NVIDIA FLARE Roadmap

This page outlines planned features and target release milestones for upcoming NVIDIA FLARE versions. Dates and features are subject to change.

Note

This roadmap reflects current planning and is provided for informational purposes. Feature scope and release timing may shift as development progresses.

FLARE 2.9.1 — Target: Q4 2026

Confidential Federated AI Support

Building on existing support for AMD SEV-SNP CPU CVMs with NVIDIA GPUs:

  • Intel TDX CPU support for CPU-based confidential computing workloads

  • CoCo (Confidential Containers) support for container-level confidential execution

FLARE 2.10.0 — Target: Early 2027

Possible Features

  • Documentation and Tutorial Transformation — Redesign the onboarding path so a first-time user runs a real federation before seeing a feature list, and bring the 100+ notebook self-paced course and research-project guides up to a verified, reproducible standard.

  • Better GPU resource utilization — Release GPUs during aggregation, barriers, and downloads instead of holding them idle. A durable, GPU-free site supervisor keeps a job’s identity alive across time-bounded Slurm allocations via checkpoint-and-resume while avoiding or minimizing changes to the federated training algorithm.

  • Job Statistics — A usage proxy for cost: FLARE has no access to billing rates, so the goal is to report compute hours, GPU hours, storage volume/time, and ingress/egress size per job, so real cost can be mapped in later once rates are supplied.

  • Operation and Security — Federated identity via OpenID Connect (sites federate through their existing enterprise IdP) plus short-lived, job-scoped ephemeral keys, so a leaked credential’s blast radius is limited to one job and one window.

  • Audit log enhancements — Still a brainstorm, not yet scoped. Today’s per-site audit log (unchanged since FLARE 2.2.1) captures event ID, user identity, action name, and job ID as one flat text line; possible gaps include artifact-level detail, site identity, and a structured, queryable format.

  • Memorization in LLMs — Understanding what a federated LLM training run memorizes from participant data, a prerequisite for defensible privacy claims about federated training itself.

  • FedRevive (Async FL Algorithm) — An asynchronous federated learning algorithm where participants contribute without waiting on a synchronous round barrier.