Analog Memory Complex MACC Engine for Crossbar Neural Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional crossbar arrays primarily focus on real-valued MACC operations, limiting their ability to efficiently perform complex-valued operations essential for advanced machine learning applications like complex neural networks.
Innovation Solution
An analog memory-based complex MACC compute engine is developed, utilizing pulse-width modulators, differential circuits with configurable resistors, and analog-to-digital converters to enable simultaneous computation of real and imaginary terms in complex MACC operations, allowing for the implementation of complex-valued MACC operations within crossbar arrays.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional crossbar arrays are used for MACC operations, then energy efficiency and throughput are improved, but the ability to perform complex-valued operations is limited
Solution Approach 1:
The crossbar array is segmented into multiple independent sub-arrays, where each sub-array handles a specific component of complex-valued MACC operations (real or imaginary parts). This segmentation allows parallel processing of different components while maintaining the energy efficiency of crossbar architecture.
Solution Approach 2:
The crossbar array is designed to perform multiple functions: it can handle both real-valued and complex-valued MACC operations using the same hardware infrastructure. By configuring the conductance values and control signals appropriately, the same crossbar array can compute real parts, imaginary parts, and their combinations.
2Adaptability or versatility
If complex-valued MACC operations are implemented, then adaptability for deep learning is improved, but device complexity increases
Solution Approach 1:
Complex-valued MACC operations are implemented through periodic switching between different computational modes. The system alternates between computing real parts and imaginary parts by periodically changing the control signals and conductance configurations, which simplifies the overall design compared to simultaneously handling all components.
Solution Approach 2:
Pulse-width modulators serve as intermediary components that translate digital complex-valued inputs into analog pulse widths, which then control the differential circuits. This intermediary layer simplifies the interface between digital control logic and analog computation, reducing overall system complexity.
3Adaptability or versatility
If pulse-width modulators and differential circuits are added, then complex-valued computation capability is improved, but device complexity increases
Solution Approach 1:
The pulse-width modulator and differential circuit functionalities are merged into integrated circuit blocks that can be directly coupled to the crossbar array. This merging reduces the number of discrete components and interconnections, thereby reducing overall device complexity while maintaining complex-valued computation capability.
Solution Approach 2:
The differential circuits automatically handle the computation of real and imaginary parts by utilizing the inherent differential nature of the circuit topology. The circuits self-regulate their operation based on the input signals, reducing the need for additional control logic and simplifying the overall system.
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 extends the energy efficiency and throughput benefits of crossbar arrays into the complex domain, providing flexibility for deep learning and enabling the handling of complex-valued synaptic weights and activations with low overhead, thus supporting the training of deep neural networks.
Implementation Method 1
a first pulse-width modulator configured to generate a first pulse based on a first input, a second pulse-width modulator configured to generate a second pulse based on a second input
Implementation Method 2
Crossbar arrays also support the performance of MACC operations based on the physics of Ohm's law and Kirchhoff s law
Implementation Method 3
the first resistor of the first differential circuit has a configurable resistance, a first terminal of the first resistor is coupled to a voltage via a voltage supply node
Implementation Method 4
the circuit further comprises an analog-to-digital converter having an input coupled to the voltage supply node
Data Source
AI summary
A circuit comprises a first pulse-width modulator configured to generate a first pulse based on a first input, a second pulse-width modulator configured to generate a second pulse based on a second input, a first differential circuit comprising a first transistor, a second transistor, a first resistor, and a second resistor, and a second differential circuit comprising a first transistor, a second transistor, a first resistor, and a second resistor. A gate of the first transistor of the first differential circuit and a gate of the second transistor of the first differential circuit, and a gate of the first transistor of the second differential circuit and a gate of the second transistor of the second differential circuit are configured to be controlled by the first and second pulse width modulators based on the first input and the second input.