Thinned GST Bridge Cell for Gradual PCM Weight Updates in AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Phase change memory (PCM) devices struggle to achieve gradual conductance changes, which are essential for analog representation in AI applications, as existing technologies do not efficiently allow for gradual phase transitions between amorphous and crystalline states.
Innovation Solution
A PCM device with a bridge cell architecture featuring a thinned germanium-antimony-tellurium (GST) portion, where the thickness varies gradually from the electrodes towards the center, enabling a gradual phase change during programming, facilitating both SET and RESET operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a conventional PCM device structure is used, then the device can store multiple resistance levels, but the conductance change is not gradual enough for analog representation in AI applications
Solution Approach 1:
The patent applies local quality by creating a non-uniform GST bridge structure where the thickness varies along the bridge length. The bridge is thinnest at the center and thicker near the electrodes, creating a gradient structure. This localized thickness variation enables the phase change to propagate gradually from the center towards the electrodes during programming, achieving analog-like conductance changes while maintaining overall device functionality
2Measurement precision
If the GST layer thickness is uniform, then the device structure is simple to manufacture, but the phase change cannot propagate gradually for analog weight representation
Solution Approach 1:
The patent employs asymmetry by deliberately creating a non-uniform GST bridge thickness profile. Instead of a uniform thickness, the bridge has a symmetric asymmetric profile where the center portion is thinned relative to the ends. This asymmetric thickness distribution causes the phase change to initiate at the thinnest center region and propagate outward, enabling gradual conductance modulation for analog AI applications
3Speed
If high current is applied to achieve phase change, then the phase transition is fast, but the operating current and voltage are too high for efficient AI applications
Solution Approach 1:
The patent utilizes parameter changes by modifying the physical dimensions of the GST bridge, specifically the thickness parameter. By creating a thinned center region with reduced thickness, the resistance and required programming current are reduced. The phase change propagates through the bridge at controlled current levels, achieving both reasonable switching speed and reduced operating power for AI applications
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 design allows for efficient gradual phase change of the GST material, enhancing the ability of PCM devices to represent multiple resistance levels, which is crucial for AI applications by enabling precise weight updates in machine learning algorithms with reduced operating current and voltage.
Implementation Method 1
Phase change materials can change phase between an amorphous state and a crystalline state by application of specific levels of electrical current or voltage
Implementation Method 2
The different voltage or current levels being applied to change phase among an off state (e.g., no voltage or current applied), the amorphous state, the crystalline state
Implementation Method 3
During PCM programming, corresponding to the weight update in machine learning, the phase change of GST occurs at the thinnest portion of the bridge and gradually propagates towards the electrodes
Implementation Method 4
The amorphous state can be characterized by a relatively higher electrical resistivity than the crystalline state, causing different levels of voltages or current being used for setting the phase
Data Source
AI summary
A memory cell structure includes a substrate having formed thereon a first electrode and second electrodes physically spaced apart. A phase change material (PCM) cell is formed on the substrate and forms a bridge extending between the first and second electrodes, the phase change material including a first end electrically contacting the first electrode and a second end contacting the second electrode. The phase change material cell includes a thinned surface portion where a surface topography of the phase change material cell is decreased relative to a surface topography of the phase change material cell surface at the first and second ends. The PCM thickness is intentionally gradually tapered to localize the formation of the phase change region. During PCM programming, corresponding to the weight update in machine learning, the phase change of the PCM occurs at the thinnest surface portion and gradually propagates towards the electrodes.