Lookup table access for quantization and dequantization operations
By representing quantized floating-point numbers as integer indices in a LUT and using a multiplexer hierarchy, the computational overhead of ML operations is minimized, enhancing efficiency in processing large datasets.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- ADVANCED MICRO DEVICES INC
- Filing Date
- 2025-12-16
- Publication Date
- 2026-07-02
AI Technical Summary
Conventional quantization and dequantization algorithms for machine learning (ML) operations consume significant computational resources due to the use of mathematical functions, especially when processing large datasets in different number formats, leading to increased processing overhead.
Represent quantized floating-point numbers as integer indices in a lookup table (LUT) to reduce computational overhead, using a hierarchy of multiplexers to access cache lines based on these indices, allowing conversions to be performed with a single instruction.
Reduces storage and bandwidth requirements while minimizing computational overhead by converting or dequantizing data using a LUT, enabling efficient matrix operations with reduced resource consumption.
Smart Images
Figure US2025059834_02072026_PF_FP_ABST