A new open-source framework promises faster federated learning without sacrificing flexibility.
Researchers have released FastFederatedLearning (FFL), a federated learning framework built in C/C++ instead of the usual Python-heavy stack. FFL lets developers define any communication graph between clients and servers, rather than locking them into a fixed client-server topology. In benchmarks against Intel's OpenFL framework, FFL delivered speedups ranging from 2.5x to 3.69x across x86-64, ARM-v8, and RISC-V hardware. The team plans to add a Python interface and a middleware layer to support different communication backends.
Federated learning already trades raw speed for privacy, since it trains models across scattered devices instead of pooling data in one place. A framework that claws back performance while also supporting arbitrary network topologies matters for anyone running federations on constrained or mixed hardware, from edge devices to multi-architecture server fleets. The bigger ambition here is dynamic federations, where the relationships between clients and servers shift over time instead of staying fixed.
That 2.5x to 3.69x figure is measured against one incumbent, Intel OpenFL, not the entire federated learning field, so treat it as a floor rather than a universal verdict.