The Calibrator for AI-on-Chips epitomizes precision in maintaining high accuracy for AI System-on-Chips through advanced post-training quantization (PTQ) techniques. It offers architecture-aware quantization that sustains accuracy levels up to 99.99% even in fixed-point architectures like INT8. This ensures that AI chips deliver maximum performance while staying within defined precision margins.
Central to its operation, Calibrator uses a unique precision simulator to emulate various precision-change points in a data path, incorporating control information that synchronizes with ONNC's compiler for enhanced performance. The integration with ONNC's calibration protocols allows for the seamless refinement of precision, thereby reducing precision drop significantly.
Highly adaptable, the Calibrator supports multiple hardware architectures and bit-width configurations, ensuring robust interoperability with various deep learning frameworks. Its proprietary entropy calculation policies and architecture-aware algorithms ensure optimal scaling factors, culminating in a deep learning model that is both compact and precise.
