DINOv3 ViT-B/16

facebook/dinov3-vitb16-pretrain-lvd1689m

embedl/dinov3-vitb16
FP16
AGX Orin

Labels

Image

Build

Baseline TensorRT

Precision

FP16

Runtime

TensorRT 10.16.2.10

Output

768-dim embedding and patch features

Run DINOv3 ViT-B/16 on AGX Orin

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: median GPU time per image, batch 1, timed with CUDA events around each run.
Accuracy: k-NN top-1 on Imagenette (k = 20), from the model's own embeddings: 9,469 training images as the reference set, 3,925 validation images classified.
Memory: peak memory while running the model.

Files

Image	Size
Model docker.io/embedl/models:hf.embedl.dinov3-vitb16.fp16.agx-orin.20261102130709a5	165.9 MiB
Runtime docker.io/embedl/models:runtimes.tensorrt.arm64.2026110212523521	703 MiB

Variants

Every build of DINOv3 ViT-B/16 on AGX Orin.

Variant	Precision	Speed	Peak memory	Runtime
Baseline TensorRT This page huggingface_fp16_tensorrt	FP16	3.023 ms Latency: Median GPU time	877 MiB	TensorRT 10.16.2.10 runtimes.tensorrt.arm64.2026110212523521
Embedl-optimized · TensorRT huggingface_int8_embedl_tensorrt	INT8	2.560 ms Latency: Median GPU time	706 MiB	TensorRT 10.16.2.10 with Embedl optimizations runtimes.tensorrt.arm64.2026110212523521