Benchmark shows new e-means clustering 100x slower than Lance's hierarchical k-means
leland_mcinnes · x · 2026-11-17
FAISS recently shipped e-means, claiming it beats k-means with 10x fewer epochs. A quick benchmark by changhiskhan's team tells a different story.
- Setup: 33.55M real DINO vectors, K=131,072, d=1,024, 32 CPU threads, one run per method, sampling/loading excluded
- Results: Lance's default hierarchical k-means finished in 182.1s, fastest by far; E-means IVF4 took 508.4s, IVF16 991.3s, and the official E-means implementation a whopping 23,063.8s
- Final MSE was roughly comparable across methods (E-means slightly better at 0.98-0.99 relative), nowhere near enough to offset the speed gap
Takeaway: on real large-scale data, classic hierarchical k-means remains dramatically faster, and e-means's claimed speedups did not reproduce.
More from Infra
- Altman explains 'double counting' warning: chipmakers all assume OpenAI will buy their supply — toptickcrypto · 2026-11-17
- RightNow launches inference API, claims 14x NVIDIA performance via GPU kernels — ycombinator · 2026-11-17
- Baseten open-sources Rust/CUDA DeepSeek engine; agent on Vera Rubin finds optimizations speeding Blackwell by 10% — baseten · 2026-11-17
- Wici Infer offers free large-model serving on consumer hardware via VRam+Ram+SSD — AlexDorofeev · 2026-11-17
- 943 agents, 11.5B tokens, $5,900: 97% of tokens went to re-reading context — chrisfirst · 2026-11-17
- MCP server returns parquet handles instead of rows, cutting 3.2M tokens to 1,700 — Daniel_Aperiodic · 2026-11-17