Switched Capacitor Cores Scaling for ADC Truncation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Low-precision ADC truncation in switched capacitor computation cores for neural network inference reduces accuracy due to reduced precision of analog MACC outputs, leading to decreased performance.
Innovation Solution
An auto-search algorithm and hardware implementation to determine optimal integer scalars for ADC truncation, incorporating scaling options such as amplifiers, input multipliers, and charge sharing to modify analog signals during switched-capacitor computation, ensuring accurate neural network inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If low-precision ADC is used in switched capacitor computation cores, then device complexity and manufacturing cost are reduced, but measurement precision of analog MACC outputs deteriorates
Solution Approach 1:
The patent applies preliminary action by determining optimal integer scalars during a training phase before actual inference. The auto-search algorithm pre-computes scaling factors that will be applied during inference to compensate for ADC truncation effects, ensuring accuracy is maintained without requiring complex real-time adjustments
Solution Approach 2:
The patent changes the parameter of scaling factors applied to MACC operations. By dynamically adjusting these scalar values based on the specific neural network layer and input characteristics, the system compensates for precision loss from low-precision ADC while maintaining compatibility with simplified hardware
2Manufacturing precision
If ADC truncation is applied to reduce precision errors, then manufacturing precision is improved, but loss of information in analog MACC outputs increases
Solution Approach 1:
The patent introduces an intermediary scaling mechanism between the analog MACC output and the ADC conversion. The optimal integer scalars act as mediators that pre-condition the analog signals to maximize the effective use of limited ADC resolution, thereby reducing information loss during the truncation process
Solution Approach 2:
The system implements feedback through the auto-search algorithm that determines optimal scalars based on training data characteristics. This feedback loop allows the system to learn and adapt scaling factors that minimize information loss for specific neural network workloads
3Reliability
If scaling factors are applied to MACC inputs or outputs, then neural network accuracy is maintained, but device complexity increases
Solution Approach 1:
The patent applies local quality by implementing scaling operations only at specific critical points in the computation pipeline - namely at the input of certain neural network layers where precision is most impactful. This selective application maintains accuracy while minimizing the overall complexity burden
Solution Approach 2:
The system uses simple digital scaling factors that can be replicated and applied through straightforward multiplication operations. These digital scalars are much simpler to implement than analog scaling circuits, achieving the same accuracy improvement with minimal complexity increase
Data Source
AI summary
An apparatus comprising: a first plurality of inputs representing an activation input vector; a second plurality of inputs representing a weight input vector; an analog multiplier-and-accumulator to generate a first analog voltage representing a first multiply-and-accumulate result for the said first inputs and the second inputs; a voltage multiplier that takes the said first analog voltage and produces a second analog voltage representing, a second multiply-and-accumulate result by multiplying at least one scaling factor to the first analog voltage; an analog to digital converter configured to convert the said second analog voltage multiply-and-accumulate result into a digital signal using a limited-precision operation during a neural network inference operation; and a hardware controller configured to determine the at least one scaling factor based on the first multiply-and-accumulate result, or a software controller configured to determine the at least one scaling factor based on the first multiply-and-accumulate result.