NVIDIA Open-Sources NVCF for GPU-Accelerated AI Workloads

Key takeaway

• NVIDIA has open-sourced NVCF, a platform that deploys and scales GPU-accelerated AI workloads across multiple clusters and regions.

• The platform provides unified control, load-balanced routing, multi-cluster autoscaling, and support for mixed GPU types—features that allow developers to manage compute-heavy AI tasks without building custom infrastructure.

• The code is now publicly available for inspection, modification, and community contribution.

3 Key Points

FAQ

What types of workloads does NVCF support?

NVCF supports long-running, invokable workloads called functions (for inference, streaming, and service-style GPU workflows) and asynchronous, run-to-completion workloads called tasks (for batch inference, evaluation, fine-tuning, and data preparation). Both can be packaged as containers or Helm charts.

How do I get started with NVCF?

After installing a self-managed NVCF deployment and configuring the CLI, you initialize a project with nvcf-cli init, generate an API key, create a function from a configuration file, deploy it, and invoke it via command line. Full setup and quickstart guides are in the docs/user/ directory.

What is the architecture of NVCF?

NVCF runs as Kubernetes services with three main planes: a control plane that manages function state and secrets, an invocation plane that routes HTTP, streaming, and gRPC requests, and a compute plane that integrates GPU clusters through the NVIDIA Cluster Agent (NVCA). Observability and telemetry help operators monitor health and debug workloads.