Overview
Scalar Quantization (SQ) is a simpler compression technique than Product Quantization that reduces the precision of each vector component, typically from 32-bit floating point to 8-bit integers, achieving 4x memory reduction.
How Scalar Quantization Works
Quantization Process
- Find Range: Determine min and max values for each dimension
- Map: Map the continuous range to discrete integer values (e.g., 0-255 for 8-bit)
- Encode: Convert each float component to its quantized integer
- Store: Store compact 8-bit representations
Search Process
- Quantize query vector using same mapping
- Compute approximate distances using quantized vectors
- Optionally rerank top candidates using original vectors
Memory Reduction
- 32-bit float → 8-bit int: 4x compression
- 32-bit float → 4-bit int: 8x compression
- Example: 768-dim vector: 3,072 bytes → 768 bytes (8-bit) or 384 bytes (4-bit)
Variants
Asymmetric Scalar Quantization
- Database vectors: quantized
- Query vectors: kept in full precision
- Better accuracy than symmetric quantization
Per-Dimension vs Global
- Per-dimension: Different min/max for each dimension
- Global: Single min/max across all dimensions
Trade-offs
Advantages:
- Simple to implement
- Minimal accuracy loss (1-2% typical)
- Fast compression and decompression
- Works well with modern SIMD instructions
Disadvantages:
- Less compression than Product Quantization
- Still requires significant memory for large datasets
Comparison with Other Methods
vs Product Quantization:
- Less compression but better accuracy
- Simpler and faster
- No training required
vs Binary Quantization:
- Better accuracy but less compression
Use Cases
- When 4x compression is sufficient
- Applications requiring high accuracy
- Real-time search scenarios
- When simplicity is valued
Pricing
Implemented in most vector databases (Qdrant, Milvus, Weaviate, etc.)
