Storage Compute Device with Configurable Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional data storage devices lack the ability to perform complex computations and dynamic reconfiguration of probabilistic inference networks, limiting their functionality beyond basic data storage and retrieval.
Innovation Solution
A storage compute device with a mass storage unit, machine learning module, programmable state machine module, and input/output interfaces, where switching circuitry selectively couples circuit modules based on configuration data stored in the mass storage unit, enabling dynamic reconfiguration and internal computations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional data storage devices are used, then data storage and retrieval functions are provided, but complex computations and dynamic reconfiguration of probabilistic inference networks cannot be performed
Solution Approach 1:
The patent merges storage and computation functions into a single integrated device. The storage device includes processing circuits that can execute computations directly within the storage array, eliminating the need for separate computation devices. This combining of storage and computation capabilities enables the device to perform complex probabilistic inference networks while maintaining data storage functionality.
Solution Approach 2:
The storage device is designed with multi-functionality to perform both data storage and complex computational tasks. The processing circuits are configured to execute various computation types including probabilistic inference networks, allowing the same device to serve multiple purposes: storage, computation, and dynamic reconfiguration of computational graphs.
2Adaptability or versatility
If dynamic reconfiguration is enabled, then adaptability and computational flexibility improve, but configuration management complexity increases
Solution Approach 1:
The system enables dynamic reconfiguration of the computational graph by allowing the processing circuits to be selectively activated or deactivated based on configuration data. The switching circuitry can dynamically connect different processing circuits to different memory devices, enabling the system to adapt its computational architecture in real-time based on operational needs without requiring physical reconfiguration.
3Reliability
If persistent storage of configuration data is implemented, then reliability of mapping information is improved, but storage requirements increase
Solution Approach 1:
The system segments configuration data into different types: first configuration data stored in non-volatile memory for persistent storage of mapping information between processing circuits and memory devices, and second configuration data stored in volatile memory for operational parameters. This segmentation allows critical persistent configuration data to be maintained without requiring excessive storage capacity, as only essential mapping information needs to be permanently retained.
Data Source
AI summary
An apparatus comprises a mass storage unit and a plurality of circuit modules including a machine learning module, a programmable state machine module, and input/output interfaces. Switching circuitry is configured to selectively couple the circuit modules. Configuration circuitry is configured to access configuration data from the mass storage unit and to operate the switching circuitry to connect the circuit modules according to the configuration data.