Memristive Transistor Computing for In-Memory Scalar Multiplication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for processing input variables using transistors are inefficient in performing scalar multiplications and accumulations, particularly in the context of artificial neural networks, due to limitations in controlling charging currents and time behaviors of transistors.
Innovation Solution
A method utilizing memristive elements in series with the control electrode of transistors to control charging currents and time behaviors, enabling precise determination of output variables through time profiles of load path currents, allowing for efficient scalar multiplications and accumulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional transistor methods are used for processing input variables, then device simplicity is maintained, but processing efficiency and precision for scalar multiplications and accumulations deteriorate
Solution Approach 1:
The patent merges the memristive element with the transistor structure by placing it in series with the control electrode, creating an integrated device that performs both memory storage and processing functions. This combination enables efficient scalar multiplications and accumulations without requiring separate memory and processing units, thereby improving productivity while controlling complexity through functional integration.
Solution Approach 2:
The processing device with memristive elements serves multiple functions: it stores input variables through the memristive elements' resistance states, performs scalar multiplications through controlled charging currents, executes accumulations through capacitor charging, and produces output variables through time profile measurements. This multi-functionality improves processing efficiency for neural network operations without proportionally increasing device complexity.
2Measurement precision
If memristive elements are added to control charging currents, then precision in determining output variables improves, but device complexity increases
Solution Approach 1:
The memristive element acts as an intermediary component between the input variables and the transistor control electrode. It precisely controls the charging current by converting resistance state changes into current modulation, enabling accurate determination of output variables through time profile measurements while adding only a single element per transistor rather than complex control circuitry.
Solution Approach 2:
The patent utilizes parameter changes in the memristive element's resistance state to control the charging current characteristics. By programming different resistance values in the memristive elements, the system achieves precise control over charging rates and time profiles, thereby improving measurement precision for output variables without requiring complex mechanical or electronic adjustment mechanisms.
3Speed
If conventional methods are used for scalar multiplications and accumulations, then device simplicity is maintained, but processing speed and efficiency deteriorate
Solution Approach 1:
The patent replaces conventional sequential mechanical or digital computing operations with direct physical analog processes. Scalar multiplications are performed through controlled charging currents that naturally scale with input values, and accumulations are achieved through capacitor charging that inherently sums contributions. This substitution of physical processes for computational operations dramatically improves processing speed while adding only memristive elements without complex mechanical systems.
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
Enables efficient and precise scalar multiplications and accumulations, facilitating the evaluation of artificial neural networks and other compute-in-memory methods by accurately determining product values based on time measurements.
Implementation Method 1
an electrical resistance of the memristive element is specifiable, for example, by programming, whereby a value for the first input variable can be assigned to the memristive element, which value, as already described above, acts on the charging current of the capacitance associated with the control electrode
Implementation Method 2
a charging current of a capacitance associated with a control electrode of the first transistor can be influenced by means of the first memristive element
Implementation Method 3
a first output variable which characterizes at least one product of the first input variable and of the second input variable, on the basis of a first variable characterizing a time profile of a current through a load path of the first transistor
Data Source
AI summary
A method for processing input variables by means of a processing device having least a first transistor. The method including: providing the first transistor and a first memristive element, which characterizes a first input variable associated with the first transistor, wherein a charging current of a capacitance associated with a control electrode of the first transistor can be influenced using the first memristive element; applying to the control electrode of the first transistor a first output variable which characterizes a second input variable associated with the first transistor; ascertaining a first output variable which characterizes at least one product of the first input variable and of the second input variable, based on a first variable characterizing a time profile of a current through a load path of the first transistor.