Processor and method for improved matrix storage

Abstract

An adaptive quantization system and method for the efficient storage and processing of matrix data are disclosed. The system comprises a processor with circuitry configured to divide a matrix into blocks, assign quantization formats to each block based on its content, create a superblock containing the quantized blocks and associated format information, and store the superblock in memory. The method uses an index to specify quantization formats, typically employing 3 bits per block, and can utilize common exponents for certain formats. A metadata buffer stores compression information and enables the gradual introduction of compression.The invention optimizes matrix data storage at a fine-grained level, ensures efficient random access, adapts to varying accuracy requirements within a matrix, and can be integrated into existing storage hierarchies. This approach significantly improves compression efficiency and processing speed for large-scale matrix operations in fields such as artificial intelligence, machine learning, and high-performance computing.

Images