Transpose Memory Unit for Multi-Bit CNN Computing-In-Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current memory solutions for multi-bit convolutional neural network (CNN) based computing-in-memory (CIM) applications face challenges in reducing latency and power consumption while maintaining accuracy, particularly due to frequent weight updates and potential write disturb issues.
Innovation Solution
A transpose memory unit and array structure controlled by first and second word lines, incorporating memory cells and transpose cells that allow for simultaneous operation of multi-bit input values and weights without write disturb, utilizing local bit lines and multiple input/output word lines to generate multiply-accumulate output values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the number of operations is increased to enhance performance, then the processing capability is improved, but the latency and power consumption increase due to frequent weight updates
Solution Approach 1:
The memory array is segmented into multiple banks (first bank, second bank, third bank, fourth bank) that can operate independently and simultaneously. This segmentation allows parallel processing of multiple operations without requiring sequential weight updates, thereby reducing latency while maintaining high operation numbers for enhanced productivity.
Solution Approach 2:
Weights are pre-loaded into the memory banks before computation begins. The first weight is loaded into the first bank and the second weight is loaded into the second bank in advance, allowing the computing unit to perform multiple operations using these pre-loaded weights without requiring frequent updates during computation, thus reducing latency.
2Productivity
If the number of operations is increased to enhance performance, then the processing capability is improved, but the power consumption increases due to frequent weight updates
Solution Approach 1:
The memory array is segmented into multiple banks that can operate independently and simultaneously. This segmentation allows parallel processing of multiple operations without requiring sequential weight updates, thereby reducing power consumption while maintaining high operation numbers for enhanced productivity.
Solution Approach 2:
Weights are pre-loaded into the memory banks before computation begins. The first weight is loaded into the first bank and the second weight is loaded into the second bank in advance, allowing the computing unit to perform multiple operations using these pre-loaded weights without requiring frequent updates during computation, thus reducing power consumption.
3Measurement precision
If weights are frequently updated to maintain accuracy, then the computation accuracy is improved, but the write disturb issue occurs
Solution Approach 1:
The memory array is segmented into multiple banks that can operate independently. This segmentation allows the system to maintain accuracy by having dedicated weight storage in each bank without requiring frequent writes to the same location, thereby preventing write disturb issues while maintaining computation accuracy.
Solution Approach 2:
Weights are pre-loaded into the memory banks before computation begins and remain stable during the computation process. This preliminary loading eliminates the need for frequent weight updates during computation, preventing write disturb issues while maintaining accuracy through the use of these stable pre-loaded weights.
Data Source
AI summary
A transpose memory unit for a plurality of multi-bit convolutional neural network based computing-in-memory applications includes a memory cell and a transpose cell. The memory cell stores a weight. The transpose cell is connected to the memory cell and receives the weight from the memory cell. The transpose cell includes an input bit line, at least one first input word line, a first output bit line, at least one second input word line and a second output bit line. One of the at least one first input word line and the at least one second input word line transmits at least one multi-bit input value, and the transpose cell is controlled by the second word line to generate a multiply-accumulate output value on one of the first output bit line and the second output bit line according to the at least one multi-bit input value multiplied by the weight.