Computing-in-Memory Data Conversion Mesh for Accurate MAC
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Advanced AI edge chips require high accuracy in multi-bit input and weight data for MAC operations, but existing technologies face challenges in achieving efficient energy consumption and silicon area optimization for these operations.
Innovation Solution
A data converter system using a mesh structure with Digital-to-Analog Converters (DAC) and Analog-to-Digital Converters (ADC) coupled with memory elements, enabling efficient data conversion and multiplication in computing-in-memory architecture, which includes a single-end current steering DAC, Trans-impedance Amplifier (TIA), and Variable Gain Amplifier (VGA) to enhance signal integrity and reduce power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-bit input data and weight data are used for MAC operations to achieve specified accuracy, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The data converter is segmented into multiple receive paths and transmit paths arranged in a mesh structure. Each receive path includes one DAC and multiple memory elements, while each transmit path includes one ADC and multiple memory elements. This segmentation allows the system to handle multi-bit data conversions through parallel processing across multiple paths, achieving high accuracy without requiring a single complex converter.
Solution Approach 2:
The patent transitions from a single-path sequential conversion approach to a multi-dimensional mesh structure with N rows and M columns. This dimensional expansion enables parallel data conversion operations, where multiple DACs and ADCs operate simultaneously across different paths, reducing the complexity burden on any single converter while maintaining overall high accuracy through the collective capability of the mesh network.
2Use of energy by stationary object
If computing-in-memory is implemented to improve energy efficiency of MAC operations, then power consumption is reduced, but manufacturing precision requirements increase
Solution Approach 1:
The patent merges the computing function with the memory function by implementing MAC operations directly within the memory array. Memory elements store weight data and perform multiplication with input data through analog current modulation, while the accumulation is performed through current summation in the readout circuitry. This merging eliminates the need for separate compute units, significantly reducing power consumption while the mesh structure with multiple DACs and ADCs provides the necessary precision through parallel sampling and digital conversion.
3Productivity
If multiple DAC and ADC operate with multiple memory elements to perform multiplication function, then productivity is improved, but device complexity increases
Solution Approach 1:
The mesh structure of DACs and ADCs serves multiple functions simultaneously: DACs convert digital input data to analog currents for memory element modulation, ADCs convert analog memory outputs back to digital results, and the mesh topology enables both parallel processing for high productivity and systematic organization for manageable complexity. Each component type (DAC, ADC, memory element) is designed to perform its specific function with standardized interfaces, reducing overall system complexity despite the large number of components.
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
The solution achieves high accuracy, low power consumption, and small silicon area, resulting in superior performance and efficiency for AI edge chips by effectively performing MAC operations with improved signal integrity and conversion rates.
Implementation Method 1
employs single-end current steering Digital-to-Analog Converter (DAC) to convert digital input data to analog current
Implementation Method 2
The current is converted to voltage through a Trans-impedance Amplifier (TIA) or Integrator
Implementation Method 3
The converted voltage is amplified by Variable Gain Amplifier (VGA) then sampled by Successive-Approximation-Register Analog-to-Digital Converter (SARADC) and converted into final digital output data
Data Source
AI summary
Computing-in-memory utilizes memory as weight for multiply-and-accumulate (MAC) operations. Input data multiplies weights to produce output data during the operation. Method of data conversion from input data, memory element to output data is described to enhance the computing efficiency.