Static Lookup Table MAC Device for Compute-in-Memory Power Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current compute-in-memory systems face high power consumption and data transfer bottlenecks due to the need for dynamic multiplier and adder circuits in MAC operations, especially in deep neural networks, which limits their performance and efficiency.
Innovation Solution
The implementation of static lookup table (LUT) circuits replaces dynamic multiplier and adder circuits to reduce peak and average current consumption by performing MAC operations based on pre-defined truth tables, eliminating the need for dynamic logic operations in most cases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If dynamic multiplier and adder circuits are used in MAC operations, then computation capability is improved, but power consumption increases significantly
Solution Approach 1:
The patent creates simplified copies of multiplication and addition operations using lookup tables that store pre-computed results. Instead of implementing full dynamic multiplier and adder circuits, the system uses memory cells to store weight values and input signals, then performs simplified MAC operations by reading from these stored values and using basic memory cell operations to compute results, dramatically reducing power consumption while maintaining computational capability
Solution Approach 2:
The patent changes the operational parameters of the MAC unit by operating memory cells in specific read modes (e.g., differential read, single-ended read) and using voltage level adjustments to perform multiplication and addition. By adjusting voltage levels and using different readout schemes, the system achieves MAC functionality without requiring traditional dynamic arithmetic circuits, thereby reducing power consumption
2Measurement precision
If dynamic logic operations are performed for MAC operations, then computational accuracy is maintained, but current consumption increases
Solution Approach 1:
The patent uses lookup tables to store pre-computed weight values and input signal combinations, copying results directly from memory rather than performing dynamic computations. This approach maintains computational accuracy by retrieving exact pre-computed values while avoiding the high current consumption associated with dynamic logic operations in traditional MAC units
3Adaptability or versatility
If traditional MAC unit architecture is used, then functionality is complete, but device complexity and power consumption are high
Solution Approach 1:
The patent merges the functions of multipliers, adders, and memory into a single integrated MAC unit structure. Memory cells serve dual purposes as both storage elements and computation elements, with the same memory infrastructure used for both weight storage and input signal processing. This merging eliminates the need for separate dynamic arithmetic circuits, reducing device complexity while maintaining complete MAC functionality
Solution Approach 2:
The patent implements a universal MAC unit architecture where memory cells perform multiple functions: storing weights, storing input signals, and performing MAC operations through their readout mechanisms. The same physical infrastructure supports all computational operations, making the device highly versatile while minimizing circuit complexity and power consumption
Data Source
AI summary
A multiply-accumulate (MAC) device for compute-in-memory (CIM) includes an input driver configured to provide a plurality of input signals including a first input signal and a second input signal. A lookup table (LUT) stores or accesses a plurality of CIM weight signals including a first CIM weight signal and a second CIM weight signal. The LUT is configured to receive the first input signal and the second input signal and provide a sum output based on the first and second input signals and the first and second CIM weight signals.