Meta FAIR's e-means beats k-means with 10-45x GPU speedups on embeddings
meta · hf · 2026-11-11
Meta FAIR introduces e-means, a clustering algorithm that replaces Lloyd's classic iterations with an exponential moving average optimization using mini-batch updates built on modern deep learning machinery.
Key points:
- The paper derives scaling laws for how decay rate and batch size affect final performance, yielding a formulaic rule to fix hyperparameters without any sweep.
- Across a wide range of embedding datasets, e-means matches Lloyd k-means' loss at 25 epochs within just 1.4-2.6 epochs, translating to 10-45x GPU wall-clock speedups over state-of-the-art implementations, frequently beating k-means++.
- Example: partitioning DINO10B (10 billion DINO embeddings) into 16 million clusters in only 2.5 epochs.
- Code is being integrated into FAISS.
Related event: Meta FAIR's e-means outperforms k-means in one-tenth the epochs(7 posts)→
More from Infra
- AI networking market to hit $278.9B by 2027 at 95% CAGR, TrendForce forecasts — Beth_Kindig · 2026-11-11
- GPU Cluster Financing Shifts to Prepayments, Making OCF/EBITDA the Key Metric — GavinSBaker · 2026-11-11
- DeepMind researcher: multimodal compute will be priced in world units, not tokens — prateekj · 2026-11-11
- Report: Alice Huang leaned on her close ties with Jensen Huang to win Nvidia chip customers — GaryMarcus · 2026-11-11
- FAA's SMART system rerouted a 119-passenger flight around storms, saving 33 minutes — deanwball · 2026-11-11
- Qwen Image 2.1 runs 30-40% slower on Windows than Linux, user reports — wacomlover · 2026-11-11