Mixed-Signal Bitwise Multiplication With Split Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Mixed signal circuit designs face limitations due to the energy constraints of analog-to-digital conversion (ADC) in performing computations, particularly in neural network workloads where matrix multiplications are energy-inefficient in the digital domain but more efficient in the analog domain.
Innovation Solution
Implementing mixed signal multiply-accumulate (MS-MAC) circuitry that combines digital and analog processing to perform bitwise multiplication with varying accuracies, using different architectures for different dot product engines based on noise contributions to enhance computation accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If ADC conversion is used to perform computations in mixed signal circuitry, then digital computation capability is achieved, but energy consumption increases significantly
Solution Approach 1:
The patent segments the multiplication computation into two distinct portions: a first portion performed with higher accuracy and a second portion performed with lower accuracy. This segmentation allows the system to allocate computational resources differently across different parts of the same operation, reducing overall energy consumption while maintaining necessary precision for critical calculations.
Solution Approach 2:
The patent applies local quality by using different accuracy levels for different portions of the multiplication operation. The first portion uses higher accuracy (more bits) where precision is critical, while the second portion uses lower accuracy (fewer bits) where approximate results are sufficient. This non-uniform quality distribution optimizes the trade-off between computation capability and energy consumption.
2Measurement precision
If uniform high accuracy is used for all bitwise multiplication portions, then computation precision is maximized, but energy consumption and circuit complexity increase
Solution Approach 1:
The patent implements local quality by assigning different accuracy levels to different portions of the multiplication operation. The first portion is computed with higher accuracy using more bits, while the second portion uses lower accuracy with fewer bits. This approach maintains necessary precision for critical calculations while reducing energy consumption and circuit complexity for less critical portions.
Solution Approach 2:
The patent divides the multiplication operation into two segments with different accuracy requirements. This segmentation allows the system to apply high precision only where necessary and use lower precision elsewhere, optimizing the balance between measurement precision and energy consumption.
3Device complexity
If all bitwise multiplication portions are performed with the same accuracy, then circuit design is simplified, but noise and bit errors increase in mixed-signal domain
Solution Approach 1:
The patent applies local quality by using different accuracy levels for different portions of the multiplication operation. This differential approach reduces noise and bit errors in the mixed-signal domain by ensuring that critical portions are computed with sufficient precision while allowing non-critical portions to use lower precision, thereby improving overall computation reliability.
4Reliability
If higher accuracy is used for all multiplication operations, then noise and bit errors are reduced, but energy consumption increases
Solution Approach 1:
The patent implements local quality by assigning different accuracy levels to different portions of the multiplication operation. Critical portions that require high reliability are computed with higher accuracy, while non-critical portions use lower accuracy. This selective approach reduces noise and bit errors where necessary while minimizing energy consumption overall.
Solution Approach 2:
The patent segments the multiplication operation into portions with different reliability requirements. By computing only the critical portions with high accuracy and using lower accuracy for non-critical portions, the system achieves the necessary computation reliability without the prohibitive energy cost of uniformly high-precision computation throughout.
Data Source
AI summary
An apparatus comprises at least one processor and at least one memory including instruction code configured to, with the at least one processor, cause the apparatus at least to perform, with a first accuracy, a first portion of a bitwise multiplication of first and second digital inputs and to perform, with a second accuracy different than the first accuracy, at least a second portion of the bitwise multiplication.