LSTM MAC Circuit with Threshold-Based Input Reuse
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
LSTM neural networks suffer from long latencies and inefficiencies due to repeated calculations of unchanged input values across consecutive time steps, leading to wasted power and time consumption.
Innovation Solution
Implementing an improved MAC circuit within LSTM cells that stores previously calculated weight and input value product terms, allowing multiplication operations to be skipped if input values change by less than a threshold, and performing fresh calculations only when necessary.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If traditional MAC circuits perform multiplication operations at every time step, then computational accuracy is maintained, but power consumption and latency increase due to redundant calculations
Solution Approach 1:
The patent implements a storage element that retains the previous input value and a comparator that checks whether the current input differs from the previous input before triggering multiplication operations. This preliminary check prevents redundant calculations when inputs remain unchanged, directly reducing power consumption while maintaining computational accuracy when needed.
2Loss of time
If multiplication operations are performed at every time step, then computational accuracy is ensured, but processing time increases due to unnecessary calculations
Solution Approach 1:
The patent extracts the essential computational step (multiplication) from the continuous time-step processing by introducing a conditional trigger mechanism. The comparator isolates the decision logic, and the storage element isolates the state retention, allowing multiplication to occur only when explicitly triggered by actual input changes, thereby reducing latency without compromising accuracy.
3Productivity
If selective input computation is implemented, then power consumption and latency are reduced, but device complexity increases due to additional circuit components
Solution Approach 1:
The patent segments the MAC circuit into distinct functional modules: a storage element for retaining previous input values, a comparator for detecting input changes, and the multiplication unit that operates conditionally. This segmentation allows each component to perform a specific function efficiently, and the modular structure enables selective activation of computational resources, improving overall efficiency despite increased component count.
Data Source
AI summary
An apparatus is described. The apparatus includes a long short term memory (LSTM) circuit having a multiply accumulate circuit (MAC). The MAC circuit has circuitry to rely on a stored product term rather than explicitly perform a multiplication operation to determine the product term if an accumulation of differences between consecutive, preceding input values has not reached a threshold.