Field-programmable ferro-diodes for reconfigurable in-memory-computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing architectures struggle to efficiently integrate data-intensive operations such as on-chip storage, content search, and matrix multiplication due to the inefficiencies of von Neumann architectures, particularly in handling big data and AI tasks, where memory access and data movement consume significant energy and time.
Innovation Solution
The development of field-programmable ferroelectric diodes (FeDs) made from Aluminum Scandium Nitride (AlScN) that enable reconfigurable circuits and architectures without transistors, supporting storage, search, and neural network operations with high on/off ratios and self-rectifying properties, compatible with CMOS BEOL processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional von Neumann architectures are used for data-intensive operations, then device complexity and data movement overhead increase, but computing performance and energy efficiency deteriorate
Solution Approach 1:
The patent merges storage and computing functions into a single integrated structure by implementing memory cells that can perform both data storage and logical operations (AND, OR, NOT, XNOR) without requiring separate processing units. This eliminates the von Neumann bottleneck by allowing computations to be executed directly within the memory array, thereby improving computing performance while reducing architectural complexity.
Solution Approach 2:
The memory cells are designed with multi-functionality, capable of performing various logical operations (AND, OR, NOT, XNOR) and serving as both storage elements and computing units. This universal design allows the same hardware structure to handle different computational tasks, reducing the need for specialized components and simplifying the overall architecture while maintaining high productivity.
2Productivity
If data movement between memory and processor is increased to handle big data tasks, then data processing capability improves, but energy consumption increases
Solution Approach 1:
The patent extracts the computing function from the traditional processor and embeds it directly within the memory cells. By taking out the need for data movement between separate storage and processing units, the system performs computations in-place within the memory array, thereby maintaining high data processing capability while eliminating the energy overhead associated with data transfer.
Solution Approach 2:
The memory cells are designed to perform logical operations autonomously without requiring external processing assistance. Each memory cell can independently execute AND, OR, NOT, and XNOR operations on stored data, enabling the memory system to serve its own computing needs and eliminating energy-consuming data movement to external processors.
3Reliability
If transistor-based memory structures are used to achieve high on/off ratios, then storage reliability improves, but device area and manufacturing complexity increase
Solution Approach 1:
The patent removes the transistor component from the memory cell structure, relying instead on the inherent properties of magnetic tunnel junctions (MTJs) to achieve the required on/off ratios and storage reliability. This extraction of the transistor eliminates the need for complex gate control structures, significantly reducing device area while maintaining reliable data storage through the MTJ's natural resistance switching behavior.
Solution Approach 2:
The invention employs composite material structures, specifically magnetic tunnel junctions composed of multiple magnetic layers separated by thin insulating barriers. These composite materials provide high on/off ratios and stable storage states without requiring transistors, thereby achieving reliable storage in a compact form factor that reduces overall device area.
4Adaptability or versatility
If field-programmable ferroelectric diodes are implemented for reconfigurable computing, then adaptability and operational flexibility improve, but device complexity and programming overhead increase
Solution Approach 1:
The ferroelectric diodes are designed to be self-programmable through simple voltage pulses that directly switch the ferroelectric polarization state without requiring complex external control logic. This self-service capability allows the device to reconfigure its operational mode (storage, search, neural network) autonomously based on applied pulses, thereby improving adaptability while minimizing programming overhead and avoiding the need for complicated control circuits.
Solution Approach 2:
The same ferroelectric diode structure serves multiple functions including data storage, content-addressable memory operations, and neural network synapse emulation. This universal design allows a single device type to provide operational flexibility across different computing paradigms without requiring separate specialized components, thereby achieving high adaptability while keeping the device structure relatively simple.
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
The FeDs facilitate compact, fast, and energy-efficient implementations of TCAMs and neural networks, achieving search delays under 100 ps and inference accuracy of ~97.5% on the MNIST dataset, addressing the inefficiencies of traditional architectures.
Implementation Method 1
field-programmable ferroelectric diode (ferro-diode), constructed from an emerging ferroelectric material Aluminum Scandium Nitride (AlScN)
Implementation Method 2
ferro-diodes which are non-volatile and have a high on/off ratio over 10^4 and a high self-rectifying ratio over 10^4
Data Source
AI summary
Ferroelectric diodes comprising materials such as aluminum scandium nitride (AlScN) or hafnium zirconium oxide (HfZrO2) may be formed atop semiconductor structures such as CMOS wafers to create storage memory cells, search Ternary Content Addressable Memory (TCAM) cells, and/neural circuitry. The diodes are non-volatile and field programmable via pulsing to a pulse-number-dependent analog state, with high on/off and self-rectifying ratios. Cells may be formed, for example, with two diodes that are oppositely polarized, and may be achieved without transistors to form, for example, 0T-2R structures.