Performance

Training Performance Results

We measured the training throughput of the NSFW model on a single DGX A100 node. The table below shows the performance results.

  • NVIDIA DGX A100 (8 x A100 80GB for CLIP L/14 base model)

Model

Metric

Nodes

1

NSFW L/14

Samples per Second

1680

Inference Performance Results

Because the NSFW model is built on top of pretrained CLIP with a small linear probing layer, inference performance is expected to match CLIP inference performance. See CLIP Inference Performance Results for details.