- INT8
- Runtime
- TensorRT 10.16.2.10 with Embedl optimizations
- Output
- 768-dim embedding and patch features
DINOv3 ViT-B/16 is Meta's self-supervised vision backbone: it turns an image into a 768-dimensional embedding and patch features, for search, clustering or a downstream head.
The optimized model (embedl/dinov3-quantized-tensorrt) is under the Embedl Models Community Licence v1.0; the upstream weights (facebook/dinov3-vitb16-pretrain-lvd1689m) are under the DINOv3 License. Both texts ship beside the engine.
Methodology
- Build: The published Embedl Deploy INT8 ONNX (embedl/dinov3-quantized-tensorrt at 4006876933ef), compiled on this device with TensorRT's FP16 and INT8 flags and optimization level 3; nothing re-quantized.
- Latency: CUDA events around one engine execution, batch 1, host-device copies excluded; 1 s warm-up, 5 s measured. Socket time is five requests to the served model, decode, PCA, PNG and JSON included.
- Accuracy: the DINOv3 model card's k-NN protocol (k=20, cosine, τ=0.07) on ImageNette, run on this device with this engine: it embeds the 9,469 train images as its own bank and the 3,925 val images as queries, each through the served model's own preprocessing (shorter side 256, centre 224 crop). It measures whether the embedding geometry survives quantization; ImageNette is close to saturation, so a few tenths of a point is a handful of images.
- Memory: peak RSS of the serving process while it answers requests.
- Size: the serialized TensorRT engine.