AIPS BOOM

NGC

6 min read

NGC (NVIDIA GPU Cloud)

Type: Platform Tags: NVIDIA, GPU, containers, model registry, software catalog, cloud, MLOps, NGC Related: NVIDIA-AI-Enterprise, NVIDIA-AI-Enterprise-Quick-Start-Guide, NVIDIA-AI-Enterprise-Software, NVIDIA-NIM, NVIDIA-NGC-Catalog, NVIDIA-Optimized-Frameworks, PyG, NVIDIA-DGL, NVIDIA-RAPIDS, NVIDIA-Merlin, NVIDIA-TAO, Nemotron, NeMo-Platform, NVIDIA-BioNeMo, BioNeMo-Recipes, Transformer-Engine, NVIDIA-Dynamo, Triton-Inference-Server, NVIDIA-NeMo, TensorRT, NVIDIA-HPC-SDK, cuQuantum-Appliance, Nsight-Cloud, NVIDIA-DGX-Spark, NVIDIA-DGX-Station, NVIDIA-Base-Command, NVIDIA-GPU-Operator Sources: NVIDIA official documentation (live fetch attempted 2026-04-10; written from verified knowledge), https://docs.nvidia.com/deeplearning/frameworks/pyg-release-notes/running.html, https://docs.nvidia.com/deeplearning/frameworks/dgl-release-notes/index.html, https://developer.nvidia.com/topics/ai/data-science/cuda-x-data-science-libraries Last Updated: 2026-04-30

Summary

NGC (NVIDIA GPU Cloud) is NVIDIA’s centralized hub for GPU-optimized software, providing a curated catalog of pre-built containers, pre-trained AI models, Helm charts, and SDKs. It eliminates the friction of configuring GPU computing environments by delivering containers that are NVIDIA-tested, CVE-scanned, and refreshed monthly across the full CUDA software stack. NGC serves both individual developers and enterprises, including as the primary distribution channel for NVIDIA AI Enterprise software.

Detail

Purpose

Setting up a functional GPU computing environment — with correct CUDA versions, cuDNN, framework builds, and dependencies — is time-consuming and error-prone. NGC solves this by providing a curated catalog of pre-built, NVIDIA-tested containers and models that work out of the box on any NVIDIA GPU or cloud instance, dramatically reducing time-to-experiment and time-to-production. For enterprises, NGC also acts as the licensing and distribution layer for NVIDIA AI Enterprise software.

Key Features

  • Container Registry: GPU-optimized Docker containers for PyTorch, PyG, legacy NVIDIA-DGL, TensorFlow, JAX, TensorRT, NeMo, Triton, NVIDIA-RAPIDS, cuQuantum-Appliance, and dozens of other frameworks, hosted at nvcr.io
  • Model Catalog: Pre-trained AI models spanning NLP, computer vision, speech, medical imaging, biology (protein structure, genomics), and generative AI
  • Helm Charts: Kubernetes-ready deployment charts for NVIDIA platforms (Triton, Riva, Metropolis, etc.)
  • Software SDKs: Direct downloads of NVIDIA SDKs and toolkits (cuDNN, TensorRT, Nsight, CUDA Toolkit, etc.)
  • Security Scanning: All containers are regularly scanned for CVEs and cryptographically signed by NVIDIA
  • Monthly Container Updates: Refreshed monthly with the latest CUDA, cuDNN, and upstream framework versions (e.g., 23.10-py3 versioning scheme)
  • NGC Private Registry: Enterprise customers can maintain private registries co-located with the public NGC catalog for custom models and containers
  • NGC CLI: ngc command-line tool for pulling containers, models, datasets, and managing registry credentials programmatically
  • Cloud Integration: Available natively on AWS Marketplace, GCP Marketplace, Azure Marketplace, and NVIDIA LaunchPad
  • NIM Catalog: As of 2024, NGC hosts the full NVIDIA NIM microservice catalog for one-click LLM and AI model deployment
  • Collections: Curated groupings of containers, models, and Helm charts for specific use cases (e.g., “LLM Inference Collection,” “Medical Imaging Collection”)

Use Cases

  • Rapidly bootstrapping deep learning training and inference environments without manual CUDA/cuDNN installation
  • Pulling NVIDIA-Optimized-Frameworks containers for versioned PyTorch, PyG, TensorFlow, JAX, CUDA Deep Learning, and related framework environments
  • Deploying production inference servers with verified, optimized Triton Inference Server containers
  • Accessing pre-trained foundation models (LLMs, vision-language models, protein structure models) for fine-tuning or deployment
  • Pulling BioNeMo-related containers, model artifacts, and NIM assets for life-sciences model training and inference workflows
  • Enterprise software distribution — NVIDIA AI Enterprise software delivered and licensed via NGC
  • Reproducible research and MLOps pipelines using pinned NGC container versions
  • Running GPU workloads on DGX, cloud GPU instances (A100, H100, H200, Blackwell B100/B200), or NVIDIA-certified servers
  • Deploying NIM inference microservices via Helm on Kubernetes

Hardware Requirements / Compatibility

  • All modern NVIDIA data center GPUs: V100, A100, H100, H200, B100, B200, GH200
  • NVIDIA RTX desktop and laptop GPUs (Turing, Ampere, Ada Lovelace, Blackwell)
  • CUDA 10.x and newer (container-specific; most current containers require CUDA 11.8+ or 12.x)
  • Compatible with NVIDIA GPU Operator for Kubernetes-based deployment pipelines
  • NGC Private Registry deployable on-premises for DGX systems and SuperPODs

Language Bindings / APIs

  • NGC CLI: Python-based command-line tool (pip install ngc-cli); supports ngc registry image pull, ngc registry model download, etc.
  • REST API: Full programmatic access to catalog browsing, artifact downloads, and private registry management
  • Docker: Standard docker pull nvcr.io/<org>/<image>:<tag> workflow
  • Helm: Kubernetes deployment via helm install with NGC-hosted charts using helm repo add ngc-stable
  • Python SDK: Integration with MLOps platforms (MLflow, Weights & Biases, Kubeflow)

Connections

  • NVIDIA-AI-Enterprise — AI Enterprise software is licensed, distributed, and updated exclusively through NGC
  • NVIDIA-AI-Enterprise-Quick-Start-Guide — first-run path for enterprise account setup, NGC sign-in, API key use, and initial container access.
  • NVIDIA-AI-Enterprise-Software — current AI Enterprise software catalog maps supported application and infrastructure components to NGC and documentation.
  • NVIDIA-NIM — All NIM containers are hosted in the NGC catalog and deployed via NGC credentials
  • NVIDIA-NGC-Catalog — public catalog surface for containers, models, Helm charts, and SDKs
  • NVIDIA-Optimized-Frameworks — deep learning framework containers are distributed through NGC and documented through NVIDIA’s framework support matrix.
  • PyG — NVIDIA PyG containers use the nvcr.io/nvidia/pyg:<xx.xx>-py3 image pattern and are distributed through the NGC container registry path.
  • NVIDIA-DGL — legacy NVIDIA DGL container release notes point to NGC security scanning and now redirect new GNN work toward PyG.
  • NVIDIA-RAPIDS — RAPIDS containers and packages are part of the NVIDIA accelerated data science distribution story.
  • NVIDIA-Merlin — Merlin recommender workflows historically used NGC containers for training, inference, and HugeCTR/NVTabular components.
  • NVIDIA-TAO — TAO containers, pretrained CV models, and model artifacts are distributed through NGC.
  • cuQuantum-Appliance - NGC-distributed container workflow for Qiskit/Cirq quantum simulation on NVIDIA GPUs.
  • NVIDIA-HPC-SDK - current HPC SDK docs include NGC/containerization as a distribution path.
  • Nemotron — NVIDIA model family whose downloadable and API-facing artifacts appear through NVIDIA model catalogs
  • NeMo-Platform — NeMo microservices rely on NGC credentials and NVIDIA-hosted artifacts
  • NVIDIA-BioNeMo and BioNeMo-Recipes — BioNeMo models, containers, recipes, and NIM assets sit in the same NGC-centered distribution pattern.
  • Transformer-Engine — current installation docs note that TE is preinstalled in NVIDIA PyTorch containers from NGC.
  • NVIDIA-Dynamo — inference-serving stack that fits the NGC/NIM deployment ecosystem
  • Triton-Inference-Server — Official Triton containers updated monthly on NGC; primary distribution channel
  • NVIDIA-NeMo — NeMo framework containers and checkpoint models (GPT, BERT, Llama variants) hosted on NGC
  • TensorRT — TensorRT containers, model optimization pipelines, and ONNX model zoo distributed via NGC
  • Nsight-Cloud - Nsight Cloud collections, containers, and Helm charts are distributed through NGC for cloud-native profiling workflows.
  • NVIDIA-DGX-Spark - DGX Spark user guide includes NGC integration for local AI development.
  • NVIDIA-DGX-Station - DGX Station’s preconfigured software stack is adjacent to NGC containers and AI software.
  • NVIDIA-Base-Command — Base Command Platform uses NGC containers and credentials for multi-node job scheduling
  • NVIDIA-GPU-Operator — GPU Operator pulls NVIDIA driver, plugin, and toolkit containers from NGC (nvcr.io/nvidia/)
  • NVIDIA-Riva — Riva speech AI containers and pre-trained acoustic/language models distributed via NGC

Resources

Graph View

Backlinks