Mixed-Technology Memory Architecture for Neural Network Calculators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network calculator systems face challenges in improving energy performance and technological robustness, particularly in handling a large number of write and read operations during learning and inference phases, due to the limitations of non-volatile memory technologies.
Innovation Solution
The proposed solution involves a digital neuromorphic network computer architecture that utilizes two types of non-volatile memory technologies: OxRAM oxide-based resistive memories for high read endurance and low read energy, and FeRAM ferroelectric polarization memories for high write endurance and low write energy, allowing for adaptive memory configuration based on the specific operation phase.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single type of non-volatile memory is used to store synaptic coefficients, then the device complexity is reduced, but the energy performance and technological robustness deteriorate due to the inability to optimize for both read and write operations
Solution Approach 1:
The memory system is segmented into two distinct types of non-volatile memory: first non-volatile memory optimized for read operations (high read endurance, low read energy) and second non-volatile memory optimized for write operations (high write cyclability). This segmentation allows each memory type to be specialized for its strengths, resolving the contradiction between simplicity and energy efficiency.
Solution Approach 2:
Different memory locations are assigned different quality characteristics based on their operational requirements. The first non-volatile memory provides high read endurance for synaptic coefficients requiring frequent reading, while the second non-volatile memory provides high write cyclability for coefficients requiring frequent updating. This local differentiation optimizes energy performance without requiring a completely complex unified architecture.
2Productivity
If memory cells undergo intensive rewriting operations during learning phase, then the learning operation can be performed, but the write cyclability and technological robustness of the memory cells deteriorate
Solution Approach 1:
The memory system segments write-intensive operations from read-intensive operations by providing a second non-volatile memory specifically optimized for write operations with high write cyclability. This allows learning operations to proceed with intensive rewriting without compromising the reliability of the overall system, as the first non-volatile memory retaining synaptic coefficients is protected from excessive write cycles.
Solution Approach 2:
The system creates a copy mechanism where synaptic coefficients can be stored in the first non-volatile memory with high read endurance, while allowing intensive write operations to occur in the second non-volatile memory. This copying approach enables learning operations to proceed intensively while preserving the reliability of the primary storage medium.
3Productivity
If memory cells undergo a large number of read operations during inference phase, then the inference operation can be performed, but the read endurance and technological robustness of the memory cells deteriorate
Solution Approach 1:
The system assigns local quality characteristics to different memory types, with the first non-volatile memory specifically optimized for read operations (high read endurance, low read energy). This allows inference operations requiring large numbers of reads to be performed on this specialized memory without compromising the overall reliability of the system, as the second non-volatile memory handles write operations.
4Device complexity
If the neural network calculator is designed for mobile applications with limited complexity, then the system meets embedded constraints, but the energy performance deteriorates due to the inability to optimize memory for specific operations
Solution Approach 1:
The calculator employs a segmented memory architecture with two types of non-volatile memory, each optimized for specific operational phases. This segmentation enables the system to meet embedded constraints while improving energy performance, as each memory type handles the operations it is best suited for, reducing overall energy consumption without requiring a completely complex unified memory 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 approach enhances the energy performance and robustness of synaptic coefficient storage by optimizing memory usage during both learning and inference operations, extending the lifespan of the neural network calculator and reducing energy consumption.
Implementation Method 1
OxRAM oxide-based resistive memories for high read endurance and low read energy
Implementation Method 2
FeRAM ferroelectric polarization memories for high write endurance and low write energy
Data Source
Figure 1
Figure 2a~2b
Figure 3a
AI summary
The invention relates to a computer for executing a calculation algorithm involving a numerical variable according to at least two phases of operation. The computer comprises: - a memory stage comprising: o a first set of memories for storing a first subword of each numerical variable; each memory of said first set being non-volatile and having a first read endurance and a first write cycleability; o a second set of memories for storing a second subword of each numerical variable; each memory of said second set having a second read endurance and a second write cycleability; the first read endurance being greater than the second read endurance and the first write cycleability being less than the second write cycleability.