Embedl's build embedl/dinov3-vitb16 of the original model facebook/dinov3-vitb16-pretrain-lvd1689m
Embedl-optimized
- embedl/dinov3-vitb16
- FP8
- AGX Thor
- Labels
- Image
- Build
- Embedl-optimized
- Precision
- FP8
- Runtime
- TensorRT 10.13.3.9 with Embedl optimizations
- Input size
- 224×224
- Output
- 768-dim image embedding
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.
Methodology
- Latency: trtexec GPU compute time of one 224×224 image, batch 1, CUDA graphs; host↔device copies, preprocessing and the k-NN search are not included.
- Accuracy: the DINOv3 model card’s k-NN protocol (k=20, cosine, τ=0.07) on ImageNette: each engine embeds the 9,469 train images as its own bank and the 3,925 val images as queries. It measures whether the embedding geometry survives quantization; ImageNette is close to saturation, so 0.3 points is a handful of images.
- Memory: peak device memory of the engine during the run.
- Size: the serialized TensorRT engine on disk.
- Baseline TensorRT: the same model built FP16 with trtexec on upstream TensorRT, same device.