Quantization
Definition
Quantization maps values onto fewer possible numerical levels.
Mapping values onto a smaller set of discrete representation levels, often with lower bit width. The goal is reduced memory and bandwidth requirements and, on suitable hardware, possibly less computational effort.
Simply explained
Like coarser markings on a scale, a simplified number representation saves space.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
Weights are mapped from FP16 to INT8 and tested against quality and latency targets.
Limits and pitfalls
- Automatically equating lower bit width with real acceleration or ignoring quantization granularity, scales, zero points, outliers, activations, and target hardware.
Origin & history
Quantization originates from signal processing and has been further developed for efficient neural inference.
Use in mid-sized companies
- Possible application: weights are mapped from FP16 to INT8 and tested against quality and latency targets.
Frequently asked questions
What does quantization mean?
The mapping of values onto a smaller set of discrete representation levels, often with lower bit width. The goal is reduced memory and bandwidth requirements and, on suitable hardware, possibly less computational effort. What mistakes should be avoided? Automatically equating lower bit width with real acceleration or ignoring quantization granularity, scales, zero points, outliers, activations, and target hardware.
Sources
Source checked on 10.10.2026. Editorial use cases are hypothetical; no evidence of real customer results.
Editorially reviewed on: