CiM HDR Digitization Using Tunable Gain for Sparse MAC Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing compute-in-memory (CiM) solutions face challenges in achieving high-precision analog computing due to limited digitization accuracy, particularly when using conventional fixed-resolution analog-to-digital converters (ADCs), which results in significant data truncation and reduced MAC computation accuracy, especially when input activation vectors are sparse and expected MAC values are smaller than the ADC quantization step.
Innovation Solution
A high dynamic range (HDR) digitization scheme is introduced, which increases the digitization dynamic range by up to 2M times through analog amplification of ADC inputs, allowing for a variable quantization granularity without increasing power consumption or ADC complexity, using a tunable amplifier and M-bit exponent quantizer to adjust the quantization step based on the magnitude of MAC output values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional fixed-resolution ADCs are used for digitization, then device complexity is reduced, but measurement precision deteriorates due to data truncation when MAC values are smaller than the ADC quantization step
Solution Approach 1:
The patent implements a tunable amplifier with variable gain that can dynamically adjust the amplification factor based on the magnitude of the analog MAC output signal. This dynamic adjustment allows the system to optimize the signal range for different input conditions, thereby improving digitization accuracy without requiring a higher-resolution fixed ADC. The amplifier transitions from a static component to an adaptive one that responds to signal characteristics in real-time.
Solution Approach 2:
The system changes the gain parameter of the amplifier based on the exponent quantizer output, which reflects the magnitude of the MAC result. By adjusting this parameter dynamically, the system effectively scales the signal to match the ADC's optimal input range, reducing quantization error and improving measurement precision without increasing ADC resolution or complexity.
2Measurement precision
If higher-resolution ADCs are used to improve digitization accuracy, then measurement precision is improved, but use of energy increases due to higher power consumption
Solution Approach 1:
Instead of increasing ADC resolution, the system changes the gain parameter of the amplifier to scale the signal appropriately. This parameter adjustment allows the same ADC to achieve effective higher precision by optimizing the input signal range, thereby maintaining digitization accuracy without the energy penalty of higher-resolution ADCs.
Solution Approach 2:
The tunable amplifier acts as an intermediary between the analog MAC output and the ADC. It preprocesses the signal by applying variable gain, ensuring that the ADC operates at its optimal range. This intermediary approach allows the system to achieve high measurement precision using a lower-resolution, lower-power ADC.
3Measurement precision
If higher-resolution ADCs are used to reduce data truncation, then measurement precision is improved, but productivity decreases due to reduced conversion speed
Solution Approach 1:
The system adjusts the amplifier gain parameter dynamically based on signal magnitude, allowing a fixed-resolution, high-speed ADC to achieve effective higher precision. This avoids the speed penalty associated with higher-resolution ADCs while maintaining accurate digitization through signal preprocessing.
Solution Approach 2:
The amplifier performs preliminary signal conditioning before the ADC conversion by scaling the analog MAC output to the optimal input range. This preliminary action ensures that the ADC operates at peak efficiency with maximum utilization of its dynamic range, achieving high precision without sacrificing conversion speed.
4Measurement precision
If analog amplification with tunable gain is implemented, then measurement precision is improved through variable quantization granularity, but device complexity increases
Solution Approach 1:
The system segments the digitization process into two stages: analog amplification with variable gain followed by fixed-resolution ADC conversion. The exponent quantizer divides the dynamic range into discrete segments, each corresponding to a specific amplifier gain setting. This segmentation allows complex variable-precision digitization to be achieved through coordinated simple components.
Solution Approach 2:
The patent replaces a mechanical or complex variable-resolution ADC system with an electronic amplifier controlled by digital exponent quantizer outputs. This substitution uses electronic gain control instead of mechanical or complex switching mechanisms, reducing overall device complexity while achieving variable quantization granularity.
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 enhances the quantization granularity of analog MAC values, reducing data truncation and maintaining energy efficiency, thereby improving the accuracy of MAC computation without the need for higher-resolution ADCs, which would increase power consumption or reduce conversion speed.
Implementation Method 1
analog amplification of ADC inputs, allowing for a variable quantization granularity without increasing power consumption or ADC complexity
Data Source
AI summary
Systems, apparatuses and methods may provide for compute-in-memory (CiM) accelerator technology that includes a multiply-accumulate (MAC) computation stage, an analog amplifier stage coupled to an output of the MAC computation stage, and an analog to digital conversion (ADC) stage coupled to an output of the analog amplifier stage, wherein a gain setting of the analog amplifier stage modifies a quantization granularity of the ADC stage.