NVIDIA DALI
Type: Technology Tags: CUDA, NVIDIA, GPU, Data Loading, Image Processing, Deep Learning, Open Source, Training Related: cuDNN, TensorRT, nvComp, Megatron-Energon, CV-CUDA, nvImageCodec Sources: NVIDIA official documentation Last Updated: 2026-04-09
Summary
NVIDIA DALI (Data Loading Library) is an open-source, portable library for GPU-accelerated decoding and augmentation of images, videos, and audio, designed to eliminate the data preprocessing bottleneck in deep learning training pipelines. As a drop-in replacement for built-in data loaders in PyTorch, TensorFlow, and MXNet, DALI overlaps data preprocessing with model training to maximize GPU utilization.
Detail
Purpose
Deep learning training pipelines often stall waiting for preprocessed batches from CPU-bound data loaders (decode, crop, resize, normalize). DALI moves these preprocessing operations to the GPU, overlapping data preparation with model computation so the GPU never sits idle waiting for data.
Key Features
- GPU-accelerated image, video, 3D volume, and audio preprocessing
- Drop-in replacement for PyTorch DataLoader, TensorFlow data pipeline, and MXNet data iterator
- Python APIs for rapid pipeline construction with transparent multi-GPU scaling
- Flexible graph-based pipeline architecture supporting custom operators
- Extensive format support: JPEG, PNG, TIFF, JPEG2K, BMP, WebP, LMDB, TFRecord, COCO, WAV, FLAC, OGG, H.264, HEVC
- Optical flow computation via NVIDIA Optical Flow SDK integration
- Super-resolution and video-specific operations
- Compatible with TensorRT, Triton Inference Server, TAO Toolkit, NVDEC
Use Cases
- Accelerating image classification, detection, and segmentation training
- Video model training (action recognition, object tracking)
- Audio model training (speech recognition, sound classification)
- Medical imaging data pipelines (DICOM, TIFF volumes)
- Multi-modal training pipelines combining image, video, and text
Hardware Requirements
- NVIDIA GPU with CUDA support
- NVDEC-capable GPU for hardware video decode (Volta+ preferred)
- Optical flow requires Turing, Ampere, or Ada architecture
- Available on Linux (primary) and Windows
Language Bindings
- Python (primary API)
- C++ (underlying pipeline implementation)
Connections
- CV-CUDA — CV-CUDA handles post-processing for inference; DALI handles pre-processing for training
- nvComp — DALI can decompress compressed image/video data via nvComp integration
- Megatron-Energon - complementary multimodal data loader for Megatron-scale WebDataset/JSONL data pipelines.
- nvImageCodec — nvImageCodec provides codec plugins that DALI uses for image decode
- TensorRT — DALI preprocessed data feeds into TensorRT inference pipelines
- cuDNN — cuDNN consumes DALI-preprocessed tensors for model execution