In-Memory Computation with Ferroelectric Transistors for Signed Weight Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computation apparatuses experience time delays and significant power consumption due to frequent data transfer between memory and computation units, particularly in neural networks, where unsigned weights complicate accuracy in complex networks.
Innovation Solution
A computation apparatus within a memory module that uses ferroelectric transistors and source follower amplifiers to perform multiplication operations with signed weights, reducing data transfer and power consumption by integrating computation and memory, and employing a pre-charge and discharge circuit to manage accumulation lines.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data is frequently transferred between memory and computation apparatus, then computation can be performed, but time delay increases and power consumption increases
Solution Approach 1:
The patent combines memory and computation apparatus into an integrated structure where computation units are embedded within the memory module. This merging eliminates the need for frequent data transfers between separate memory and computation components, thereby reducing time delay while maintaining computation performance.
2Productivity
If data is frequently transferred between memory and computation apparatus, then computation can be performed, but power consumption increases
Solution Approach 1:
The integrated memory-computation structure reduces power consumption by eliminating frequent data transfers between separate memory and computation components. The computation units perform operations directly on data stored in the memory array, significantly reducing the energy required for data movement.
3Device complexity
If unsigned weight values are used in neural networks, then implementation is simpler, but accuracy cannot be ensured in complex networks
Solution Approach 1:
The patent changes the parameter of weight representation from unsigned to signed values. The computation units are designed to handle signed weight values through differential voltage representations, enabling accurate computation in complex neural networks while maintaining manageable implementation complexity through the integrated architecture.
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 solution reduces time delays and power consumption while maintaining high accuracy in neural network computations, as demonstrated by improved inference accuracy with signed weights compared to unsigned weights.
Implementation Method 1
a source follower amplifier configured with a ferroelectric transistor to output a voltage corresponding to a result of the multiplication operation with respect to an input voltage provided to the word line
Implementation Method 2
the ferroelectric transistor has a threshold voltage corresponding to the weight stored in the ferroelectric transistor
Data Source
AI summary
There is provided a computation apparatus located in a memory module and configured to perform computation with data stored in the memory, the computation apparatus including: a plurality of word lines to which an input is provided; a plurality of unit arrays which store a weight having a sign and perform a multiplication operation on the input provided from the word line and the weight; and an accumulation line connected to the plurality of unit arrays and on which results of the multiplication operations performed by the plurality of unit arrays are accumulated, wherein each of the plurality of unit arrays includes a source follower amplifier including a ferroelectric transistor configured to output a voltage corresponding to a result of the multiplication operation with respect to an input voltage provided to the word line.