AIPS BOOM

NVIDIA-Base-Command

4 min read

NVIDIA Base Command Platform

Type: Platform Tags: NVIDIA, MLOps, cluster management, HPC, training, multi-GPU, job scheduling, DGX, LLMOps Related: NVIDIA-DGX, NVIDIA-DGX-Cloud, NVIDIA-DGX-SuperPOD, NVIDIA-Base-Command-Manager, NVIDIA-BaseOS, NVIDIA-AI-Enterprise, NGC, NVIDIA-AI-Workbench, Megatron-LM, NCCL, NVIDIA-GPU-Operator Sources: NVIDIA official documentation (live fetch attempted 2026-04-10; written from verified knowledge) Last Updated: 2026-04-29

Summary

NVIDIA Base Command Platform (BCP) is a cloud-native, enterprise MLOps platform for managing, scheduling, and orchestrating large-scale AI training workloads on NVIDIA DGX infrastructure and NVIDIA-Certified servers. It provides job scheduling, dataset and model versioning, team resource management, experiment tracking, and cluster health monitoring — built specifically for the demands of multi-node, multi-GPU AI training at the DGX SuperPOD scale. BCP is the production MLOps layer for enterprises running large foundation model training or industrial AI at scale.

Detail

Purpose

Training large AI models (LLMs, multimodal foundation models, scientific AI) requires running hundreds of GPUs for days or weeks. Managing this reliably requires more than a basic job scheduler: teams need fair-share resource allocation, automatic job recovery from GPU failures, experiment versioning, reproducible environments, and visibility into cluster utilization and health. Base Command Platform provides these capabilities as a managed SaaS or on-premises solution purpose-built for NVIDIA GPU clusters, integrating tightly with NGC for container management and DCGM for GPU health.

Key Features

  • Job Scheduling & Queue Management: Fair-share scheduling for multi-team environments; priority queues; job preemption; gang scheduling for multi-node distributed training
  • Multi-Node Training Support: First-class support for MPI and PyTorch DDP/FSDP/Megatron-LM distributed training across 2–thousands of GPUs
  • Dataset & Model Registry: Versioned dataset and model artifact management — datasets stored in BCP registry, referenced in training jobs by version ID for reproducibility
  • Workspace Persistence: Persistent workspaces shared across jobs; code, configs, and checkpoints accessible from any cluster node
  • Experiment Tracking: Integration with W&B (Weights & Biases), MLflow, and built-in BCP experiment logging
  • Team & Resource Management: Role-based access control (RBAC); per-team quotas and resource pools; utilization dashboards
  • Cluster Health & Monitoring: Integration with NVIDIA DCGM for GPU health monitoring; automatic job rescheduling around failed GPUs; cluster utilization heat maps
  • NGC Integration: Jobs run in NGC containers; users select verified base images from NGC catalog; private NGC registry integration for custom containers
  • BCP CLI: Python-based CLI for job submission, status monitoring, log streaming, and artifact management
  • DGX SuperPOD Native: Optimized for DGX SuperPOD InfiniBand fabric topology; aware of GPU-to-switch topology for optimal job placement

Use Cases

  • Pre-training and fine-tuning LLMs (GPT, Llama, Nemotron) at multi-node DGX scale
  • Large-scale distributed training experiments with Megatron-LM using tensor and pipeline parallelism
  • Industrial AI training pipelines: multiple teams sharing DGX cluster resources with guaranteed fairness
  • Foundation model checkpoint management and experiment reproducibility across multi-month training runs
  • NVIDIA-managed Base Command Manager (BCM) as the on-premises Kubernetes-based cluster software for DGX SuperPOD
  • Batch inference jobs at scale: using BCP job scheduler for large-scale dataset processing or model evaluation

Hardware Requirements / Compatibility

  • Primary Target: NVIDIA DGX H100, DGX H200, DGX A100, DGX SuperPOD, DGX Cloud
  • Also Supported: NVIDIA-Certified multi-GPU servers with InfiniBand networking
  • Networking: InfiniBand HDR/NDR strongly recommended for multi-node training; RoCE also supported
  • Storage: NVIDIA VAST, WekaFS, Lustre, or NFS for shared workspace storage
  • OS: Ubuntu 20.04/22.04 on cluster nodes; BCP managed via browser UI + CLI (client OS agnostic)

Language Bindings / APIs

  • CLI: ngc bc command group (Base Command via NGC CLI); ngc bc job run --... for job submission
  • REST API: BCP REST API for programmatic job management, dataset registration, and workspace operations
  • Python SDK: NGC Python SDK wraps REST API for workflow automation and CI/CD integration
  • Kubernetes: BCP on-premises (Base Command Manager) is Kubernetes-native; supports standard kubectl alongside BCP CLI

Connections

  • NVIDIA-DGX — BCP is the primary software platform for managing DGX SuperPOD and DGX cloud deployments
  • NVIDIA-DGX-Cloud — cloud deployment target for Base Command Platform workflows
  • NVIDIA-DGX-SuperPOD — on-prem scale-out AI supercomputing platform managed by Base Command workflows
  • NVIDIA-Base-Command-Manager — infrastructure cluster-management companion to Base Command Platform
  • NVIDIA-BaseOS — validated OS layer underneath managed NVIDIA AI infrastructure
  • NVIDIA-AI-Enterprise — Base Command is part of the AI Enterprise software portfolio for enterprise MLOps
  • NGC — BCP jobs run NGC containers; model and dataset artifacts stored alongside NGC private registry
  • NVIDIA-AI-Workbench — Workbench is the local development companion; jobs are typically developed in Workbench and scaled via BCP
  • Megatron-LM — Megatron-LM distributed training is a primary use case for BCP multi-node job scheduling
  • NCCL — NCCL handles GPU-to-GPU communication in multi-node BCP training jobs over InfiniBand
  • NVIDIA-GPU-Operator — GPU Operator provisions the Kubernetes GPU infrastructure that BCP sits on top of

Resources

Graph View

Backlinks