Meta FAIR's e-means beats k-means in 1/10 epochs with 10-45x GPU speedups
burny_tech · x · 2026-11-10
Meta FAIR (authors include Hugo Touvron, Matthijs Douze, Hervé Jégou) published "e-means: outperforming k-means (in one-tenth the epochs)".
- Method: replaces k-means centroid updates with an exponential moving average optimization over mini-batches, built on modern deep learning machinery; scaling laws over decay rate and batch size yield a formulaic rule to set hyperparameters without sweeps.
- Results: matches Lloyd k-means' 25-epoch loss in 1.4–2.6 epochs across embedding datasets, translating to 10–45x GPU wall-clock speedups over state-of-the-art implementations, often beating k-means++. It partitions DINO10B (10B embeddings) into 16M clusters in 2.5 epochs with lower loss than Lloyd at 25 epochs; 10x compute savings vs Faiss-GPU/Flash-KMeans at billion scale.
Clustering underpins retrieval, quantization, and data curation for LLM pipelines. Code is open-sourced in FAISS.
Related event: Meta FAIR's e-means Beats k-means with Up to 45x GPU Speedup(3 posts)→
More from Infra
- GB300 Rents for $12-15/hr as Inference Supply Outstrips a Thin Buyer Market — teodorio · 2026-11-10
- Report: NVIDIA May Cut TSMC Orders as AMD and ASIC Makers Circle for Capacity — zephyr_z9 · 2026-11-10
- RTX 5090 local LLM math: a day of inference outputs less than $1 worth of API usage — PurzBeats · 2026-11-10
- CoreWeave needs ~350MW more live by year-end — the exact spot it slipped last year — tengyanAI · 2026-11-10
- Memphis data centers led by xAI generated $54M in tax revenue amid pause proposal — pickover · 2026-11-10
- Memphis mayor says SpaceX data center taxes more than doubled to $54M — elonmusk · 2026-11-10