In-Memory Computing Macro for Reduced SNN/ANN Data Movement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing architectures face limitations in performance and power efficiency due to frequent data movements between processor and memory, which hinder the efficient execution of neural network operations.
Innovation Solution
An in-memory computing (IMC) macro that alternates between a spiking neural network (SNN) mode and an artificial neural network (ANN) mode, utilizing a crossbar array and post-arithmetic circuit to perform operations directly on data stored in memory, reducing data movement and enhancing power efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If data is frequently moved between processor and memory for neural network operations, then computational accuracy is maintained, but power consumption increases and operational efficiency decreases
Solution Approach 1:
The patent merges memory storage and computational processing into a single integrated structure. The crossbar array simultaneously functions as memory for storing weights and as a processing unit for performing multiply-accumulate operations, eliminating the need to move data between separate memory and processor components.
Solution Approach 2:
The crossbar array acts as an intermediary between traditional memory and processor functions. It receives input signals, performs computational operations with stored weights through conductance modulation, and outputs results, thereby mediating the data flow and computation that would otherwise require separate memory access and processing steps.
2Device complexity
If a single IMC macro supports both SNN and ANN operations, then hardware complexity is reduced, but operational mode switching requires additional control logic
Solution Approach 1:
The IMC macro employs dynamic configurability through operating mode signals that alter the functional behavior of its components. The same physical hardware structure can dynamically switch between SNN and ANN operational modes, with the conductance modulation mechanism adapting its characteristics based on the selected mode rather than requiring separate dedicated hardware for each neural network type.
3Adaptability or versatility
If an additional row is added to memory cells for feedback storage, then SNN operations are enabled, but memory cell structure becomes more complex
Solution Approach 1:
The additional row in the crossbar array serves multiple functions: it stores feedback signals for SNN operations, maintains weight values for computational operations, and can be configured for different neural network types. This multi-functional design enables the memory structure to support both SNN and ANN operations without requiring entirely separate storage mechanisms.
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
An in-memory computing (IMC) macro has a mode alternating between a first mode and a second mode, and the IMC macro includes: an input control circuit configured to be capable of generating a signal in which a predefined pattern is applied to an input signal and of transmitting a previous operation result that is fed back, and which is performed depends on which mode the operating mode is in; a crossbar array including memory cells including an additional row that processes and stores the fed-back previous operation result, and columns including an adder tree corresponding to the memory cells; and a post arithmetic circuit configured to be capable of performing a first operation corresponding to a spiking neural network (SNN) and a second operation corresponding to an artificial neural network (ANN), wherein which of the first and second operations is performed depends on which mode is in effect.