Mala Mahadevan tries something out:
I was experimenting with vector search in SQL Server 2025 using approximately 3.8 million Hacker News articles that that I used for my talk on SQL Server-based vector search feature at EightKB. I generated two sets of embeddings from the same articles: one stored as FP32 (32-bit floating point) and the other as FP16 (16-bit floating point). Both sets contained the same articles and used DiskANN vector indexes.I used a query as below to find articles similar to a selected example.
What Mala found was that FP16 queries could sometimes take longer and require more reads than FP32 queries, which doesn’t make sense on its face. Click through to see two experiments she devised to come up with the answer.