Quantisierung

Nvidia application: Shared number scales for compact AI calculations

An approach published on October 1 aims to support addition and subtraction in neural networks at reduced numerical precision. Model benchmarks are missing.