Deep Learning Accelerator Memory Integration for Real-Time Medical Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in efficiently processing medical images in real-time with reduced energy consumption and computation time using Artificial Neural Networks (ANNs), particularly in medical imaging devices like ultrasound probes and CT scanners.
Innovation Solution
An integrated circuit with a Deep Learning Accelerator (DLA) and random access memory is used to process medical images, where the DLA performs parallel vector and matrix calculations optimized for ANN operations, reducing reliance on a Central Processing Unit (CPU) and minimizing data access bottlenecks through high-bandwidth communication with the memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional CPU-based processing is used for medical images, then general-purpose computing flexibility is maintained, but processing speed and energy efficiency deteriorate
Solution Approach 1:
The system segments processing tasks by separating deep learning inference operations from general-purpose CPU tasks. A dedicated Deep Learning Accelerator (DLA) handles ANN computations while the CPU manages system control and non-DL tasks, resolving the contradiction by specializing processing components for specific workloads.
Solution Approach 2:
The DLA acts as an intermediary component between the CPU and memory system, offloading computationally intensive DL operations from the CPU. This mediator approach enables faster processing without requiring the CPU to handle all computations, thus improving productivity while maintaining system manageability.
2Measurement precision
If deep learning models with large parameters are deployed, then processing accuracy is improved, but memory bandwidth requirements and energy consumption increase
Solution Approach 1:
The DLA merges computation and memory access functions into a unified architecture. By integrating on-chip memory with processing units, the system reduces data movement between separate memory and processing components, thereby lowering energy consumption while supporting large-scale deep learning models for accurate medical diagnosis.
Solution Approach 2:
The architecture transitions from traditional von Neumann architecture to a more integrated compute-memory architecture. This dimensional change in system organization allows simultaneous access to model parameters and computation resources, reducing the energy cost of data transfer while maintaining high diagnostic accuracy through comprehensive model processing.
3Loss of time
If real-time processing is implemented, then feedback speed is improved, but computational complexity and energy usage increase
Solution Approach 1:
The system performs preliminary actions by pre-loading deep learning model parameters into on-chip memory before processing begins. This pre-positioning of data eliminates the need for repeated high-energy memory accesses during real-time processing, enabling fast feedback with reduced energy consumption through advance preparation.
Solution Approach 2:
The DLA maintains continuous processing capability by keeping model parameters readily available in on-chip memory, eliminating idle time between computations. This continuous operation mode allows real-time medical image processing with consistent energy efficiency, as the system avoids the energy spikes associated with intermittent high-speed memory access.
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 an image generated in an imaging apparatus configured to image a portion of a person, parameters of an artificial neural network, and instructions executable by the Deep Learning Accelerator to perform matrix computation to generate an output of the artificial neural network. The output can include a feature identified by the artificial neural network and a diagnosis determined by the artificial neural network to assist or guide the imaging of the portion of the person.