Integrated DLA Memory for Parallel Sensor Data Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sensor fusion systems face challenges in reducing energy consumption and computation time, particularly in processing data from multiple sensors using Artificial Neural Networks (ANNs), due to inefficiencies in data access and overlapping computations.
Innovation Solution
An integrated circuit with a Deep Learning Accelerator (DLA) and random access memory is used to optimize sensor fusion by reducing overlapping processing and enhancing data access, allowing for parallel computations and efficient coordination of sensor data from multiple sources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If sensor data is processed using traditional ANNs with sequential processing, then computation accuracy is maintained, but energy consumption and computation time increase
Solution Approach 1:
The patent segments the ANNs into separate processing units, with each unit handling specific sensor data independently before merging results. This allows parallel computation across multiple sensors, reducing total computation time and energy consumption while maintaining accuracy through coordinated processing of divided tasks
Solution Approach 2:
The patent introduces a temporal dimension to data processing by buffering sensor data in memory and processing it in coordinated time windows. This allows overlapping computations across different sensors and time periods, transforming sequential processing into parallel temporal processing that reduces overall computation time and energy usage
2Productivity
If overlapping computations are performed across multiple ANNs, then computation time is reduced through parallel processing, but data access inefficiencies increase
Solution Approach 1:
The patent implements preliminary actions by pre-loading sensor data into memory buffers before processing begins. This allows the ANNs to access data directly from memory rather than requiring continuous data transfers during computation, significantly reducing data access time and enabling efficient overlapping computations
Solution Approach 2:
The patent introduces memory as an intermediary between sensors and processing units. Memory buffers serve as a mediator that stores sensor data locally, allowing fast access during parallel ANN computations without requiring constant communication with the sensor sources, thus reducing data access time and enabling efficient overlapping processing
3Productivity
If separate ANNs process sensor data independently, then processing efficiency is maintained, but sensor fusion accuracy decreases
Solution Approach 1:
The patent merges the outputs of separate ANNs by collecting results from each sensor's processing unit and feeding them into a fusion ANN. This combining approach maintains the processing efficiency of independent ANNs while achieving accurate sensor fusion through the coordinated integration of multiple results
Solution Approach 2:
The patent creates a multi-functional system where individual ANNs serve dual purposes: they process their specific sensor data independently for efficiency, and their outputs are simultaneously used as inputs for the fusion ANN to achieve accurate sensor fusion. This universal approach allows the same processing units to fulfill multiple functions without compromising either efficiency or accuracy
Data Source
AI summary
Systems, devices, and methods related to a Deep Learning Accelerator and memory are described. For example, an integrated circuit device may be configured to execute instructions with matrix operands and configured with random access memory. The random access memory is configured to store input data from a sensor, parameters of a first portion of an Artificial Neural Network (ANN), instructions executable by the Deep Learning Accelerator to perform matrix computation of the first portion of the ANN, and data generated outside of the device according to a second portion of the ANN. The Deep Learning Accelerator may execute the instructions to generate, independent of the data from the second portion of the ANN, a first output based on the input data from the sensor and generate a second output based on a combination of the data from the sensor and the data from the second portion of the ANN.