Computational Storage Drives With ML Coprocessors for Local Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems face prohibitive costs and complexities due to high-end server requirements and large data storage needs, with network bandwidth inefficiencies arising from vast data transmission.
Innovation Solution
Incorporation of computational storage drives (CSDs) with integrated machine learning coprocessors in storage devices for local execution of machine learning operations, enabling parallel processing and reducing network data transfer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If high-end servers and GPUs are used to execute deep learning algorithms, then machine learning computation capability is improved, but system cost and complexity increase prohibitively
Solution Approach 1:
The patent combines storage functions and machine learning computation functions into a single integrated device. The computational storage drive includes persistent storage media for data storage and a machine learning coprocessor for executing deep learning algorithms, eliminating the need for separate high-end servers and GPUs, thereby reducing system complexity and cost while maintaining computation capability
Solution Approach 2:
The computational storage drive serves multiple functions: it acts as both a storage device for persistent storage of raw data and a computation device with a machine learning coprocessor that can execute various deep learning algorithms. This multi-functionality replaces the need for dedicated storage servers and separate GPU computing systems
2Quantity of substance
If massive amounts of persistent storage are equipped for vast data corpus, then data storage capacity is improved, but network bandwidth consumption and bottlenecks increase
Solution Approach 1:
The patent segments the machine learning processing function from the host system and places it directly in the storage device. Each computational storage drive can independently execute machine learning algorithms on its locally stored data, dividing the overall processing task across multiple distributed drives and eliminating the need to transfer vast amounts of data over the network
Solution Approach 2:
The machine learning coprocessor in the computational storage drive performs preliminary data processing and filtering operations locally before data needs to be accessed by the host. This preliminary action reduces the volume of data that needs to be transferred over the network, as only processed or relevant results need to be transmitted
Data Source
AI summary
Use of computational storage drives (CSDs) in a machine learning pipeline. The CSD may include a machine learning coprocessor capable of natively executing a machine learning model on raw data stored locally at the CSD. In turn, one or more lower order machine learning operations may be executed at a CSD in response to a read/transform command issued by a host. In turn, the CSD may return transformed data comprising data or metadata that is an output of the lower order machine learning operations. This approach may allow for application of a machine learning model locally to input data stored on the CSD having the machine learning coprocessor. This may avoid network bandwidth associated with traditional read and write operations for input and output data from a machine learning pipeline. Moreover, use of CSDs may provide highly parallelized processing using CSDs for application of machine learning operations at the edge of a network.