FFT Circuit Grouping Data for Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT) computations face accuracy issues due to uneven data distribution, leading to loss of smaller data when the dynamic range exceeds the data bitwidth, resulting in reduced computation accuracy.

Innovation Solution

The method involves grouping data based on bitwidth distribution, assigning different data representations with effective bits and group indices, decomposing data into multi-bit parts, and calculating FFT/IFFT results for each part, then adding them, allowing for accurate computation even with non-uniform data distribution.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If conventional automatic gain method is used to scale data to save bitwidth, then resource consumption is reduced, but computation accuracy deteriorates dramatically when data distribution is uneven

Engineering Contradiction:
ImprovebitwidthVSAvoidcomputation accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent segments the FFT computation process into multiple stages, with each stage performing partial FFT operations on subsets of data. This segmentation allows different data representations (including effective bit and group index) to be assigned to different groups, enabling accurate representation of both large and small values without uniform scaling, thus resolving the contradiction between reducing bitwidth and maintaining computation accuracy.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If data is uniformly scaled to represent maximum value, then dynamic range is covered, but smaller data values are completely lost

Engineering Contradiction:
Improvedynamic rangeVSAvoidsmaller data
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent applies local quality by assigning different data representations to different groups of data based on their local characteristics. Instead of uniformly scaling all data to cover the maximum value, each group can have its own effective bit width and exponent, allowing small values in certain groups to be preserved with adequate precision while still representing large values in other groups, thus preventing complete loss of smaller data.

Inventive Principle:
Principle #3Local quality

3Measurement precision

If grouping data is implemented with different data representations, then computation accuracy is improved, but device complexity increases

Engineering Contradiction:
Improvecomputation accuracyVSAvoidcircuit complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements dynamic data representation where the effective bit width and group indices are determined based on the actual data distribution characteristics. The system dynamically adjusts the number of groups and their respective parameters rather than using a fixed complex structure, allowing accurate computation while adapting the circuit complexity to the actual needs of the data being processed.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentUS9832060B2Method and circuit for fast fourier transform
Publication Date: 2017.11.28 MONTAGE TECHNOLOGY CO LTD
  • US9832060B2 patent drawing
  • US9832060B2 patent drawing
  • US9832060B2 patent drawing

AI summary

A method for FFT/IFFT computation, comprising: identifying whether grouping data is needed based on data bitwidth distribution in a set of data, wherein the set of data includes data in a stage of a FFT/IFFT computation; assigning different data representations including effective bit and group index for data identified in different groups if grouping is needed, wherein data in a group have same exponent, and data in different groups have different exponents; and outputting a signal indicating the exponent; for each of a plurality of short sequence FFT/IFFT computation—decomposing data used in present short sequence FFT/IFFT computation into at least a first multi-bit part and a second multi-bit part; respectively calculating FFT/IFFT computation results for the first multi-bit part and the second multi-bit part; adding the FFT/IFFT computation results for the first and the second multi-bit part; scanning a plurality of short sequence FFT/IFFT computation added results.