Neuromorphic Memory Management via Prescience Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neuromorphic memory allocation methods, such as the daisy chain method, are inefficient and limit the application of neuromorphic devices in low-power devices due to high power consumption and performance dependency on external memory units, preventing the use of recursive operations and deep learning algorithms.
Innovation Solution
A memory management system with a prescience estimator and manager that dynamically allocates memory in a multi-dimensional lattice structure, allowing for variable memory access and efficient use of memory space, enabling recursive operations and deep learning algorithms by storing and processing multiple data sets simultaneously.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a daisy chain memory allocation method is used, then data is sequentially stored and processed, but memory use efficiency is low and power consumption is high
Solution Approach 1:
The patent implements dynamic memory allocation where the memory structure changes based on input data characteristics. The prescience estimator dynamically determines memory requirements, and the prescience manager dynamically reconfigures memory allocation, allowing the system to adapt memory usage to actual needs rather than following a fixed sequential pattern
Solution Approach 2:
The patent transitions from a one-dimensional sequential daisy chain allocation to a multi-dimensional memory structure. Multiple data sets are allocated across different memory regions simultaneously, enabling parallel processing and improving memory utilization efficiency
2Adaptability or versatility
If the whole memory is reset for continuous processing, then another input data can be processed, but processing time is increased
Solution Approach 1:
The patent performs preliminary memory allocation based on prescience estimation before data processing begins. By pre-determining memory requirements and allocating appropriate regions, the system avoids time-consuming reset operations and enables continuous processing without interruption
Solution Approach 2:
The patent enables continuous data processing by maintaining active memory allocations across multiple data sets. Instead of resetting memory between processing tasks, the system keeps memory regions allocated and ready, allowing seamless transition from one data set to the next
3Productivity
If a high system bus rate is required for rapid data exchange, then processing rate is improved, but power consumption increases
Solution Approach 1:
The patent allocates memory regions precisely matching the actual data size requirements rather than using excessive memory capacity. This partial allocation approach reduces the amount of data that needs to be exchanged over the system bus, thereby reducing power consumption while maintaining adequate processing rates
4Use of energy by moving object
If only small area of memories is used due to small input data, then power consumption is reduced, but memory use efficiency is low
Solution Approach 1:
The patent makes memory regions serve multiple functions by allocating them for different data sets simultaneously. A single memory region can be used for storing input data, intermediate results, and final outputs across multiple processing tasks, improving memory utilization efficiency without increasing overall memory size or power consumption
Data Source
AI summary
Provided are a system and method for managing a neuromorphic memory. The system includes a memory having a multi-dimensional lattice structure, a prescience estimator (PSE) configured to separately estimate memory use of two or more data sets and determine a sequence of inputting the two or more data sets to the memory, and a prescience manager (PSM) configured to allocate memory regions in which the two or more data sets will be stored according to estimation results of the PSE, store the data sets, and perform learning and inference.