arXiv benchmark finds int8 quantization nearly matches float32 on text embeddings
Original title: Balancing hierarchical representation and quantization: A performance trade-off in text embeddings
int8 quantization retains 94.4% baseline effectiveness at 32x storage reduction across 38 MTEB tasks.
The study empirically evaluates four transformer-based embedding models on 38 MTEB tasks using 1D and 2D hierarchical representations combined with 32-bit, 8-bit, and binary quantization. Int8 quantization is statistically indistinguishable from float32 for pair classification and re-ranking, costing at most 1.5 percentage points overall. Binary quantization performs significantly worse across every category, while combining 8x dimensionality reduction with int8 yields 32x storage savings at 94.4% effectiveness. Retrieval tasks degrade fastest; no evidence shows quantization speeds up evaluation.
- int8 quantization statistically similar to float32 on pair classification and re-ranking
- halving embedding dimension once not detectable above m=16
- binary quantization significantly worse across all task categories
- 32x storage reduction at 94.4% baseline effectiveness
- no evidence quantization speeds up evaluation time
From the original article
1 min read at arxiv.orgHierarchical text embedding models offer variable-length representations, while quantization reduces numerical precision; both aim to improve computational efficiency. This work is an empirical benchmark study of how these two orthogonal, already-established techniques interact.
By Kai Ponel · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.