Neural Network MAC Array Controller for Elementwise Sum Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for performing elementwise sum operations in neural networks require large weight sizes, leading to worsened operational performance and increased memory consumption.
Innovation Solution
A neural network operation apparatus that includes a receiver for input feature maps, a controller for controlling multiplier-accumulators (MACs) in MAC arrays, and operation engines to perform depthwise convolution and elementwise sum operations efficiently by controlling the operation states of MACs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If elementwise sum operation is performed by concatenating two feature maps in channel direction and performing convolution operation, then the operation can be executed using existing convolution hardware, but large weight size is required which worsens operational performance and increases memory consumption
Solution Approach 1:
The patent extracts only the essential computation needed for elementwise sum by setting most MACs to disable state and enabling only one MAC per channel group. This removes unnecessary weight computations while maintaining the elementwise sum functionality, directly reducing weight size requirements.
Solution Approach 2:
The patent changes the operation parameters by controlling MAC enable/disable states dynamically. By setting operation states of MACs to disable except for one MAC that performs the actual addition, the system transforms a full convolution operation into an efficient elementwise sum operation, reducing memory consumption.
2Ease of manufacture
If elementwise sum operation is performed by concatenating two feature maps and performing convolution, then the operation can be implemented using standard convolution units, but memory consumption increases due to large weight size
Solution Approach 1:
The patent extracts only the necessary MAC operations for elementwise sum, disabling unnecessary MACs. This extraction approach maintains compatibility with standard convolution hardware while eliminating redundant memory accesses and computations, directly reducing memory consumption.
Solution Approach 2:
Instead of enabling all MACs for a full convolution, the patent enables only the necessary portion (one MAC per channel group) to perform elementwise sum. This partial action approach reduces memory consumption by avoiding excessive weight loading and computation while still achieving the desired operation.
3Productivity
If all MACs in MAC array are enabled to perform convolution operation, then comprehensive computation is achieved, but operational performance worsens due to large weight size requirements
Solution Approach 1:
The patent extracts only the essential MAC operations needed for elementwise sum by disabling unnecessary MACs. This maintains sufficient computation capability for the specific operation while dramatically reducing weight size requirements compared to enabling all MACs.
Solution Approach 2:
The patent applies different operation states to different MACs based on their specific function. By setting most MACs to disable state and enabling only the necessary one MAC per channel group, the system optimizes local resource allocation, maintaining computation capability where needed while reducing overall weight size.
Data Source
AI summary
A neural network operation apparatus includes: a receiver configured to receive a first input feature map; a controller configured to control multiplier-accumulators (MACs) included in a first MAC array; and a first operation engine comprising the first MAC array and configured to process the first input feature map based on the MACs of which operation states are controlled.