Skip to content

Perf: In‐memory providers

Jump to bottom
Harsha Simhadri edited this page May 13, 2026 · 3 revisions

Query Throughput Comparison of DiskANN3 + In-Memory Providers v0.47 with HNSW and Intel SVS

All experiments with 32 threads on Azure D32v6 VM using v0.47 with AVX512 enabled.

With Full-precision vectors

Dataset: Wikipedia articles + Cohere embeddings (768 Dimensions, float32)

wikipedia-1m-float32-32-threads.pdf image

Dataset: MSTuring (100 Dimensions, float32)

msturing-1m-float32-32-threads.pdf image

With Spherical Quantized Vectors + Re-ranking

Dataset: Wikipedia articles + Cohere embeddings (768 Dimensions, float32)

wikipedia-1m-quantized-32-threads.pdf image

Dataset: MSTuring (100 Dimensions, float32)

msturing-1m-quantized-32-threads.pdf image

Clone this wiki locally