sbert-10k-sentence-speed-comparison
IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s0-abstract.md
Created 2026-08-25T02:58:27+00:00
Finding the most similar pair among 10,000 sentences requires ~49,995,000 forward passes (~65 hours) with a BERT cross-encoder on a single V100 GPU, versus ~5 seconds with SBERT embeddings plus cosine similarity.
Summary
Comparing every pair of 10,000 sentences with a cross-encoder model is computationally infeasible, taking nearly three days on a single GPU. The practical alternative is to encode each sentence once into a fixed vector and then compare those vectors with simple math, finishing in seconds, which is what makes large-scale similarity search actually usable.