3D Lookup-Table Accelerator for Floating-Point Dot Products
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence systems, particularly neural network systems, face performance bottlenecks due to the time-consuming nature of operations like multiplication and matrix multiplication, which significantly impact overall system performance.
Innovation Solution
A computing accelerator system that utilizes a lookup table to accelerate floating-point multiplications by retrieving the fraction portion of the product of two floating-point operands, leveraging symmetry and zero-skipping to improve performance, and is implemented in a three-dimensional memory assembly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional multiplication operations are used in neural network systems, then computational accuracy is maintained, but system performance and processing speed deteriorate due to time-consuming calculations
Solution Approach 1:
The patent pre-calculates and stores multiplication results in lookup tables during system initialization or offline processing. During actual neural network inference, the system directly retrieves pre-computed results from the lookup tables based on input values, eliminating the need for real-time multiplication operations and significantly reducing computational time.
Solution Approach 2:
The patent creates simplified copies of multiplication operations by storing results in lookup tables. Instead of performing actual multiplication computations, the system uses table lookups with pre-stored results, effectively replacing complex computational operations with simple memory access operations that are much faster to execute.
2Speed
If lookup tables are used to accelerate floating-point multiplications, then computational speed is improved, but memory requirements and device complexity increase
Solution Approach 1:
The patent transitions from traditional two-dimensional memory structures to three-dimensional stacked memory architecture. Multiple memory layers are vertically stacked and interconnected, enabling higher storage capacity and faster access speeds within a compact footprint. This dimensional change allows the system to accommodate large lookup tables without proportionally increasing the physical device complexity.
Solution Approach 2:
The patent implements a hierarchical memory structure where lookup tables are organized across multiple stacked memory layers. The memory architecture nests smaller memory units within a larger three-dimensional structure, with each layer containing portions of the lookup table. This nested organization allows efficient data retrieval by accessing only the necessary memory layer, reducing overall system complexity while maintaining high-speed access.
Data Source
AI summary
A computing accelerator using a lookup table. The accelerator may accelerate floating point multiplications by retrieving the fraction portion of the product of two floating-point operands from a lookup table, or by retrieving the product of two floating-point operands of two floating-point operands from a lookup table, or it may retrieve dot products of floating point vectors from a lookup table. The accelerator may be implemented in a three-dimensional memory assembly. It may use approximation, the symmetry of a multiplication lookup table, and zero-skipping to improve performance.