Meta FAIR's e-means beats k-means in one-tenth the epochs with 10-45x GPU speedups
NielsRogge · x · 2026-11-12
Meta FAIR (including Hugo Touvron, Matthijs Douze and Hervé Jégou) released the paper "e-means: outperforming k-means (in one-tenth the epochs)".
- Core method: an exponential moving average optimization procedure for k-means with mini-batch updates, powered by modern deep learning machinery.
- Establishes scaling laws for decay rate and batch size, giving a formulaic rule to fix hyperparameters with no sweeps needed.
- Results: matches Lloyd k-means at 25 epochs in just 1.4–2.6 epochs across embedding datasets, delivering 10–45x GPU wall-clock speedups over state-of-the-art implementations, often beating k-means++.
- Scale test: partitions DINO10B — 10 billion DINO embeddings — into 16 million clusters in only 2.5 epochs.
Relevant for large-scale clustering in data analysis, retrieval, quantization and data curation.
Related event: Meta FAIR Open-Sources e-means, Beating k-means by up to 45x(3 posts)→
More from Infra
- Hornet launches with thesis: retrieval infrastructure is mandatory for the agent era — jobergum · 2026-11-12
- David Linthicum: Private Cloud Is Back and Cloud-First Thinking Is Ending — DavidLinthicum · 2026-11-12
- On-prem models can cost 10-20x less than cloud AI, says enterprise architect — DavidLinthicum · 2026-11-12
- Evercore downgrades Nvidia, slashes price target from $465 to $186 — asymco · 2026-11-12
- SpaceX's Southaven data centers net local schools and county $63M a year — 0xAllen_ · 2026-11-12
- Quantum infrastructure firm Q-CTRL announces Series C led by Salesforce Ventures — MJBiercuk · 2026-11-12