Memory Readout Using MSB Run-Length Encoding To Reduce CNN Read Energy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) face significant energy consumption due to repetitive memory reads of identical data, which amplifies energy consumption, particularly in near-memory compute macros.
Innovation Solution
A method that leverages the zero-centered Gaussian weight distribution to selectively skip reading certain bits during retrieval by extracting the run-length of the most significant bit (MSB) and encoding this information in memory addresses, reducing the overall read energy per bit.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If memory readout is performed for all inputs in a CNN layer, then complete data retrieval is achieved, but energy consumption increases significantly
Solution Approach 1:
The patent extracts and utilizes the statistical properties of weight distributions (zero-centered Gaussian) to identify and skip reading redundant bits. By extracting the run-length of the most significant bit and encoding it in memory addresses, the system selectively retrieves only necessary data portions, eliminating unnecessary memory read operations while maintaining computational accuracy.
Solution Approach 2:
The patent changes the parameter representation by encoding weight features (run-length of MSB) into memory addresses. This parameter transformation allows the system to leverage statistical characteristics of weight distributions to optimize memory access patterns, reducing the number of bits read from memory while preserving the ability to reconstruct complete weight values when needed.
2Reliability
If all bits of weights are read from memory, then accurate neural network computation is ensured, but read energy per bit increases
Solution Approach 1:
The patent extracts only the essential information (MSB run-length) needed to represent weight characteristics and stores it in memory addresses. During computation, the system reads this compressed representation rather than full weight values, significantly reducing energy consumption while maintaining sufficient accuracy for near-memory compute operations.
Solution Approach 2:
The patent performs preliminary encoding of weight features into memory addresses before neural network operations. This advance preparation allows the system to quickly retrieve compressed weight representations during computation, reducing real-time energy consumption without compromising computational reliability.
Data Source
AI summary
A method of operating a memory device is provided, including operations: generating, based on at least one weight stored in a first memory, a weight feature to be stored in a second memory different from the first memory, wherein the weight feature is associated with a number of repetitious bits, that are in neighbor positions of and the same as a most significant bit, in the at least one weight; and accessing, according to the weight feature and an address of the at least one weight, the first memory and the second memory to transmit the at least one weight to a multiply and accumulate circuit for a first neural network layer operation.