Transposable Synapse Array Access via Column Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional random access memories (RAMs) are limited to row or column access, leading to inefficiencies in neuromorphic and synaptronic computations, such as increased power consumption and reduced performance, especially in event-driven neural architectures that require transposable access for learning rules like spike-timing dependent plasticity (STDP).
Innovation Solution
Implementing a transposable random access memory using column aggregation, which allows for simultaneous row and column read/write access through a crossbar array with electronic synapses, reducing the number of memory accesses per spike and enabling efficient synaptic weight updates, and also utilizing a recursive array layout to optimize memory access operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If traditional random access memory is used with row or column access only, then the memory structure is simple, but power consumption increases and performance decreases in neuromorphic computations
Solution Approach 1:
The memory array is segmented into multiple banks, where each bank can be independently accessed. This segmentation allows parallel read operations across different banks, reducing the total number of access operations needed for transposable access patterns while maintaining a manageable structure within each bank.
Solution Approach 2:
The patent introduces a bank dimension to the traditional row-column memory structure, creating a three-dimensional access hierarchy (bank, row, column). This additional dimension enables transposable access by allowing the system to switch between row-major and column-major access patterns across different banks, reducing power consumption without excessive complexity.
2Reliability
If transposable access is implemented for STDP learning rules, then learning accuracy improves, but the number of memory access operations increases
Solution Approach 1:
The memory system performs preliminary organization of synaptic weights into banks with specific access patterns. By pre-organizing data in a manner that facilitates both row and column access, the system enables STDP learning rules to execute with fewer access operations, maintaining learning accuracy while improving productivity.
Solution Approach 2:
The memory banks are designed to support multiple access modes (row-major and column-major) universally. This multi-functionality allows the same hardware structure to efficiently support both forward propagation and backward propagation operations in neuromorphic networks, improving access efficiency without compromising learning accuracy.
3Speed
If column aggregation is used for transposable access, then access speed increases, but write operation complexity increases
Solution Approach 1:
Write operations are segmented and distributed across multiple banks rather than requiring simultaneous updates across the entire array. This segmentation allows write operations to be performed in parallel on different banks, increasing access speed while managing write complexity through distributed processing.
Solution Approach 2:
The patent introduces bank-level control logic as an intermediary between the row/column access interfaces and the actual memory cells. This intermediary manages the complexity of write operations by coordinating updates across banks, enabling fast transposable access while abstracting the complexity of write operations from the core memory array.
Data Source
AI summary
Embodiments of the invention relate to providing transposable access to a synapse array using column aggregation. One embodiment comprises a neural network including a plurality of electronic axons, a plurality of electronic neurons, and a crossbar for interconnecting the axons with the neurons. The crossbar comprises a plurality of electronic synapses. Each synapse interconnects an axon with a neuron. The neural network further comprises a column aggregation module for transposable access to one or more synapses of the crossbar using column aggregation.