Lookup-Table Computing Accelerator for Faster AI Multiplication

The patent pre-calculates and stores multiplication results in lookup tables before runtime. During neural network inference, the system retrieves pre-computed values from the lookup table based on input quantized values, avoiding real-time multiplication operations and significantly reducing computational time while maintaining acceptable accuracy through controlled approximation.

The patent creates simplified copies of multiplication operations by storing results in lookup tables. Instead of performing actual multiplication computations during inference, the system uses table lookups with quantized indices, effectively replacing complex arithmetic operations with simpler memory access operations that are much faster to execute.

The patent divides the lookup table into multiple segments or banks organized in a three-dimensional memory structure. This segmentation allows parallel access to different portions of the lookup table, improves memory utilization efficiency, and reduces access conflicts. The segmented structure also enables better distribution of memory resources and facilitates hierarchical memory organization.

The patent transitions from traditional two-dimensional lookup tables to a three-dimensional memory assembly structure. This dimensional change enables more efficient memory organization, allows for better spatial locality exploitation, and provides additional degrees of freedom for memory access patterns. The third dimension can represent different precision levels, different operation types, or different data groups, enabling multi-functional access from a single memory structure.

The patent changes the precision parameter dynamically by using different quantization levels for different input ranges or different neural network layers. For less critical computations, lower precision (fewer bits) is used to maximize speed, while for more critical computations requiring higher accuracy, the system uses higher precision levels. This adaptive parameter change allows the system to optimize the trade-off between accuracy and speed based on specific computational requirements.

The patent uses partial precision instead of full precision for lookup table operations. By storing and retrieving only the most significant bits of multiplication results, the system achieves sufficient accuracy for neural network inference while dramatically reducing the memory footprint and access time. This partial action approach recognizes that full precision is often unnecessary for the approximate nature of neural network computations.