---
title: "NVIDIA's DALI 2.1 Cuts Training Time by 72% Moving Data Prep to GPU"
subtitle: "NVIDIA DALI v2.1 brings a new imperative Dynamic Mode API, torchvision compatibility, and CUDA 13 support to GPU-accelerated data pipelines"
url: "https://alphasignal.ai/news/nvidia-s-dali-2-1-cuts-training-time-by-72-moving-data-prep-to-gpu"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "NVIDIA AI Developer"
company_url: "https://alphasignal.ai/company/nvidiaaidev"
published: "2026-06-16T13:41:18.000Z"
updated: "2026-06-20T08:32:00.015Z"
type: "Repo"
topics:
  - "Training Infra"
  - "Open Source"
subtopics:
  - "Distributed Training"
upvotes: 3382
source_url: "https://github.com/NVIDIA/DALI"
comments: 0
truncated: true
---

# NVIDIA's DALI 2.1 Cuts Training Time by 72% Moving Data Prep to GPU

> NVIDIA DALI v2.1 brings a new imperative Dynamic Mode API, torchvision compatibility, and CUDA 13 support to GPU-accelerated data pipelines

[NVIDIA DALI](https://github.com/NVIDIA/DALI) (Data Loading Library) has shipped version 2.1, the latest update to its GPU-accelerated data preprocessing library. The release formalizes a new imperative programming model called Dynamic Mode, adds a torchvision-compatible API, upgrades thread management, and extends support to CUDA 13.2\. If your training runs are GPU-starved because your CPU can't feed data fast enough, this is the library worth knowing.

## The Problem That Won't Go Away

GPU compute speeds have increased more than 15x from 2018 to 2023 (from NVIDIA V100 to H100), while CPU compute speeds have only grown by roughly 2.23x over the same period. That gap is the core tension DALI was built to address. DALI tackles the CPU bottleneck by offloading data preprocessing to the GPU, relying on its own execution engine built to maximize the throughput of the input pipeline. 

When deploying models in production, preprocessing often becomes the bottleneck. While TensorRT can run model inference in just a few milliseconds, CPU-based preprocessing (resize, pad, normalize) can take 2-10ms per image, especially at high resolutions. NVIDIA DALI solves this by moving the entire preprocessing pipeline to the GPU. The gains are real: benchmarks on SageMaker showed training time improvements of 72%, 37%, and 43% for ResNet18, ResNet50, and ResNet152 respectively. 

## Two APIs, One Library

DALI now ships with two distinct programming models. Understanding the difference matters for how you integrate it into your workflow.

The original **Pipeline Mode** uses a declarative, graph-based API. You define a pipeline as a computation graph using the `@pipeline_def` decorator, compile it once, and then run it repeatedly. DALI optimizes the graph at compile time, enabling aggressive prefetching and overlap between CPU and GPU work. This is the right choice for stable, high-throughput training loops where the pipeline shape doesn't change between batches.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [NVIDIA/DALI GitHub Repository](https://github.com/NVIDIA/DALI)
- [NVIDIA DALI Documentation](https://docs.nvidia.com/deeplearning/dali/user-guide/docs/index.html)
- [DALI v2.1.1 Release Notes](https://docs.nvidia.com/deeplearning/dali/release-notes/index.html)
- [GPU-Accelerated Preprocessing with NVIDIA DALI — Ultralytics Docs](https://docs.ultralytics.com/guides/nvidia-dali)
- [Accelerate computer vision training with NVIDIA DALI on Amazon SageMaker](https://aws.amazon.com/blogs/machine-learning/accelerate-computer-vision-training-using-gpu-preprocessing-with-nvidia-dali-on-amazon-sagemaker/)