Hardware Recommendations

We tested LightlyTrain in various setups, the following training examples with hardware setups are provided as a performance reference:

• Distillation YOLOv8l on 2 × NVIDIA RTX 4090

  • method='distillation'

  • batch_size=512

  • model='ultralytics/yolov8l.yaml'

  • Dataset Size: 1 million images

  • GPUs: 2 × NVIDIA RTX 4090

  • GPU Memory: ~16GB per GPU

  • Time per Epoch: ~19 minutes

• Distillation YOLO11x on 4 × NVIDIA H100

  • method='distillation'

  • batch_size=2048

  • model='ultralytics/yolov11x.yaml'

  • Dataset Size: 1 million images

  • GPUs: 4 × NVIDIA H100

  • GPU Memory: ~80GB per GPU

  • Time per Epoch: ~6 minutes

Those setups deliver efficient training while handling large-scale datasets with high throughput. While smaller setups can be used, training times may increase significantly and batch size might need to be reduced to fit GPU memory.