AI & ML Benchmarks

Benchmarks for AI / ML workloads that modern cloud databases increasingly handle natively - currently vector similarity search, with SQL-based AI functions covered under Primitives.

Why AI/ML Benchmarks?

Analytical databases are absorbing workloads that used to require dedicated ML infrastructure. Vector similarity search - kNN / ANN over embedding vectors - is the most mature case: DuckDB, PostgreSQL (pgvector), ClickHouse, Snowflake, StarRocks, and Doris all ship array / vector types with distance functions.

Vector workloads differ from OLAP:

• High-dimensional distance computations dominate CPU cost
• Recall vs. latency trade-offs depend on ANN index choice (HNSW, IVF)
• Data shapes are different - fixed-length arrays and large row counts

BenchBox ships repeatable vector-search queries with synthetic embeddings so you can compare engines without depending on an external embedding service.

AI/ML Benchmarks in BenchBox

Benchmark	Focus	Target Platforms
Vector Search	kNN / ANN similarity search over embedding vectors	DuckDB, pgvector, ClickHouse, Snowflake, StarRocks, Doris

Synthetic embeddings are generated locally with a fixed seed, so there is no per-query API cost regardless of scale.

Included Benchmarks

• Vector Search Benchmark

Note

AI Primitives - SQL-based AI functions (Snowflake Cortex, BigQuery ML, Databricks AI) - is a related but distinct workload registered under Primitives in the benchmark registry. See BenchBox Primitives for coverage.

See Also

• BenchBox Primitives - Core SQL operation primitives (includes AI Primitives)
• BenchBox Experimental - Other emerging benchmarks