MASAR Column Using Analog Charge Summation for AI MAC
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning and neural network models require significant memory and computational power for large numbers of weights, leading to increased energy and time costs due to the need for extensive data transfer and multiplication-accumulate operations in digital hardware implementations.
Innovation Solution
The implementation of a Multiply-Accumulate Successive Approximation (MASAR) column, which uses digital and analog circuit techniques to perform digital multiplication and analog summation, storing results as analog charge in capacitors and converting them back to digital using a SAR ADC, allowing for efficient MAC operations without additional ADCs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If digital hardware implementations are used for MAC operations, then computational power is provided, but energy and time costs increase due to extensive data transfer and multiplication-accumulate operations
Solution Approach 1:
The patent replaces digital multiplication and accumulation operations with analog circuit operations. Unit capacitors perform analog multiplication by storing charge proportional to the product of input signals, and analog summation is performed by connecting capacitors in parallel. This substitution of digital computation with analog physics-based operations reduces energy consumption and computational time while maintaining the required computational power for neural network inference.
Solution Approach 2:
The patent changes the operational parameters from digital domain to analog domain. By representing computational values as analog voltages and charges rather than digital bits, the system enables parallel MAC operations to be performed simultaneously in analog circuits, dramatically reducing the energy and time required compared to sequential digital computation.
2Productivity
If digital hardware implementations are used for MAC operations, then computational functionality is achieved, but time costs increase due to extensive multiplication-accumulate operations
Solution Approach 1:
The patent substitutes iterative digital multiplication and accumulation with direct analog circuit operations. The analog MAC cell uses capacitors to perform multiplication (charge = voltage × capacitance) and summation (parallel capacitor connections) simultaneously in a single operation, eliminating the sequential processing steps required in digital implementations and thus reducing computation time while improving productivity.
Solution Approach 2:
The patent performs preliminary configuration of the analog circuit before computation. Unit capacitors are pre-configured with specific capacitance values corresponding to weight parameters, and the circuit topology is prepared in advance. This preliminary setup enables the actual MAC operation to be executed instantly when input signals are applied, significantly reducing computation time compared to digital systems that must sequentially process each operation.
3Adaptability or versatility
If large numbers of weights are stored on ASIC, then neural network capacity increases, but memory requirements and data transfer increase
Solution Approach 1:
The patent replaces digital memory storage with analog physical representation. Weight parameters are encoded as capacitance values of unit capacitors rather than being stored as digital data in memory cells. This physical encoding allows the ASIC to inherently store and process weight parameters without requiring separate memory structures, reducing memory requirements while maintaining high neural network capacity.
Solution Approach 2:
The patent makes the unit capacitors multi-functional. The same capacitors serve both as storage elements for weight parameters and as computational elements for performing MAC operations. This eliminates the need for separate memory and computation units, reducing overall memory requirements while enabling the ASIC to handle large numbers of weights for high-capacity neural networks.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces energy and time costs by enabling efficient MAC operations and multibit precision computations, suitable for AI/ML applications, while minimizing memory requirements and data transfer.
Implementation Method 1
a unit capacitor configured to store the result as analog charge
Implementation Method 2
digital logic configured to perform analog summation of the analog charge of the unit capacitors
Implementation Method 3
configuring the unit capacitors as a capacitive digital to analog converter (CDAC) in a successive approximation register (SAR) analog to digital converter (ADC)
Data Source
AI summary
A multiply-accumulate successive approximation (MASAR) column is provided. The MASAR column includes a plurality of MASAR cells, each including a multiplier configured to perform digital multiplication between an input activation received to an input and an operand to compute a result, and a unit capacitor configured to store the result as analog charge. The MASAR column further includes digital logic configured to perform analog summation of the analog charge of the unit capacitors of the plurality of MASAR cells to determine a digital output of the multiplication.