Partial-Polarization Resistive Devices for Low Mismatch Neural Network Weight Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Resistive Processing Units (RPUs) face challenges in achieving 1000 levels of conductance with minimal systematic mismatch, which is difficult with phase-change memory (PCM) and resistive random access memory (RRAM) due to their fundamental asymmetry, limiting their effectiveness in Deep Neural Networks (DNNs) for tasks like weight update operations.

Innovation Solution

The implementation of partial-polarization resistive electronic devices using a semiconductor material with a channel region, source and drain contacts, and a partial-polarization material, where programming voltage pulses set a partial-polarization state, enabling a diode-connected configuration in neural network circuits for efficient weight updates through stochastic multiplication.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If phase-change memory (PCM) or resistive random access memory (RRAM) is used as the partial-polarization material, then the device can achieve nonvolatile memory functionality, but the fundamental asymmetry of these materials makes it difficult to achieve 1000 levels of conductance with systematic mismatch below 5%

Engineering Contradiction:
Improveconductance level precisionVSAvoidmaterial symmetry
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies asymmetry principle by intentionally designing a symmetric device structure (semiconductor channel with partial-polarization material on one side and gate contact) that can achieve symmetric conductance characteristics despite using asymmetric materials like PCM or RRAM. The gate contact configuration and semiconductor channel structure compensate for material asymmetry, enabling the system to achieve 1000 conductance levels with <5% systematic mismatch.

Inventive Principle:
Principle #4Asymmetry

2Measurement precision

If more conductance levels are achieved to increase neural network accuracy, then the complexity of achieving systematic mismatch below 5% increases significantly

Engineering Contradiction:
Improveneural network accuracyVSAvoidconductance level control complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent uses dynamic control through gate voltage adjustment to achieve precise conductance modulation. The gate contact can dynamically adjust the conductance state of the semiconductor channel by controlling carrier concentration, enabling fine-grained control over 1000 conductance levels without increasing manufacturing complexity. This dynamic control mechanism allows systematic mismatch to be kept below 5% while achieving high neural network accuracy.

Inventive Principle:
Principle #15Dynamics

3Adaptability or versatility

If a gate contact is added to control the conductance state, then programming flexibility is improved, but device structure complexity increases

Engineering Contradiction:
Improveprogramming flexibilityVSAvoidcontact structure complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The gate contact serves multiple functions: it controls conductance state during programming, enables read operations, and can adjust device characteristics dynamically. This multi-functional design provides programming flexibility without proportionally increasing device complexity, as the same gate structure supports multiple operational modes (programming, reading, and conductance adjustment).

Inventive Principle:
Principle #6Universality (Multi-functionality)

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 approach allows for accurate and efficient weight updates in DNNs, reducing training errors and achieving low test errors with improved tolerance to non-systematic asymmetry, while maintaining realistic area constraints, outperforming prior art RPU specifications.

Implementation Method 1

a partial-polarization material on the semiconductor material between the source contact and the drain contact opposite the channel region

Methodology Applied
Scientific EffectPartial-polarization: Polarisation

Implementation Method 2

a semiconductor material including a channel region configured to conduct a current

Methodology Applied
Scientific EffectElectrical conduction: Conduction (electrical)

Data Source

PatentUS20200143879A1Partial-polarization resistive electronic devices, neural network systems including partial-polarization resistive electronic devices and methods of operating the same
Publication Date: 2020.05.07 UNIV OF NOTRE DAME DU LAC
  • US20200143879A1 patent drawing
  • US20200143879A1 patent drawing
  • US20200143879A1 patent drawing

AI summary

An electronic device can include a semiconductor material including a channel region configured to conduct a current, a source contact electrically coupled to the channel region at a first location, a drain contact electrically coupled to the channel region at a second location spaced apart from the first location, a partial-polarization material on the semiconductor material between the source contact and the drain contact opposite the channel region and a gate contact on the partial-polarization material opposite the channel region and ohmically coupled to the drain contact or ohmically coupled to the source contact.