Quantization
is a technique used to reduce the precision of a 's weights and activations to decrease memory footprint and increase inference speed. By mapping high-precision values, typically 32-bit floating-point (), to lower-precision formats such as 8-bit integers (), , or even 4-bit representations, it enables the deployment of on resource-constrained hardware like devices or mobile processors. Common strategies include (PTQ) and (QAT), both aiming to minimize the loss in caused by discretization errors. Major industry players like , , and utilize this technique within frameworks like and to optimize performance on and architectures.
[PREVIEW MODE] Definitions streamed at five depths, references, and related terms are available to signed-in readers — [SIGN IN]
Contextual terminology map
KEEP IN VIEW
Ethics, safety & society
Moving as fast as the field, and easier to overlook
- Explainable AI
- Interpretability
- Model Card
- AI Audit
- Algorithmic Bias
- Disparate Impact
- Digital Divide
- WCAG (Web Content Accessibility Guidelines)
- AI Alignment
- Red Teaming
- EU AI Act
- NIST AI RMF (AI Risk Management Framework)
- Frontier Model
- Deepfake
- Content Credentials
- AI Watermarking
- Job Displacement
- Prompt Injection
- Data Poisoning
- Differential Privacy
- GDPR (General Data Protection Regulation)
- Zero-Day
- End-to-End Encryption
- Data Broker
Infrastructure, markets & the economy
The compute, power and capital behind the boom